<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Blackbee AI on Medium]]></title>
        <description><![CDATA[Stories by Blackbee AI on Medium]]></description>
        <link>https://medium.com/@blackbeeai?source=rss-4cb0b92797f2------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*vcdSCQv-k6i7gxul4I5KDw.png</url>
            <title>Stories by Blackbee AI on Medium</title>
            <link>https://medium.com/@blackbeeai?source=rss-4cb0b92797f2------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 20:55:36 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@blackbeeai/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[SAP ECC Support Ends Soon. Fix These Finance Processes Before You Migrate.]]></title>
            <link>https://medium.com/@blackbeeai/sap-ecc-support-ends-soon-fix-these-finance-processes-before-you-migrate-34011cf94708?source=rss-4cb0b92797f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/34011cf94708</guid>
            <category><![CDATA[erp-software]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[sap]]></category>
            <dc:creator><![CDATA[Blackbee AI]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 08:08:30 GMT</pubDate>
            <atom:updated>2026-11-17T08:08:58.544Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Your migration will copy every messy finance habit you have into S/4HANA. Here’s what to clean up first.</em></p><p>If you run finance on SAP ECC, you’ve probably heard the date by now. December 31, 2027. That’s roughly 13 months away.</p><p>And if your S/4HANA project is still a slide in someone’s roadmap deck, you’re in good company. Plenty of finance teams are in exactly the same spot.</p><p>So what actually happens at the end of 2027, and what should finance do about it before the migration starts? Let’s get into it.</p><h3>What actually ends (and what doesn’t)</h3><p>Here’s the timeline, straight from SAP.</p><p><a href="https://news.sap.com/2020/02/sap-s4hana-maintenance-2040-clarity-choice-sap-business-suite-7/">SAP’s official maintenance announcement</a> confirms that mainstream maintenance for Business Suite 7 core applications, which includes ECC, runs until the end of 2027. After that, you can buy optional extended maintenance until the end of 2030. It costs an extra two percentage points on top of your current maintenance fee.</p><p>There’s also a newer escape hatch. SAP’s <a href="https://news.sap.com/2025/08/rise-with-sap-journey-sap-erp-private-edition-transition-option-updates/">ERP, private edition, transition option</a> gives businesses continuity from 2031 to 2033. The catch is that your systems need to move to SAP ERP, private edition on SAP HANA before the end of 2030, and they have to be at least 2 TB in size. Put simply, it’s a paid bridge for large companies, and you have to sign up for the cloud route to get it.</p><p>One thing that trips people up: your ECC system doesn’t switch off on January 1, 2028. It keeps running. What you lose is standard support at standard pricing, and that gets more expensive the longer you wait.</p><h3>Where everyone actually is</h3><p>You might think most companies have already moved. They haven’t.</p><p><a href="https://www.theregister.com/2026/02/05/sap_migrations_research/">The Register reported</a> that, according to Gartner, 39% of SAP’s 35,000 ECC customers worldwide still hadn’t migrated as of Q4 2024, nearly ten years after S/4HANA launched. A <a href="https://www.precisely.com/press-release/new-research-reveals-sap-s-4hana-migration-momentum-despite-ongoing-automation-challenges">Precisely and ASUG survey</a> paints a slightly brighter picture, with 59% of companies fully or partly live on S/4HANA, up 13 percentage points from 2024.</p><p>And plenty of companies have already decided they’ll pay to wait. In a <a href="https://www.theregister.com/2026/02/27/dsag_ecc_sap/">survey by the German-speaking SAP user group DSAG</a>, almost half of the members still investing in ECC said they plan to move to S/4HANA by the end of 2030. That means paying the extended maintenance premium for up to three years.</p><p>That’s a perfectly reasonable choice, by the way. But if you’re going to pay extra for time, you’d better use that time well. And finance is where a lot of the useful prep work sits.</p><h3>Why so many migrations go sideways</h3><p>Here’s the uncomfortable bit.</p><p>That same <a href="https://www.theregister.com/2026/02/05/sap_migrations_research/">ISG research</a> of 200 senior decision-makers found that nearly 60% of SAP migration projects end up delayed and over budget. More than half of the respondents admitted they’d over-customised their legacy ERP to the point where standardising now feels risky.</p><p><a href="https://www.horvath-partners.com/en/press/detail/study-shows-sap-s-4hana-transformations-rarely-go-as-planned-60-percent-exceed-budget-and-schedule-two-thirds-dissatisfied-with-result-quality">Horváth’s study</a> found S/4HANA projects take an average of 30% longer than planned. Fewer than one in ten companies that had finished their transformation did it on schedule.</p><p>It isn’t only the tech, either. An <a href="https://www.theregister.com/2025/11/05/two_years_out_from_support/">ASUG survey reported by The Register</a> found that business process change was the single biggest barrier to migration for 49% of users. The people and process side is just as hard as moving the system.</p><p>And the data? <a href="https://sapinsider.org/blogs/sap-data-migration-delays-xmateria-research/">Xmateria’s research, reported by SAPinsider</a>, found that 57% of SAP customers and experts say data migration regularly delays their projects.</p><p>So why does this matter to finance specifically? Because a huge share of that customisation and messy data lives in your world. Vendor records, approval logic, PO workarounds and exception handling were all built up over 15 years of “we’ll fix it later.” A migration moves all of it, unless you fix it first.</p><h3>Five finance processes to fix before you migrate</h3><h3>1. Clean up your vendor master</h3><p>This is the big one. In S/4HANA, customers and vendors get merged into Business Partners. According to <a href="https://www.ibsolution.com/academy/blog_en/data-and-analytics/why-sap-s4hana-data-migrations-often-take-longer-than-planned">IBsolution</a>, that’s exactly when hidden duplicates surface. Think of the same supplier entered under slightly different names, or one vendor with a separate record for every location. Your AP team has quietly worked around these for years because they know which record is the “real” one. The migration doesn’t know that.</p><p>Start de-duplicating now. Agree on who owns vendor data and who gets the final say on merges.</p><h3>2. Pull your approval rules out of custom code</h3><p>Plenty of ECC systems have approval workflows hard-coded into custom Z-programs. Who approves what, and at which threshold. That’s the kind of over-customisation that makes ISG’s numbers so ugly.</p><p>Write your approval policies down in plain English. Then decide where they should live. If they’re buried in ERP code, every migration (and every policy change) means rebuilding them. If they sit in a layer above the ERP, they carry over untouched.</p><h3>3. Fix intake before it becomes a PO</h3><p>Here’s a question worth asking. How much of your spend starts in an email or a Teams chat and only reaches SAP once someone raises a PO after the fact?</p><p>A <a href="https://blog.basware.com/en/creating-a-positive-spend-culture-to-put-an-end-to-maverick-buying">Hackett Group study with Basware</a> found that many companies lose 10–20% of their targeted savings to maverick buying. A shiny new ERP won’t change that on its own, because the decision happens before anything touches your ERP. Fix where requests come in, and S/4HANA inherits cleaner data from day one.</p><h3>4. Get a grip on your exceptions</h3><p>Ardent Partners’ <em>AP Metrics That Matter in 2025</em> (summarised by <a href="https://www.tungstenautomation.com/blog/ai-in-accounts-payable-metrics-that-matter">Tungsten Automation</a>) puts the average invoice exception rate at 14%, compared with 9% for best-in-class teams. 53% of AP leaders called high exception rates a major hurdle.</p><p>Exceptions don’t disappear when you change ERPs. If your team handles price mismatches and missing receipts through tribal knowledge and spreadsheets today, they’ll be doing the same thing on S/4HANA tomorrow. Map your top exception types now and decide how each one should be resolved.</p><h3>5. Stop reconciling month-end in spreadsheets</h3><p>Here’s a slightly painful truth. You can go live on S/4HANA on time and on budget and still have your team reconciling by hand at every close. SAPinsider covered <a href="https://sapinsider.org/blogs/sap-s4hana-transformation-decisions-kellton/">Kellton’s take on exactly this</a>, and the argument is that the problem usually comes from decisions that were delayed or never made, long before anyone touched the technology.</p><p>So make a list of every spreadsheet that touches your close. Who owns it, and which ERP gap is it covering for? Some of those gaps S/4HANA will fix. Others it won’t, unless someone writes them into the project scope.</p><h3>Migration season is fraud season</h3><p>This one doesn’t get talked about enough.</p><p>During a migration, vendor records get merged and bank details get re-entered by a team that’s already stretched thin. That’s exactly the kind of chaos fraudsters love. <a href="https://www.pymnts.com/news/security-and-risk/2025/fake-invoices-hit-47percent-mid-market-companies-past-year/">PYMNTS Intelligence</a> found that 47% of the mid-sized companies it surveyed had been hit by fake invoice scams in the past year. Many of those started with a supplier whose own email or billing system had been compromised.</p><p>So while you’re cleaning up the vendor master, lock down how bank details can change. Every change should be verified with a call-back to a number you already have on file, never one from the email asking for the change. And keep a clear audit trail of who changed what during cutover. Your auditors will thank you.</p><h3>Five questions finance should ask before kickoff</h3><p>If you’re in finance and someone hands you a migration plan, these are worth raising early:</p><ol><li><strong>Who owns the vendor master clean-up, and when does it finish?</strong> If the answer is “the SI will handle it in month three,” push back.</li><li><strong>Which custom finance programs are we retiring, and which are we rebuilding?</strong> Every rebuild adds time and cost.</li><li><strong>Where will our approval rules live after go-live?</strong> Inside custom code again, or somewhere you can change them without a project?</li><li><strong>What happens to open POs and unpaid invoices at cutover?</strong> You don’t want to find out during your first close.</li><li><strong>How are we testing month-end close before go-live?</strong> A dry run with real data beats a sign-off on a test script.</li></ol><h3>Where Blackbee AI fits in</h3><p>Full disclosure: I work at <a href="https://blackbeeai.com/">Blackbee AI</a>, so here’s how we think about this.</p><p>Blackbee AI is an <a href="https://blackbeeai.com/blackbee-intake-to-pay">agentic Intake-to-Pay platform</a> that sits above your ERP. It handles everything from identifying spend intent at its source (emails, chats, Teams and so on) to routing approvals and sorting out invoices, until clean, structured data hits your ERP.</p><p>For SAP teams, that matters because it <a href="https://blackbeeai.com/integrations/sap">works above SAP ECC today and S/4HANA after you move</a>. The approval rules and exception handling you set up now don’t need rebuilding after go-live. Your SAP system stays your system of record the whole way through.</p><p>Of course, all of those steps are within your explicit control. You choose what Blackbee AI handles, and it does only that, and does it well.</p><h3>The clock’s ticking, so start with finance</h3><p>You don’t need to wait for the migration kickoff to start fixing things. Clean vendor data, documented approval rules, a proper intake process, a handle on your exceptions and fewer close spreadsheets will make your migration faster, whichever route you take. They’ll also make your finance team’s life a lot easier long before go-live.</p><p>And if you’d like to see how this looks on top of your own SAP setup, <a href="https://www.blackbeeai.com/">book a demo.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=34011cf94708" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What AI Agents Actually Do in a Finance Team (And What You Should Never Let Them Do Alone)]]></title>
            <link>https://medium.com/@blackbeeai/what-ai-agents-actually-do-in-a-finance-team-and-what-you-should-never-let-them-do-alone-97c6e04bac38?source=rss-4cb0b92797f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/97c6e04bac38</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[ai-in-finance]]></category>
            <dc:creator><![CDATA[Blackbee AI]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 10:41:40 GMT</pubDate>
            <atom:updated>2026-11-09T10:41:55.149Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A practical look at where AI agents help a finance team, and where people should stay firmly in charge.</em></p><p>Everyone in finance is talking about AI agents right now. Your LinkedIn feed is full of them, and every software vendor you’ve ever bought from has suddenly “gone agentic.”</p><h3>So what’s actually real?</h3><p>The adoption numbers are. A <a href="https://www.wolterskluwer.com/en/news/pr-2025-wolters-kluwer-survey-increasing-adoption-agentic-ai">Wolters Kluwer survey of 392 finance leaders</a> found that only 6% were using agentic AI, but another 38% planned to adopt it within 12 months. That puts 44% of finance teams on agentic AI in 2026, which is more than a sixfold jump. <a href="https://www.gartner.com/en/newsroom/press-releases/2024-09-12-gartner-predicts-that-90-percent-of-finance-functions-will-deploy-at-least-one-ai-enabled-tech-solution-by-2026">Gartner</a> goes further and predicts that 90% of finance functions will deploy at least one AI-enabled technology solution by 2026.</p><p>Here’s the part fewer people mention, though. Gartner also predicts that <a href="https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027">over 40% of agentic AI projects will be canceled by the end of 2027</a>, mostly because costs climb and the business value never becomes clear. It also called out something it named “agent washing,” where vendors rebrand chatbots and assistants as agents. By Gartner’s estimate, only about 130 of the thousands of vendors claiming agentic AI actually offer it.</p><p>So the useful question for a finance team is where agents belong, and where they don’t.</p><h3>First, what is an AI agent? (Put simply)</h3><p><a href="https://www.ibm.com/think/topics/agentic-ai">IBM</a> describes agentic AI as a system that can accomplish a specific goal with limited supervision. That “limited supervision” bit is the whole point.</p><p>A traditional workflow follows rules you wrote in advance. If the invoice is under $5,000, it goes to the department head. If it’s over, it goes to the controller. It does exactly that every single time, even when something about the invoice is clearly off.</p><p>An agent looks at the situation and decides what to do within the boundaries you’ve set. Then it tells you why it made that call.</p><p>Put simply, a workflow follows the map, and an agent reads the road.</p><h3>Why finance teams need this now</h3><p>Most finance teams I talk to aren’t short on software. They’re short on time, and a huge chunk of that time disappears into exceptions.</p><p>In Ardent Partners’ <em>AP Metrics That Matter in 2025</em> report (summarised by <a href="https://www.tungstenautomation.com/blog/ai-in-accounts-payable-metrics-that-matter">Tungsten Automation</a>), 53% of AP leaders named high exception rates as a major hurdle. The average invoice exception rate sits at 14%, while best-in-class teams get it down to 9%. Those same best-in-class teams move an invoice from receipt to payment-ready in <a href="https://europe.thomsonreuters.com/blog/3-ap-metrics-that-matter">3.1 days, compared with an industry average of 17.4 days</a>.</p><p>Then there’s the spend that never touches a formal process at all. A <a href="https://blog.basware.com/en/creating-a-positive-spend-culture-to-put-an-end-to-maverick-buying">study by The Hackett Group with Basware</a> found that many companies lose 10–20% of their targeted savings to maverick buying. That’s people buying outside your approved contracts and suppliers.</p><p>Fraud keeps getting sneakier too. <a href="https://www.pymnts.com/news/security-and-risk/2025/fake-invoices-hit-47percent-mid-market-companies-past-year/">PYMNTS Intelligence</a> found that 47% of the mid-sized companies it surveyed had dealt with fake invoice scams in the past year. Many of those scams started with a supplier whose own email or billing system had been compromised.</p><p>So what ties all of this together? Timing. By the time an invoice lands in your inbox, the spending decision was made days or weeks earlier, usually in an email thread or a Teams chat that finance never saw. Processing that invoice faster won’t change the decision behind it. You have to start earlier.</p><h3>Where AI agents earn their keep</h3><p>These are the jobs where agents genuinely help a finance team. They involve a lot of reading and cross-checking against policy you’ve already written down.</p><p><strong>1. Catching spend intent where it starts.</strong> Someone in marketing emails a vendor about a new campaign. Someone in ops asks on Teams whether they can renew a software licence. An agent can pick up that request at the source and bring it into a governed process before anyone signs anything.</p><p><strong>2. Checking invoices against what was actually agreed.</strong> Your contract says $40 per unit with a 2% early-payment discount. The invoice says $44 and has no discount. A rules engine can only check the invoice against the PO. An agent can read the contract and point to the exact clause the invoice breaks.</p><p><strong>3. Routing approvals by risk instead of dollar amount.</strong> A $2,000 invoice from a brand-new vendor with freshly changed bank details is riskier than a $20,000 invoice from a supplier you’ve paid for six years. Amount-based routing treats the second one as the scary one. An agent weighs vendor history and your policy before it routes anything.</p><p><strong>4. Spotting anomalies before month-end.</strong> A supplier’s prices creep up quarter after quarter. A duplicate comes in once by email and once through a portal. One category runs well ahead of budget. A long-standing vendor’s risk profile suddenly changes. Agents can watch for these continuously, so you find them in week two instead of in the middle of close.</p><h3>Where humans stay in charge</h3><p>This is the part I care about most, and honestly, it’s the part most vendors skip.</p><p><a href="https://kpmg.com/ca/en/insights/2026/05/the-agentic-shift-ai-next-phase.html">KPMG’s 2026 research on the agentic shift</a> found that 51% of organisations cite trust and transparency concerns as the reason employees hesitate to use AI agents. The same research found that 41% use human-in-the-loop models to validate AI outputs, and 38% restrict autonomous decision-making in high-risk use cases.</p><p>That caution makes complete sense to me. Here’s where I’d keep a person firmly in the loop:</p><ul><li><strong>Setting policy.</strong> Approval thresholds and tolerance levels are yours to set. Agents work inside the rules. They shouldn’t write them.</li><li><strong>Big or unusual exceptions.</strong> Anything above a threshold you choose, or anything the agent isn’t confident about, goes to a person.</li><li><strong>Vendor decisions.</strong> Onboarding a new supplier or changing their bank details should always end with a human sign-off. An agent can flag the risk, and you make the call.</li><li><strong>Anything you can’t explain to an auditor.</strong> If an agent can’t show you the reasoning behind a decision, it shouldn’t be making that decision.</li></ul><p>Put simply, you move from approving every transaction to supervising the decisions agents make on your behalf. And you decide how much they handle.</p><h3>How Blackbee AI handles this split</h3><p>Full disclosure: I work at <a href="https://blackbeeai.com/">Blackbee AI</a>, so this is the part where I tell you how we approach it.</p><p>Blackbee AI is an <a href="https://blackbeeai.com/blackbee-intake-to-pay">agentic Intake-to-Pay platform</a>. It sits above your ERP, whether that’s <a href="https://blackbeeai.com/integrations/netsuite">NetSuite</a>, <a href="https://blackbeeai.com/integrations/sage-intacct">Sage Intacct</a>, <a href="https://blackbeeai.com/integrations/dynamics-365">Dynamics 365</a>, <a href="https://blackbeeai.com/integrations/workday">Workday</a> or <a href="https://blackbeeai.com/integrations/sap">SAP</a>, and it governs spend from the moment someone signals they need to buy something.</p><p>That includes picking up intent at its source (emails, chats, Teams and so on), managing commitments and POs, checking contracts, routing approvals, scoring vendor risk, validating invoices and handling everything else in between, until clean, structured data hits your ERP. Eight specialist AI agents, each supported by sub-agents, do that work, and <a href="https://blackbeeai.com/blackbee-agentic-ai">every decision comes with its reasoning</a> so you’re never left guessing why an invoice got flagged.</p><p>Of course, all of those steps are within your explicit control. You choose what Blackbee AI handles, and it does only that, and does it well. Your ERP stays your system of record the whole time.</p><h3>So, should your finance team use AI agents?</h3><p>Probably, yes. Just be picky about where.</p><p>Start with the work that eats your team’s week, like exceptions and contract checks, and keep people on policy and the high-stakes calls. And when a vendor tells you their product is agentic, ask them one question: can it show me why it made that decision? If the answer is vague, you may have found one of those agent-washed tools Gartner warned about.</p><p>If you want a quick gut check on where your team stands, try our <a href="https://blackbeeai.com/agentic-readiness-quiz">Agentic Readiness Quiz</a>. Or if you’d rather see it in action, <a href="https://calendly.com/gautham-blackbeeai/30min">book a 30-minute demo</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=97c6e04bac38" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Only 4% of Companies Manage Their Tail Spend. Here’s Why the Rest Can’t]]></title>
            <link>https://medium.com/@blackbeeai/only-4-of-companies-manage-their-tail-spend-heres-why-the-rest-can-t-59dcc88f9d32?source=rss-4cb0b92797f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/59dcc88f9d32</guid>
            <category><![CDATA[cfo]]></category>
            <category><![CDATA[procurement]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Blackbee AI]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 11:01:01 GMT</pubDate>
            <atom:updated>2026-10-26T11:01:01.661Z</atom:updated>
            <content:encoded><![CDATA[<h3>The money in the tail has no owner. That’s the part no tool has fixed yet.</h3><p>Someone in marketing needs a webinar tool by Friday. It’s $400. They put it on a card, expense it, and get on with their week. Procurement never hears about it.</p><p>Next quarter, a teammate two desks away buys a different webinar tool. Same story.</p><p>Neither of them did anything wrong. They needed something to do their job, and they got it. Now multiply that by every team, every month. That’s your tail spend. And almost nobody is managing it.</p><h3>The number that should bother every CFO</h3><p>In its <a href="https://www.zycus.com/knowledge-hub/research-report/hackett-group-2025-tail-spend-management-study">2025 Tail Spend Management Study</a>, The Hackett Group found that only <strong>4% of companies actively manage most of their tail spend</strong>. Another 64% of procurement leaders said they were dissatisfied or neutral about how they handle it today. (A quick note on sources: the study is distributed by Zycus, a procurement software vendor, which <a href="https://www.zycus.com/blog/spend-management/tail-spend-opportunity-hackett-2025-report">summarised the findings here</a>. Worth knowing when you read the numbers.)</p><p>So what counts as tail spend? The usual rule of thumb is about 20% of total spend, spread across about 80% of your suppliers. McKinsey frames it from the item side: in <a href="https://www.mckinsey.com/capabilities/operations/our-insights/long-tail-big-savings-digital-unlocks-hidden-value-in-procurement">their analysis</a>, the tail covers 80 to 90% of everything a company buys but only 10 to 20% of what it spends.</p><p>And the money is real. McKinsey puts the savings potential at 5 to 15% for companies that haven’t managed the tail well before. Their example is a company with $3 billion in direct spend, $400 million of it in the tail. Leave that tail alone, and you leave around $40 million on the table.</p><p>Hackett’s 2025 numbers tell the same story from a different angle. Companies say they capture 7 to 10% savings on tail spend today, and they believe 16 to 20% is possible.</p><p>So everyone knows the money is there. Why does almost nobody go and get it?</p><h3>Everyone has a good reason the tail isn’t theirs</h3><p>Ask around your own company, and you’ll hear some version of this:</p><ul><li><strong>Procurement</strong> is busy with the big contracts. <a href="https://www.bcg.com/publications/2019/taming-tail-spend">BCG</a> points out that procurement teams naturally focus on the 20% of suppliers that carry 80% of spend. Tail items might get bought once every two or three years, so they rarely come up in supplier conversations.</li><li><strong>Finance and AP</strong> see the purchase when the invoice lands. By then, the money is already committed.</li><li><strong>The person who bought it</strong> just wanted the thing. Nobody told them there was a preferred supplier, and the approved route felt slower than a card.</li><li><strong>IT</strong> often finds out about new software last. <a href="https://zylo.com/blog/hidden-cost-of-shadow-it">Zylo’s research</a> found that 45% of SaaS transactions happen through employee expenses.</li></ul><p>Every one of those reasons makes sense on its own. Put them together, and you get a category of spend that belongs to everybody and nobody at the same time.</p><p>Corpay’s <a href="https://www.corpay.com/it/resources/blog/tail-spend-management">guide to tail spend</a> lays it out in a simple table. For strategic suppliers, the owner is procurement, by name. For tail suppliers, the owner is “frequently nobody.”</p><p><strong>Put simply, tail spend fails because nobody owns it.</strong> The size of each purchase is a side issue.</p><h3>Why the usual fixes keep stalling</h3><p>Companies have thrown plenty at the tail. Marketplaces, purchasing cards, catalogs, outsourced buying desks. According to the Hackett study, those approaches haven’t closed the gap.</p><p>Here’s my read on why. Each fix handles one part of the problem and quietly hands the rest to someone else.</p><p>Catalogs are great when the item you need is in the catalog. BCG notes they can be clunky and are often updated infrequently, so they don’t always have what people need. Outsourcing takes work off your plate, though you become dependent on a provider. And cards? Corpay makes a sharp point here: put the tail on cards and all that reconciliation work piles up in one place. It doesn’t go away. Your AP team just trades invoice headaches for card-coding headaches.</p><p>There’s a people problem too. In an <a href="https://www.sdcexec.com/sourcing-procurement/news/21710545/fairmarkit-tail-spend-becomes-increasing-focus-for-companies">earlier Hackett survey reported by Supply &amp; Demand Chain Executive</a>, 40% of respondents said difficulty managing end users stopped them from handling tail spend properly. That’s the marketing person with the webinar tool. They’re the start of the process, and most programs meet them at the end.</p><h3>The tail is where your risk hides, too</h3><p>Savings get all the attention. Risk deserves some.</p><p>The <a href="https://www.financialprofessionals.org/about/learn-more/press-releases/Details/over-75-percent-of-us-firms-experienced-payments-fraud-in-2025-while-ai-adoption-for-fraud-mitigation-lags">2026 AFP Payments Fraud and Control Survey</a> found that 76% of US organizations faced attempted or actual payments fraud in 2025, and 74% were hit by business email compromise.</p><p>Now think about who those fraudsters like to impersonate. A supplier you pay every month has a history you can check a change request against. A supplier you’ve paid once has no history at all. Your tail is full of suppliers nobody in finance has ever spoken to. That makes it the easiest place to slip in a fake invoice or a “new bank details” email.</p><h3>What owning the tail actually looks like</h3><p>Owning the tail is a lot less dramatic than it sounds. Nobody needs to approve every $60 purchase by hand. What you need is a small set of habits:</p><ol><li><strong>Catch the request where it starts.</strong> Spend usually begins as a message: an email, a Teams chat, a Slack thread, a quick ask in a meeting. If you only see it at the invoice, you’re reviewing a decision someone already made.</li><li><strong>Decide the rules before the request arrives.</strong> Which purchases can go straight through, and which need a second look or a supplier you already have a contract with? Write it down once and apply it every time.</li><li><strong>Name one accountable owner.</strong> Procurement and finance both have a stake, so pick one person to own the program and hold a monthly review both teams attend. Corpay recommends the same, and they’re right.</li><li><strong>Track supplier count per million dollars of spend.</strong> If that number isn’t falling quarter on quarter, the tail is still growing, whatever the invoice count says.</li></ol><p>If you want to see what your tail looks like before you fix it, I wrote about <a href="https://blackbeeai.com/blog/spend-intelligence-software">how spend intelligence turns AP data into something a CFO can act on</a>. Start there.</p><h3>Where Blackbee AI comes in</h3><p>Step one is where most programs fall over, and it’s the reason we built Blackbee AI’s <a href="https://blackbeeai.com/products/intake-agent">Intake Agent</a>.</p><p>It picks up spend intent at its source (emails, chats, Teams, Slack) before anyone has committed a dollar. From there, Blackbee AI can route the approval and check the invoice against what was agreed, all the way until clean, structured data hits your ERP. Of course, every step in between stays within your explicit control. You choose what Blackbee AI handles, and it does only that, and does it well.</p><p>That’s what owning the tail looks like in practice. Someone, or something, is watching the request from the moment it’s made.</p><h3>So, who owns the tail at your company?</h3><p>I’d love to know. Is it procurement, finance, a shared-services team, or honestly nobody? Tell me in the responses. I read every one.</p><p><em>Anupama Nair is Growth Marketing Manager at </em><a href="https://blackbeeai.com/"><em>Blackbee AI</em></a><em>, an agentic Intake-to-Pay platform that helps mid-market finance teams govern every dollar from intent to payment.</em></p><p><strong>Sources</strong></p><ul><li><a href="https://www.zycus.com/knowledge-hub/research-report/hackett-group-2025-tail-spend-management-study">The Hackett Group 2025 Tail Spend Management Study (via Zycus)</a></li><li><a href="https://www.zycus.com/blog/spend-management/tail-spend-opportunity-hackett-2025-report">Zycus: The Silent 20%, summary of the Hackett study</a></li><li><a href="https://www.mckinsey.com/capabilities/operations/our-insights/long-tail-big-savings-digital-unlocks-hidden-value-in-procurement">McKinsey: Long tail, big savings</a></li><li><a href="https://www.bcg.com/publications/2019/taming-tail-spend">BCG: Taming Tail Spend</a></li><li><a href="https://www.corpay.com/it/resources/blog/tail-spend-management">Corpay: Tail Spend Management</a></li><li><a href="https://zylo.com/blog/hidden-cost-of-shadow-it">Zylo: The Hidden Cost of Shadow IT</a></li><li><a href="https://www.sdcexec.com/sourcing-procurement/news/21710545/fairmarkit-tail-spend-becomes-increasing-focus-for-companies">SDCExec: Tail Spend Becomes Increasing Focus for Companies</a></li><li><a href="https://www.financialprofessionals.org/about/learn-more/press-releases/Details/over-75-percent-of-us-firms-experienced-payments-fraud-in-2025-while-ai-adoption-for-fraud-mitigation-lags">AFP: 2026 Payments Fraud and Control Survey</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=59dcc88f9d32" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is Maverick Spend, Really? What the Research Says About Why People Buy Around Procurement]]></title>
            <link>https://medium.com/@blackbeeai/what-is-maverick-spend-really-what-the-research-says-about-why-people-buy-around-procurement-fb78ebc13028?source=rss-4cb0b92797f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/fb78ebc13028</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[maverick-spend]]></category>
            <category><![CDATA[procurement]]></category>
            <dc:creator><![CDATA[Blackbee AI]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 11:36:57 GMT</pubDate>
            <atom:updated>2026-10-20T11:36:57.051Z</atom:updated>
            <content:encoded><![CDATA[<p>Every finance team I talk to has a story like this one. Someone in marketing needed a design agency fast. The approved route meant a request form and a supplier onboarding check that could take two weeks. So they found an agency themselves and paid the first invoice on a corporate card.</p><p>Was that wrong? By the rulebook, yes. That’s maverick spend.</p><p>But here’s the question that interests me more. Why did the rulebook lose?</p><p>Search for “maverick spend,” and you’ll find dozens of near-identical glossary pages. Most of them treat it as a discipline problem: people are breaking the rules, so write clearer rules and enforce them harder. When I went looking for the research behind the topic, I found a much more interesting picture. Academics and analysts have been studying maverick buying since at least the mid-2000s, and their findings rarely make it into the explainers.</p><p>So let’s go through what they actually found.</p><h3>What is maverick spend?</h3><p>Maverick spend is any purchase made outside an organization’s agreed buying process. That could mean buying from a supplier you don’t have a contract with, skipping a required approval, going over a spending limit without sign-off, or buying something your contracts already cover from somewhere else.</p><p>You’ll see it under a few names:</p><ul><li><strong>Maverick buying</strong> is the term most academic papers use.</li><li><strong>Rogue spend</strong> usually means the same thing, though some people reserve it for deliberate workarounds.</li><li><strong>Off-contract spend</strong> is the narrowest version. The <a href="https://www.cips.org/knowledge/glossary-of-terms/">CIPS glossary</a> defines buying off-contract as purchasing an item without a contract when a contract for that item already exists.</li><li><strong>Non-compliant purchasing</strong> is the formal phrasing you’ll find in policy documents.</li></ul><p>Put simply, maverick spend is the gap between how your organisation agreed to buy things and how people actually buy them.</p><h3>How common is it?</h3><p>Here’s where I have to be honest with you. Nobody has a clean global number for how much spend is maverick, and I’d be suspicious of anyone who quotes one confidently. Most figures come from surveys, and surveys measure what people believe as much as what they spend.</p><p>That said, the numbers we do have are worth looking at.</p><p><a href="https://get.coupa.com/rs/950-OLU-185/images/19-Maverick-Spend-Report.pdf">The Hackett Group’s 2019 Maverick Spend Report</a> found that typical organisations kept 74% of their spend on contract. The leaders in their study reached 91%. Respondents also believed that as much as 16% of negotiated savings get lost to maverick buying.</p><p>Then there’s the cost side. <a href="https://link.springer.com/article/10.1007/s10551-008-9768-2">Karjalainen, Kemppainen and van Raaij</a> cite earlier work by Angeles and Nath (2007) suggesting non-compliant purchasing can cost up to 20% more than compliant purchasing (<a href="https://scholarworks.wmich.edu/cgi/viewcontent.cgi?article=4199&amp;context=honors_theses">summarised here</a>).</p><p>A closely related problem is spend nobody can see at all. When <a href="https://www.accountancyage.com/2025/05/22/shadow-spend-black-hole-swallows-10-of-uk-business-expenditure/">Ivalua surveyed 300 UK supply chain decision-makers</a>, 87% said up to a quarter of their spend escapes procurement oversight. On average, about one in every ten pounds was “shadow spend”. Software is a big chunk of it. The <a href="https://zylo.com/news/2024-saas-management-index">Zylo 2024 SaaS Management Index</a> found that more than a third of a typical company’s applications are shadow IT, and 67% of IT leaders named rogue software purchases as a top SaaS challenge.</p><p>What does good look like? <a href="https://artofprocurement.com/blog/maverick-spend-why-its-usually-procurements-problem-not-the-buyers">Art of Procurement</a> points to Merck, which keeps maverick spend at about 2.1% with a long-running “no PO, no pay” policy and monthly tracking. That’s one company, though, so treat it as a benchmark to aim for and nothing more.</p><h3>The many forms of maverick buying</h3><p>One of the most useful ideas in the research is that maverick buying comes in very different flavours.</p><p><a href="https://link.springer.com/article/10.1007/s10551-008-9768-2">Karjalainen, Kemppainen and van Raaij’s 2009 paper in the <em>Journal of Business Ethics</em></a> combined a literature review with interviews of purchasing professionals. They describe forms of maverick buying that range from completely unintentional all the way to outright sabotage.</p><p>That range matters. Someone who didn’t know a contract existed and someone deliberately splitting orders to dodge an approval limit are both “maverick” on paper. They need completely different responses, though.</p><p>A follow-up study by <a href="https://www.ippa.org/IPPC4/Proceedings/04EconomicsofProcurement/Paper4-2.pdf">Karjalainen and van Raaij</a> looked at operative buyers in Finnish government units that used central frame agreements. They focused on three common forms:</p><ol><li><strong>Buying for better terms.</strong> The buyer believes they can get a better deal outside the agreement</li><li><strong>Keeping an existing supplier relationship.</strong> The buyer sticks with a supplier they already know and trust.</li><li><strong>Not knowing the agreement exists.</strong> The buyer simply isn’t aware there’s a contract to use.</li></ol><p>Only the third one is about missing information. The first two are judgement calls, and in some cases the buyer may even be right.</p><h3>Why do people do it?</h3><p>So what actually drives it? Two sources give us a good answer.</p><p>The Hackett Group asked procurement professionals what causes maverick spend in their organisations. Their <a href="https://get.coupa.com/rs/950-OLU-185/images/19-Maverick-Spend-Report.pdf">2019 report</a> ranked the causes like this:</p><ul><li>75% blamed a lack of self-service or guided buying tool</li><li>69% cited a non-compliant mentality</li><li>69% pointed to resistance to behavioural change, even with clear policies</li><li>58% said processes were complex enough to need training</li><li>50% said there was no catalogue of goods and services</li></ul><p>Look at that list again. The top answer is about tooling. Most of the others are about how hard the process is to follow. Only one of the five is really about attitude.</p><p>The Finnish study went a step further and tested which organisational factors actually linked to maverick buying. Karjalainen and van Raaij analysed responses from 247 buyers across six categories, from IT equipment to hotel services. Here’s what they found.</p><p>Things that showed <strong>no significant link</strong> to maverick buying:</p><ul><li>Reward and sanction systems</li><li>The general compliance climate</li><li>How specialised the buyer’s role was</li><li>How long they’d been in the job</li></ul><p>Things that <strong>did</strong> show a link:</p><ul><li><strong>Task autonomy.</strong> Buyers with less freedom to choose suppliers did less maverick buying of every type.</li><li><strong>Training and involvement.</strong> Buyers with less training were more likely to stick with old suppliers or not know about agreements. Buyers less involved with the central purchasing unit were more likely to go hunting for better terms.</li></ul><p>The authors’ recommendation follows from that. Invest in training and involvement, and think carefully about how much autonomy contract users really need. Rewards and sanctions didn’t earn a mention in their list of effective levers.</p><h3>Why sanctions don’t work, and what does</h3><p>This is the part I find most useful for anyone actually trying to fix the problem.</p><p>If you read the vendor glossaries, the fix usually sounds like enforcement. Tighten the policy and send another reminder. The Finnish data suggests that approach has limited reach, because rewards and sanctions showed no significant link to behaviour.</p><p>The practitioner world has started to say the same thing. In a <a href="https://artofprocurement.com/blog/maverick-spend-why-its-usually-procurements-problem-not-the-buyers">July 2026 piece for Art of Procurement</a>, Philip Ideson argues that maverick spend is usually a symptom of a buying process that’s harder than the alternative. His advice is to ask why going around procurement was easier, then make the compliant path faster.</p><p>That lines up with what we see elsewhere in finance. In <a href="https://www.bottomline.com/resources/state-epayables-2025-aps-unfinished-journey">Ardent Partners’ State of ePayables 2025</a>, 49% of AP teams named delayed approvals as a key barrier. If getting a yes takes a week and a corporate card takes a minute, which would you choose?</p><p>So what does work? Pulling the research together, a few levers keep coming up:</p><ol><li><strong>Make the compliant route the fastest route.</strong> Guided buying, catalogues and simple request forms address the top cause in the Hackett data.</li><li><strong>Train people on what agreements exist.</strong> Unawareness is the easiest form of maverick buying to fix, and training was linked to less of it.</li><li><strong>Involve the people who buy.</strong> Buyers who felt disconnected from the central team were more likely to go looking for their own deals.</li><li><strong>Use hard controls where they matter.</strong> A “no PO, no pay” rule, like Merck’s, stops the money moving without a record.</li></ol><p>Notice that only one of those is a control. The others are about making the right thing easier to do.</p><h3>Should you eliminate it at all?</h3><p>Here’s the question most explainers never ask. Is all maverick buying bad?</p><p><a href="https://ideas.repec.org/a/eee/proeco/v179y2016icp77-89.html">Rothkopf and Pibernik’s 2016 paper</a> in the <em>International Journal of Production Economics</em> takes it on directly. Their title says it well: “Maverick buying: Eliminate, participate, leverage?”</p><p>They model maverick buying as a hidden-action problem. Head office can’t see everything local buyers do, and local buyers often know their own markets better than head office does. The usual response is to monitor and penalize, and the authors argue that approach has clear limits.</p><p>Their alternative is what they call “participation menus”. These are contract options designed so local buyers can choose how they take part. Buyers with genuinely better outside options get a structured way to use that knowledge, and the organization gets the benefit instead of losing it.</p><p>Put simply, sometimes a maverick buyer has spotted a better deal. The smart response is to find out what they know.</p><p>That’s a useful reframe. Every maverick purchase carries information. It tells you where your contracts or your buying process are falling short.</p><h3>How to run a maverick spend analysis</h3><p>You don’t need special software to get a first read on your own maverick spend. Here’s a simple version anyone in finance can run:</p><ol><li><strong>Pull 12 months of data.</strong> Export your AP transactions alongside card and expense data.</li><li><strong>Tag each line.</strong> Mark whether it had a PO and whether the supplier and price match a contract.</li><li><strong>Flag the gaps.</strong> Anything with no PO, an unapproved supplier, an off-contract price or a split order is a candidate.</li><li><strong>Look for patterns.</strong> Group by department and by supplier, and see which teams keep showing up.</li><li><strong>Track it monthly.</strong> Report it as a share of your total addressable spend so you can see whether it’s moving.</li></ol><p>One tip from the research: when you find a cluster, talk to the people behind it before you write a new rule. Did they simply not know the contract existed, or did they think they had a better option? Each answer points to a different fix.</p><h3>Real examples of maverick spend</h3><p>Here’s what it tends to look like day to day:</p><ul><li><strong>The urgent agency.</strong> A team signs a creative or consulting agency over email because the approval route would miss a launch date</li><li><strong>The familiar supplier.</strong> An office manager keeps ordering from the stationery shop they’ve always used, even though there’s a national contract.</li><li><strong>The quiet subscription.</strong> Someone puts a software tool on their card for a trial, and it renews every month for two years.</li><li><strong>The split order.</strong> One large purchase gets broken into several smaller ones so each stays under an approval threshold.</li></ul><p>Only the last one is clearly deliberate. The other three are ordinary people trying to get their jobs done.</p><h3>What this means for finance teams</h3><p>The research is clear that maverick spend costs real money, and the losses in negotiated savings are significant.</p><p>It also changes where you should look for the fix. The studies keep pointing at process, tooling, training and involvement. Enforcement on its own reaches a limited part of the problem.</p><p>So the next time a maverick purchase turns up, try asking the question I started with. Why did the rulebook lose? The answer usually tells you exactly what to fix.</p><blockquote>Written by Anupama Nair for <a href="https://blackbeeai.com/">Blackbee AI</a>. Blackbee AI is an agentic Intake-to-Pay platform that sits above your ERP and helps finance teams govern every dollar from intent to payment.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fb78ebc13028" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Time-to-Value: Why Agentic I2P Deploys in Weeks While Legacy P2P Takes a Year]]></title>
            <link>https://medium.com/@blackbeeai/time-to-value-why-agentic-i2p-deploys-in-weeks-while-legacy-p2p-takes-a-year-a5300956ac75?source=rss-4cb0b92797f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/a5300956ac75</guid>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[accounts-payable]]></category>
            <category><![CDATA[enterprise-software]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Blackbee AI]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 11:11:01 GMT</pubDate>
            <atom:updated>2026-10-05T11:11:01.528Z</atom:updated>
            <content:encoded><![CDATA[<p>Let’s be honest about the thing nobody says out loud on the demo call. The reason a lot of mid-market finance teams stall on buying new procurement software has nothing to do with the software. It’s the memory of the last rollout. The one that got sold as six months and turned into eighteen. The consultants who practically moved into the office while your team kept doing everything the old way anyway.</p><p>So when a new tool shows up promising to fix your Intake-to-Pay process, the honest first reaction is dread. Not another year of this.</p><p>I want to take that fear head-on, because it is the single biggest thing that kills these deals, and because the answer actually depends on how the software is built. Some tools take a year because they have to. Others go live in weeks for the same reason. Let’s get into why.</p><h3>Why legacy procure-to-pay actually takes a year</h3><p>Start with the big suites, the Coupas and Aribas of the world. They are powerful, and for a global enterprise they can be the right call. But look at what they ask of you just to get running.</p><p>A full source-to-pay rollout on<a href="https://www.mindsprint.com/resources/blogs/coupa-vs-ariba"> Coupa runs 9 to 18 months</a>. SAP Ariba lands in<a href="https://www.zamp.ai/blogs/best-vendor-onboarding-software-in-2025-sap-ariba-vs-coupa-vs-zip-vs-ai-agents"> the same 12-to-18-month range</a> for a full enterprise deployment, and almost always needs a systems integrator riding shotgun. Ivalua can stretch<a href="https://procurementaiagents.com/compare/sap-ariba-vs-ivalua-vs-coupa"> up to 18 to 30 months</a>. And the services bill to get there often runs<a href="https://procurementaiagents.com/blog/sap-ariba-pricing-breakdown-2026"> one and a half to three times your first-year subscription</a>, before you have processed a single invoice.</p><p>Why so long? Because these platforms replace a huge chunk of how your company operates. You are migrating data, rebuilding approval workflows from scratch, retraining every buyer, and bending the tool to fit your policies through months of configuration. That is real work, and there is no shortcut hiding inside it. The timeline is baked into the architecture.</p><p>This is the part that stings. The business case you signed off on assumed value starting at go-live. If go-live slips by a year, that whole business case slips right along with it.</p><h3>A lot of these rollouts never actually land</h3><p>The year-long timeline would be tolerable if the finish line were guaranteed. It is not. Big enterprise software has a genuinely rough track record.</p><p>Depending on how you define failure,<a href="https://www.erpresearch.com/en-us/erp-implementation-failure-case-studies"> somewhere between 50% and 75% of ERP projects</a> blow past their original budget or their timeline. Panorama’s 2025 research puts the overall<a href="https://kreativecoretech.com/erp-statistics/"> ERP failure rate at 68%</a>. The average implementation runs about<a href="https://kreativecoretech.com/erp-statistics/"> 17 months against the 12 most teams planned for</a>, and budget overruns average 189% across industries. Only about<a href="https://concorderp.com/blogs/why-70-of-erp-implementations-fail-and-how-to-not-be-that-statistic/"> 30% of projects finish on time and on budget</a>, and Gartner expects that by 2027, more than 70% of recent ERP initiatives will still miss their original business goals.</p><p>Read those numbers as a finance person, because that is what you are. You would not approve a capital project with a one-in-three chance of landing on plan. Yet that is roughly the bet a rip-and-replace procurement rollout asks you to make.</p><h3>Why sitting above the ERP changes the whole timeline</h3><p>Agentic Intake-to-Pay goes fast for one structural reason. It does not replace anything.</p><p>That approach is what<a href="https://blackbeeai.com/products"> Blackbee AI</a> is built on. Blackbee AI puts procurement at the front of the process and handles every task that sits inside it. It picks up intent at its source, whether the ask lands in an<a href="https://blackbeeai.com/products/intake-agent"> email, in Slack, in Teams</a>, or anywhere else your people actually work. It handles the invoices and the approvals, along with everything in between, right up until clean, structured data lands in your ERP. Of course, every one of those steps stays under your explicit control. You choose what Blackbee AI handles, and it does only that, and it does it well.</p><p>Notice what is missing from that list. You are not migrating your general ledger. You are not asking three hundred people to learn a new place to do their job. Blackbee AI sits<a href="https://blackbeeai.com/integrations"> above the ERP you already run</a> and connects to it. The thing that takes a year in a legacy rollout, the ripping out and rebuilding of the core, simply never happens.</p><p>That one architectural choice is the whole difference. The big suites are slow because they swallow your ERP. Blackbee AI is quick because it respects it.</p><h3>What a few weeks actually looks like</h3><p>Fast does not mean magic, so let me give you the honest version of how it goes.</p><p>You connect Blackbee AI to your ERP through the<a href="https://blackbeeai.com/products/integration-agent"> Integration Agent</a>, which talks to NetSuite, Sage Intacct, Dynamics, or SAP out of the box. You tell it which workflow you want it to run first, maybe invoice intake and three-way matching for one entity. You set the rules it has to follow, the same policies your team already works to. Then you turn it on for that one slice and watch it run against real invoices.</p><p>Because you started with one workflow instead of boiling the ocean, you see value in that first slice within weeks instead of quarters. Once it earns your trust there, you hand it the next piece. You are expanding on your own schedule, with each step already proven before the next one starts. That is the opposite of a big-bang go-live where everything breaks at once on a date you cannot move.</p><p>For reference, the faster modern tools in this space already land in<a href="https://www.zamp.ai/blogs/best-vendor-onboarding-software-in-2025-sap-ariba-vs-coupa-vs-zip-vs-ai-agents"> weeks to a couple of months</a> rather than years. Blackbee AI is built to live at that end of the range, because it was designed to sit above your stack from day one.</p><h3>What the slow rollout quietly costs you</h3><p>There is a number nobody puts in the proposal. Every month your process stays broken is a month you keep paying for it.</p><p>Think about what a year-long rollout actually means. For that entire year, your team keeps chasing the same exceptions by hand, and the off-contract spend keeps leaking out the side. Every problem the new system was meant to fix is still costing you, every single month, and you are paying consultants on top of it.</p><p>A successful ERP project takes a<a href="https://kreativecoretech.com/erp-statistics/"> median of 16 months just to pay back</a>. A deployment measured in weeks starts paying back almost immediately, because the value shows up the moment that first workflow goes live. So treat deployment speed as a hard number in your business case. Every week you cut off the rollout is a week your savings start sooner.</p><p>When you sit through the next procurement demo, the feature checklist will look impressive. It always does. The question that actually protects you is simpler, and most vendors hate it.</p><p>How soon will this be live and saving me money, on my real workflows, with my team?</p><p>If the honest answer is a year, you already know what that year looks like, because you have lived it. If the honest answer is a few weeks, dig into why, and make the vendor show you. The ones that deploy fast can usually tell you exactly how, because the speed was designed in from the start. Agentic Intake-to-Pay sits above your ERP for precisely this reason. It gets you to value before the old approach has even finished its data migration.</p><p><em>Blackbee AI is an agentic Intake-to-Pay platform, the decision and control layer that sits above your ERP. Because it works above the systems you already run, it goes live in weeks instead of quarters.</em><a href="https://blackbeeai.com/"><em> blackbeeai.com</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a5300956ac75" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Horizontal AP Automation Fails Vertical Finance Teams: The CPG and Healthcare Case]]></title>
            <link>https://medium.com/@blackbeeai/why-horizontal-ap-automation-fails-vertical-finance-teams-the-cpg-and-healthcare-case-fae391702942?source=rss-4cb0b92797f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/fae391702942</guid>
            <dc:creator><![CDATA[Blackbee AI]]></dc:creator>
            <pubDate>Thu, 01 Oct 2026 11:39:23 GMT</pubDate>
            <atom:updated>2026-10-01T11:39:23.313Z</atom:updated>
            <content:encoded><![CDATA[<p>Every industry has its own software. Restaurants run on Toast. Construction runs on Procore. Doctors, dentists, law firms, logistics companies; they all bought tools built specifically for how they work. The data backs the shift. Vertical software companies reach $10 million in revenue about 40% faster than horizontal ones<a href="https://kgt.solutions/resources/blog/vertical-saas-trends-why-niche-platforms-win"> (Bessemer via KGT</a>), and they hold on to customers far better. Nearly half of all new SaaS unicorns are now<a href="https://www.swarnendu.de/blog/why-vertical-saas-is-winning/"> vertical</a>. Investors noticed, and poured more than<a href="https://tech-insider.org/the-rise-of-vertical-saas-why-industry-specific-software-is-winning/"> $18 billion into vertical software in 2025 alone</a>.</p><p>Which raises a question worth sitting with. If every industry deserves software that understands how it actually works, why is accounts payable, the function that literally moves your money, still mostly sold as one generic product for everyone?</p><p>The answer shows up most clearly in two industries where the money moves in genuinely strange ways: consumer packaged goods and healthcare. Let’s look at both.</p><h3>The promise horizontal AP makes</h3><p>Horizontal AP automation runs on one simple assumption. Every company gets invoices, so every company can run them through the same tool. Capture the invoice, match it to a purchase order, send it for approval, post it to the ledger. That loop is real, and for a big chunk of your spend it genuinely works. The best teams running it process an invoice for<a href="https://www.procuredesk.com/best-ap-automation-manufacturing/"> about $2.78 and clear it in roughly three days</a>.</p><p>The problem starts the moment your industry’s money stops moving through that loop. In any business built on complex commercial deals, the biggest money never touches that loop at all. It moves through trade agreements and rebate structures that a generic matching engine has never heard of.</p><p>This is not a rounding error. In 2025,<a href="https://precoro.com/blog/invoice-processing/"> invoice exceptions became the single biggest headache in accounts payable</a> for the first time in almost twenty years of tracking. Processing things faster did nothing to fix it, because the exceptions were never about speed. They were about context the generic tool simply does not have.</p><h3>Case one: when the invoice shows up as a subtraction</h3><p>In consumer packaged goods, the scariest line in accounts payable often isn’t an invoice at all. It’s a deduction.</p><p>Sell through a big retailer like Kroger or Target and the pattern is always the same. They don’t pay you in full and then send a bill when something goes wrong. They just pay you less, and staple a reason code to it. A promotional allowance one month, a shortage claim the next. Your money is already gone before anyone on your finance team gets to decide whether the claim was even fair.</p><p>And the scale is wild. Deductions and chargebacks usually run<a href="https://www.overdeduct.com/deductions/what-is-a-deduction"> between 2% and 15% of gross sales</a> for CPG brands, and climb toward<a href="https://blog.inymbus.com/cpg-trade-claims-and-deductions-explained"> 20% to 30% once you fold in trade promotions</a>. Trade spend overall is the second-biggest cost a brand carries after the product itself, at<a href="https://www.cpgvision.com/blog/deduction-management-is-a-key-component-of-profit-protection-in-cpg"> 15% to 25% of gross sales</a>. Worse, a real chunk of every deduction is just plain wrong. Estimates of the invalid or preventable share<a href="https://www.overdeduct.com/deductions/what-is-a-deduction"> run anywhere from 5% to 60%</a> depending on the retailer, and almost<a href="https://blog.inymbus.com/cpg-trade-claims-and-deductions-explained"> 40% of brands report invalid rates above 10%</a>.</p><p>Put it in money you can feel. Take a brand with $20 million in gross sales. A 10% deduction load that runs 8% invalid works out to<a href="https://www.cpgscout.ai/deduction-management"> roughly $160,000 a year</a> that belongs to the brand and sits in a retailer’s account. Mid-market CPG companies write off<a href="https://www.overdeduct.com/deductions/what-is-a-deduction"> between 1.2% and 2.4% of gross revenue</a> every year to deductions they never clawed back.</p><p>So ask what a horizontal AP tool does with any of this. Honestly, nothing. A deduction is not an invoice, so there is no document to capture. There is no PO to match it against, because the thing that authorized it is a trade promotion that lives in a sales spreadsheet, not your ERP. To validate it, you have to trace a cryptic retailer reason code back to the exact promotion that justified it,<a href="https://www.confidotech.com/blogs/the-playbook-to-manage-and-automate-deductions"> often down to the UPC and months later</a>. A generic matching engine has no field for that. It was built for the bill that lands in your inbox, not the money quietly pulled out of what you are owed.</p><p>This is why deduction validation is a properly vertical problem. A deduction is really trade spend that shows up as a subtraction instead of a bill, and winning it back takes speed and the exact documentation the retailer wants to see. That is deep industry knowledge baked into software, and it is the first thing a horizontal tool leaves out.</p><h3>Case two: when the invoice has to match a contract you did not write</h3><p>Healthcare has the opposite problem, and it lands in the same place.</p><p>Roughly<a href="https://en.wikipedia.org/wiki/Group_purchasing_organization"> 72% of a hospital’s purchases run through group purchasing organization contracts</a>, deals negotiated in advance that fix the price a provider is allowed to pay. GPOs save the system<a href="https://precoro.com/blog/group-purchasing-organizations/"> an estimated $34 billion a year</a>, and used properly, they knock<a href="https://www.getvalify.com/blog/gpo-cost-saving-strategies-hospitals/"> 10% to 18%</a> off supply costs. The catch is in that one word. Properly.</p><p>A GPO contract only pays off if the invoice actually reflects the price you are entitled to, at the right tier, for a facility that hit its compliance target. None of that happens automatically. Hospitals get<a href="https://logicsource.com/group-purchasing-organizations-spend-healthcare-supply-chain-challenges/"> locked into agreements that force most of their spend through a single portfolio</a>, with penalties when they miss targets. Meanwhile clinicians keep buying off-contract anyway. Physician preference items alone make up<a href="https://www.medigroup.com/blog/why-gpo-alignment-matters-more-than-non-acute-providers/"> 40% to 60% of total supply costs</a>, and every off-contract purchase quietly chips away at the savings the contract promised.</p><p>So the real control question in healthcare AP goes deeper than matching an invoice to a PO. Did the vendor bill you the GPO price you earned? And did this purchase even belong on-contract in the first place? That is a validation against a messy external agreement, plus a judgment call at the moment of purchase. GPO contracts on their own rarely stop an off-contract buy as it happens, and GPOs<a href="https://business.amazon.com/en/blog/healthcare-procurement"> do not give you the spend visibility or real-time compliance tracking</a> that finance needs to catch the leak.</p><p>A horizontal AP tool walks in and matches the invoice to the PO, exactly like it was designed to. The invoice can sail through that check and still be wrong, because the vendor billed a non-contract price, or because the item should never have been bought outside the GPO at all. The tool has no idea what contract tier you are on, and no idea which purchases break policy. Rising costs are the<a href="https://www.advisory.com/sponsored/healthcare-procurement"> number one worry healthcare procurement leaders name for 2026</a>, and this blind spot feeds it directly.</p><h3>The same problem wearing two costumes</h3><p>Line the two cases up, and the shape is identical. In CPG, the money hides inside a deduction that has to be matched to a trade promotion. In healthcare, it hides inside an invoice that has to be matched to a GPO contract tier. Two completely different industries, one underlying problem. The hardest and most valuable work in accounts payable was never processing the document. It was validating that document against a commercial agreement the horizontal tool knows nothing about.</p><p>This is how the vertical software story reaches your finance stack. A generic tool handles the clean, universal slice of AP, the invoice that matches its PO and glides straight through. That work is real and worth automating. But the clean slice was never where your margin leaked. It leaks in the exceptions, and the exceptions are industry-specific by their very nature. Andreessen Horowitz put the principle plainly: vertical platforms build up<a href="https://www.saasmag.com/vertical-saas-outperforming-horizontal-2026/"> industry-specific data no horizontal tool can match</a>, and that data is what turns a dumb checker into a real validator.</p><p>Horizontal software, as one good breakdown of the category points out, is<a href="https://abiraja.substack.com/p/horizontal-vs-vertical-software"> wide open to a narrowly focused competitor</a> on exactly this kind of problem. The generalist has to please everyone, so it encodes nobody’s contracts. The specialist encodes the precise agreement that decides where your money goes.</p><h3>What a vertical-aware approach actually looks like</h3><p>If the problem is validation against a domain-specific agreement, then the fix has to start earlier than the invoice. The document is the last step in the chain. By the time it arrives, the deduction is already taken, and the off-contract order is already placed. Your leverage sits upstream, inside the agreement itself.</p><p>That means treating the commercial agreement as a live control rather than a PDF in a folder. In CPG, that agreement is the trade promotion, so a deduction can be matched to the deal that authorized it and challenged the moment it cannot be. In healthcare, that agreement is the GPO contract, so an invoice gets checked against the tier the facility actually earned, and an off-contract purchase gets flagged before it is placed.</p><p>This is the whole idea behind<a href="https://blackbeeai.com/products"> Blackbee AI</a>. Blackbee AI puts procurement at the front of the process and handles every task that falls inside it. That starts with identifying intent at its source, whether it shows up in<a href="https://blackbeeai.com/products/intake-agent"> email, in Slack, in Teams</a>, or in a hallway conversation. From there, it takes the invoices and the approvals off your plate, along with everything that happens in between, right up until clean and structured data lands in your ERP. Of course, every one of those steps stays under your explicit control. You choose what Blackbee AI handles, and it does only that, and it does it well.</p><p>The part that matters for a vertical team is the contract. Blackbee AI reads your agreement and turns its terms into rules that check every charge<a href="https://blackbeeai.com/products/contract-intelligence-agent"> downstream</a>, which is the exact layer a horizontal invoice-matcher is missing. And because it sits above your ERP instead of replacing it, the same control layer bends to whatever agreement your industry runs on, whether that is<a href="https://blackbeeai.com/solutions/cpg"> trade deductions in consumer goods</a> or<a href="https://blackbeeai.com/solutions/healthcare"> GPO pricing in healthcare</a>.</p><p>None of this makes horizontal AP tools bad. They are good at the job they were built for. That job just happens to stop at your ERP’s front door, and in a vertical business, most of the money was already decided long before it got there.</p><h3>The vertical wave is coming for the back office</h3><p>For years, the back office was the last place anyone expected industry-specific software. Accounting is accounting, people figured. Debits on the left. But your money does not move in generic patterns. It moves through the specific commercial machinery of your industry, and accounts payable is the exact spot where that machinery either gets enforced or gets ignored.</p><p>The vertical wave already took the front office and the operations stack. It is moving into finance now, and deduction validation in CPG and contract compliance in healthcare are two of the clearest early signs. The teams that see it coming will stop asking which AP tool processes invoices the fastest. They will start asking the better question. Which tool actually understands how my money moves?</p><p><em>Blackbee AI is an agentic Intake-to-Pay platform, the decision and control layer that sits above your ERP. It is built for mid-market finance teams in industries where money moves through complex commercial agreements.</em><a href="https://blackbeeai.com/"><em> blackbeeai.com</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fae391702942" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[E-invoicing is being rolled out across Australia and New Zealand, and here’s half the problem that…]]></title>
            <link>https://medium.com/@blackbeeai/e-invoicing-is-being-rolled-out-across-australia-and-new-zealand-and-heres-half-the-problem-that-5deff75ced04?source=rss-4cb0b92797f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/5deff75ced04</guid>
            <category><![CDATA[e-invoicing]]></category>
            <category><![CDATA[fintech]]></category>
            <dc:creator><![CDATA[Blackbee AI]]></dc:creator>
            <pubDate>Tue, 22 Sep 2026 13:37:10 GMT</pubDate>
            <atom:updated>2026-09-22T13:37:10.886Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>E-invoicing is being rolled out across Australia and New Zealand, and here’s half the problem that only starts once the invoice lands</strong></h3><p>Right now, all across Australia and New Zealand, finance teams are busy getting themselves Peppol-ready, with Access Points being accredited, participant IDs registered, and ERPs wired into the network. It’s real work, and it’s genuinely worth doing, so if that’s where you are at the moment, you’re in very good company. There’s just one thing worth pausing on before you tick the box and move along to the next project.</p><p>Think of Peppol as the best courier you’ve ever hired. It picks up an invoice from your supplier and delivers it to your system, perfectly packed: structured, machine-readable, verified as authentic, and dropped neatly on your desk in seconds. It’s genuinely brilliant at that particular job, and it does it faster and far more reliably than anything we’ve had before. And then the courier’s work is finished, the box is sitting on your desk, and somebody still has to open it, read what’s inside, and decide what on earth to do about it. That somebody, and the quiet pile of decisions they’re holding, is really the whole story of this piece.</p><p>Here’s why it matters so much. A Peppol invoice arrives in beautiful condition, and it still has to be checked against the contract, matched to a purchase order, weighed against the budget, run past the right approver, and eventually paid. Put simply, Peppol makes your incoming invoices cleaner, while the deciding stays every bit as demanding as it always was. So let’s walk through Peppol properly together: what it actually is, how the network works under the hood, where the ANZ mandates really stand in 2026, and then the half of the problem that begins the very instant the invoice lands.</p><h3>What is Peppol e-invoicing, in plain English?</h3><p>Put simply, Peppol (it stands for Pan-European Public Procurement On-Line, though nobody actually says that out loud at parties) is a shared, governed network for sending business documents, usually invoices, straight from a supplier’s accounting system into a buyer’s accounting system as structured data that both systems can read. Your supplier’s system talks directly to yours, and the invoice appears as clean, ready-to-use data the moment it lands, which, for anyone who’s spent years squinting at PDF invoices and retyping them by hand, feels a little like a small miracle.</p><p>Two things make Peppol genuinely special, and they’re worth understanding before we go further. The first is that it’s a shared standard, so you connect once and reach everyone, meaning any business on the network can trade documents with any other business on it without anyone having to wire up each supplier by hand. The second is that it’s properly governed, so you can actually trust who’s on the other end. OpenPeppol sets the rules, while each country appoints a Peppol Authority that accredits providers and sets local requirements, and that governance is exactly what keeps the whole thing trustworthy at scale.</p><p>Across Australia and New Zealand, the format everyone uses is PINT A-NZ, the Peppol International invoice model tailored to our region, and it’s been the sole supported format since 15 May 2025. Hold onto one idea as we go, because the rest of this piece hangs off it entirely: Peppol standardises how the invoice travels and what shape it arrives in, while the question of what that invoice means for your business, and what you ought to do about it, is a completely separate matter that the network cheerfully leaves entirely to you.</p><h3>How the Peppol network actually works</h3><p>Most people describe Peppol as email for invoices and leave it there, which undersells it badly and quietly hides the exact spot where the interesting gap opens up, so here’s the real mechanism, kept friendly. The network runs on what people call a four-corner model, and the easiest way to picture it is as a relay.</p><p>It starts with your supplier at corner one, who creates the invoice in their own accounting system, and then passes it to corner two, your supplier’s Access Point, an accredited provider that puts the invoice onto the network in the exact right format. From there, it travels to corner three, your Access Point, which pulls that invoice off the network on your behalf, and it finishes at corner four, which is you, receiving a clean, structured invoice straight into your system, ready to use.</p><p>The two Access Points do all the heavy lifting of delivery, and behind the scenes, the network runs a clever shared address book (the SML and SMP, if you enjoy an acronym) that instantly answers one question: who serves this recipient, and what can they receive? You get a Peppol participant ID, usually tied to your ABN or NZBN, so you register just once, and from then on, anyone on the network can find you and deliver straight to you.</p><p>Now here’s a detail that surprises almost everyone, and it’s worth knowing. In Australia, the ATO runs the network as the Australian Peppol Authority; in New Zealand, that role sits with MBIE, whose job is to accredit Access Points, govern the network, and set the local standard. Your actual invoices, though, travel privately, straight between your Access Point and your trading partner’s, and they stay strictly between the two of you, with a handful of other countries adding a fifth corner (Singapore loops its tax authority in through InvoiceNow), while ANZ today still runs a lovely, clean four-corner network.</p><p>Read those four corners back one more time, because this really is the whole hinge of the article. Corners one through four describe a delivery pipeline, and the moment that invoice lands at corner four, structured, verified, and sitting in your system, Peppol has finished its entire job, gloriously. Everything a finance team actually loses sleep over happens in the very next heartbeat, and that heartbeat is where the rest of this piece lives.</p><h3>Where the ANZ mandates actually stand in 2026</h3><p>E-invoicing tends to arrive the way these things usually do: with the government going first and the private sector getting pulled along by what its customers start to expect. So here’s the honest lay of the land, country by country. In Australia, the story begins on 1 July 2022, when every Non-Corporate Commonwealth Entity, meaning the federal departments and agencies, had to be able to receive Peppol e-invoices, and from there the targets ratchet up steadily.</p><p>By 1 July 2026, at least 30% of the invoices those agencies receive need to come through the Peppol network, and by December 2026, they need automated processing and sending of e-invoices, with quarterly progress reports going up to the ATO. For everyone else, meaning plain old business-to-business, Peppol remains voluntary, because the proposed Business eInvoicing Right, which would have allowed any Peppol-ready business to request e-invoices from its trading partners, was explored by Treasury and then set aside in favor of gentle encouragement.</p><p>That B2B point is constantly misreported, so let’s be really clear about it.</p><p>In Australia today, using Peppol between businesses is a choice, and a strongly encouraged one, because with government agencies chasing their targets, suppliers to the public sector are increasingly expected to be Peppol-capable, and over 400,000 businesses are already sitting on the network.</p><p>It stays a choice all the same, even as the direction of travel gets more obvious by the quarter, and two extra details are worth tucking away while we’re here: an Australian e-invoice that meets the A-NZ spec counts as a valid tax invoice even without the words “tax invoice” printed on it, and federal agencies run a “pay on time or pay interest” policy that aims to pay valid e-invoices within five days, with the records kept for five years.</p><p>New Zealand has been walking a very similar path, and in a couple of ways it’s actually a step ahead. Central government agencies have had to be able to receive e-invoices via Peppol since 31 March 2022, under MBIE’s Government Procurement Rules, and from 1 January 2026, any agency handling more than 2,000 domestic trade invoices a year has to be able to send as well as receive, and to meet a five-business-day payment target.</p><p>Then, from 1 January 2027, those agencies must ask their large suppliers, meaning businesses with revenue above NZ$33 million across the two prior accounting periods, to send domestic B2G invoices via Peppol, while B2B across the wider private sector stays voluntary. That 2027 rule is the first ANZ mandate that reaches past government and touches specific private businesses, and while it’s narrow for the moment, it’s a genuine signal of where all of this is heading.</p><p>The ANZ picture looks fairly relaxed today, being mostly government-facing and mostly voluntary for private business. Look a little further afield, though, and the trajectory gets very clear indeed. Singapore has gone all in, running its Peppol-based InvoiceNow network through IMDA, and its GST InvoiceNow requirement now requires GST-registered businesses to transmit invoice data directly to the tax authority through that network.</p><p>It phases in from November 2025 for new voluntary registrants, then rolls out to all remaining GST-registered businesses in waves from 2028 to 2031, making it a genuine B2B mandate running on Peppol rails.</p><p>Europe is moving the very same way. It adopted its VAT in the Digital Age package back in March 2025, so that from 1 July 2030, structured e-invoicing and near-real-time reporting will become mandatory for cross-border B2B transactions across the EU, with member states now free to introduce their own domestic mandates. Belgium has already made Peppol its national e-invoicing backbone, with a B2B mandate taking effect on 1 January 2026, and both France and Germany are phasing theirs in too.</p><p>So here’s the real takeaway for any finance leader staring at all of this. Today’s gentle, government-facing mandate is really just the on-ramp. Treat 2026 as a box to tick, and you’ll very likely be replanning within two years. Treat it as the moment your invoice data finally turns clean, structured, and machine-readable; build for what to do with all that lovely data, and you’ll be the one who actually banks the value.</p><h3>What Peppol takes care of, and the questions that arrive with the invoice</h3><p>Let’s be properly fair about the win here, because it’s a genuine one. Peppol sorts out your inputs, and that’s no small thing, because for years, a huge chunk of the finance team&#39;s effort went into wrestling invoices into a usable state, which meant reading PDFs, fixing OCR errors, chasing missing fields, and keying it all into the ERP. The benchmarks are sobering, with 63% of AP teams still spending more than 10 hours a week on manual processing (IFOL, 2025), so a structured Peppol invoice, delivered straight in with guaranteed fields, wipes out a big slice of that, and that’s real, measurable relief you’ll feel almost immediately.</p><p>So by the time it reaches corner four, Peppol has taken care of a real and useful set of things:</p><ul><li>an invoice that’s clean, structured, and machine-readable</li><li>a sender that’s verified on a governed network</li><li>every mandatory field present and correctly formatted</li><li>valid PINT A-NZ, ready for your system to read</li></ul><p>And in the very same moment, a rather different set of questions quietly lands on your desk:</p><ul><li>Does the price match the contract you actually signed?</li><li>Does it match an approved PO or commitment?</li><li>Have you already paid this one?</li><li>Has this vendor’s rate quietly crept above your agreed terms?</li><li>Is this vendor still low-risk, or has something shifted?</li><li>Who needs to approve this, given the risk and the policy?</li><li>Does this month’s cash flow say pay now, or hold?</li><li>And where does the final, approved decision get written back into the ERP?</li></ul><p>Sit with that second list for a moment, because every single line on it is a decision, and decisions are where the money genuinely lives. A gorgeous, perfectly formatted invoice can still be an exception, which is the thing 53% of AP teams call their single biggest headache (Ardent Partners, 2025), and a flawless Peppol invoice for the wrong amount is still the wrong amount, just arriving very beautifully.</p><p>It can just as easily be a fraud attempt, given that 47% of companies encountered a fake or fraudulent invoice in the past year (MHC, 2025), because verified delivery only confirms that the invoice was delivered safely, while whether it’s actually legitimate is a separate matter.</p><p>And very often, it’s simply the tail end of ungoverned spend, since 40% to 60% of enterprise spend starts outside formal financial workflows, which means the invoice is the last step of a commitment that frequently had no controls upstream at all, so Peppol tidies up that final step, while the commitment itself still needs governing much earlier in the story.</p><p>There’s a quiet little trap hiding in the way e-invoicing tends to get sold, and it’s worth naming plainly. Structured data gets talked about as though it already means “sorted”, when in truth, clean data raises the stakes on every decision, because there’s now no messy-data excuse left for getting one wrong.</p><h3>The layer that has to sit above the network</h3><p>So if Peppol is the courier that drops a spotless invoice at corner four, something has to be standing right there ready to decide, and for most finance teams, that something has always been a person, or, more honestly, a whole relay of people, spreadsheets, and email threads.</p><p>That approach strains at 200 invoices a month and buckles somewhere around 2,000, and cleaner incoming data does nothing to ease the pressure on its own, which is exactly why a new idea has been quietly taking shape in finance circles. It’s the notion of a decision-and-control layer that sits above both the ERP and the Access Point, and some people have started calling the emerging category &quot;agentic Intake-to-Pay.&quot; The simplest way to picture it is as a brain sitting on top of the plumbing, because it takes the clean invoice the network delivers, along with every request, commitment, and contract that led up to it, and it actually reasons about the whole thing.</p><p>So what does reasoning about it look like in the real world?</p><p>Picture the very same invoice landing on a Tuesday morning. The old way, it drops into a queue for a human to open, eyeball, hunt down the PO, dredge up the contract rate from memory, wonder if it’s a duplicate, and forward on to whoever they think signs off on this one, and you know the drill, because we all know the drill.</p><p>With a decision layer doing the thinking, that same handful of seconds goes very differently, because the fields get confirmed against your own records, the line items get read against the contract, and the layer spots that the unit price is running 6% above the rate you agreed, the matching PO gets found, and a quantity variance gets flagged, that price exception trips your approval policy, so the invoice routes to the category owner for a proper look, the vendor’s risk score gets checked and comes back steady, the committed amount folds into your 30/60/90-day cash-flow view, and once it’s all resolved, the approved decision gets written straight back into the ERP.</p><p>The invoice that arrived was identical in both stories, and what happened to it afterward was worlds apart, purely because this time, something actually thought about it.</p><p>There’s an important reassurance buried in there for anyone getting Peppol-ready right now: a decision layer like this sits above your existing systems and complements the network, leaving it free to do what it does so well.</p><p>Your Access Point continues sending and receiving precisely as the mandate requires; your ERP remains your trusted system of record, and the intelligence simply lives one story up, turning clean, delivered data into governed decisions.</p><p>And there’s a deeper point tucked underneath all of that too, because the invoice is really one of the very last events in a much longer story, so that, by the time it reaches corner four, the spend was already intended, the vendor already chosen, the contract already signed, and the PO already raised.</p><p>E-invoicing gives that long story a lovely, tidy ending, while the real prize is governing the whole arc, all the way from the first flicker of spend intent to a payment decision posted back to the ERP, and Peppol turns out to be a wonderful rail for one document type near the end of it, and one genuinely useful piece of a much bigger picture.</p><h3><strong><em>Clean data is where the good part starts.</em></strong></h3><p>Peppol is well worth doing, and the ANZ mandates make it unavoidable for anyone touching government, so please don&#39;t read any of this as an argument against it, because it genuinely improves how invoices arrive. But a mandate is fundamentally a plan for your inputs, and a spotless invoice in corner four is beautiful raw material for the work that’s always been the hard part: deciding whether it’s right, whether it matches, who signs off, and whether you actually pay. Get Peppol-ready with no plan for that, and you’ll have paved a gorgeous new road that still runs straight into a manual toll booth.</p><p>The finance teams who win in 2026 will be the ones who see the mandate for what it really is: the moment their invoice data finally becomes machine-readable, and who then start building for the decisions that arrive the instant it lands. Clean data is the beginning of the interesting work, and the interesting work is where the value has quietly been hiding all along, so the smartest move you can make this year is to get the delivery sorted and then look one story up.</p><p>Peppol gets the invoice to your door in perfect shape, and the next move, made wisely and at speed, is deciding exactly what happens to it once it’s there.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5deff75ced04" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Will AI Replace Accounts Payable Jobs?]]></title>
            <link>https://medium.com/@blackbeeai/will-ai-replace-accounts-payable-jobs-75523c511e7e?source=rss-4cb0b92797f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/75523c511e7e</guid>
            <dc:creator><![CDATA[Blackbee AI]]></dc:creator>
            <pubDate>Mon, 21 Sep 2026 11:33:13 GMT</pubDate>
            <atom:updated>2026-09-21T11:33:13.964Z</atom:updated>
            <content:encoded><![CDATA[<h4>The honest answer isn’t yes or no. It’s a split, and the data shows exactly where the line falls.</h4><p>If you run an accounts payable function, you’ve probably had the thought at 6pm during a month-end close: <em>how much of this will still be my job in five years?</em> It’s a fair question, and the internet’s answer is uselessly binary. Half the headlines promise AP is about to be fully automated. The other half insist accountants are irreplaceable. Both are selling something.</p><p>The actual evidence is more interesting than either camp. It says AI is already replacing <em>tasks</em> inside AP at scale while simultaneously growing the <em>value</em> of the people who do the work. To see how both can be true, you have to stop asking “will AI replace accountants” and start asking which parts of the job it touches.</p><h3>The government data already answers the headline</h3><p>Start with the most boring, most reliable source there is: the U.S. Bureau of Labor Statistics. It splits finance work into two occupational buckets, and the two are moving in opposite directions.</p><p>Employment of<a href="https://www.bls.gov/ooh/business-and-financial/accountants-and-auditors.htm"> <strong>accountants and auditors</strong> is projected to grow 5% through 2034</a>, faster than the average occupation, with well over 100,000 openings a year. Meanwhile,<a href="https://www.bls.gov/ooh/office-and-administrative-support/bookkeeping-accounting-and-auditing-clerks.htm"> <strong>bookkeeping, accounting, and auditing clerks</strong> are projected to <em>decline</em> 6%</a> over the same period, a net loss of roughly 94,300 positions from a base of about 1.6 million. BLS is blunt about the cause: software has automated many of the tasks clerks used to do, so the same volume of work now needs fewer people.</p><p>That’s the whole debate in two numbers. The judgment role grows. The keystroke role shrinks. And most AP jobs today sit somewhere on the spectrum between them, which is exactly why the future feels uncertain from inside the seat.</p><p>There’s a wrinkle worth naming, because it looks like a contradiction. Goldman Sachs, analyzing <em>task exposure</em> rather than employment,<a href="https://curatesuite.com/accounting/articles/accounting-talent-crisis-ai-and-jobs"> rated accountants and auditors among the roles most exposed to AI</a>. How can the same job be “high growth” and “high risk”? Because task exposure measures how many of a job’s <em>tasks</em> AI could theoretically do, not how many <em>jobs</em> disappear. A role can be full of automatable tasks and still grow, if the human keeps the tasks machines can’t do and sheds the ones they can. Hold onto that idea. It’s the key to the entire question.</p><h3>What AI genuinely does well in AP</h3><p>Let’s be specific instead of hand-wavy, because vagueness is how bad predictions get made. Modern AP systems reliably handle the high-volume, rules-based layer of the work:<a href="https://verito.com/blog/will-ai-replace-accountants/"> transaction classification, bank reconciliation, invoice matching, document extraction from PDFs, and routine report generation</a>. These are the tasks that used to fill an AP clerk’s morning.</p><p>And the productivity gains are real, not marketing. According to<a href="https://ardentpartners.com/ap-metrics-that-matter-in-2025/"> Ardent Partners’ <em>Accounts Payable Metrics That Matter in 2025</em></a>, best-in-class AP teams now process an invoice for about <strong>$2.78</strong> versus a market average closer to <strong>$9 to $11</strong>, and close the cycle in <strong>3.1 days</strong> against <strong>17.4 days</strong> for everyone else. Roughly<a href="https://www.medius.com/resources/guides-reports/ardent-partners-accounts-payable-metrics-that-matter/"> <strong>75% of AP departments now use some form of AI</strong></a>. Automation has already moved from differentiator to table stakes.</p><p>So if you define the AP job as “key the invoice into the ERP,” then yes, that job is being automated, and the BLS clerk numbers are already pricing it in. Nobody should sugarcoat that.</p><p>But that definition was never the whole job. And the benchmark data itself tells you where the ceiling is.</p><h3>The exception is the job</h3><p>Here’s the number the automation vendors are quieter about. Even best-in-class teams only reach<a href="https://parseur.com/blog/ai-invoice-processing-benchmarks"> <strong>49.2% touchless (straight-through) processing</strong></a>; the all-buyer average is closer to a third. The theoretical enterprise ceiling, with heavy purchase-order discipline, tops out around 89%, and almost nobody operates there. Which means that across the entire market, somewhere between <em>half and two-thirds of invoices still require a human to touch them.</em></p><p>Why? Because most of those touches aren’t data-entry failures the software could learn to fix. As the benchmark analysts note,<a href="https://parseur.com/blog/ai-invoice-processing-benchmarks"> most exceptions are not extraction problems</a>. They’re judgment problems. A price that doesn’t match the contract. A delivery that arrived partial. An invoice with no PO. A vendor who’s disputing a charge. In fact,<a href="https://www.tungstenautomation.com/blog/ai-in-accounts-payable-metrics-that-matter"> <strong>53% of AP leaders name invoice exceptions as their single biggest challenge</strong></a>, ahead of fraud risk and slow approvals.</p><p>Exceptions are where AP stops being clerical and starts being <em>decisional</em>. Resolving one means weighing policy, chasing context, judging a vendor relationship, and taking accountability for the call. That’s not a task you route to a model and forget. It’s the part of the job that was always the actual job. It was just buried under the keystrokes.</p><p>There’s a hard technical reason the machine can’t own that layer outright, too. A<a href="https://www.accountingtoday.com/list/ai-cant-replace-accountants-when-could-it"> study published in <em>Nature</em> found even the best large language models still hallucinate at meaningful rates</a>, inventing plausible, confident, wrong answers. In a function whose entire purpose is control and accuracy, a system that fabricates at high rates on certain tasks cannot be the last line. It can draft the answer. A human still has to own it. Genuine displacement, as one accounting analyst put it, would require AI to take ambiguous problems end to end <em>with accountability</em>, and that bar is nowhere close.</p><h3>The evidence says AI makes AP people more valuable</h3><p>This is where the “robots are coming” narrative collides with the actual labor data, and the data wins.</p><p><a href="https://www.pwc.com/gx/en/news-room/press-releases/2025/ai-linked-to-a-fourfold-increase-in-productivity-growth.html">PwC’s 2025 Global AI Jobs Barometer</a> analyzed close to a billion job ads across six continents. Its headline finding is almost the opposite of the fear: job numbers are <strong>growing in every industry studied, including the roles considered most automatable</strong>, and jobs that are <em>augmented</em> by AI are growing faster than average. Workers with AI skills command a<a href="https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-jobs-barometer.html"> <strong>56% wage premium</strong></a>, more than double the premium of a year earlier. AI, in PwC’s framing, isn’t erasing workers in exposed fields. It’s raising their price.</p><p>The field studies point the same way. Research from<a href="https://www.xero.com/us/accountant-bookkeeper-guides/will-ai-replace-accountants/"> Stanford Graduate School of Business found that accountants using AI-assisted workflows supported more clients per week, finalized monthly statements about 7.5 days faster, and produced more granular reporting</a>, with no drop in quality. And adoption is moving fast: AI use in accounting firms<a href="https://verito.com/blog/will-ai-replace-accountants/"> jumped from 9% to 41% in a single year</a>, while a 2025 Intuit survey found<a href="https://www.intuit.com/blog/innovative-thinking/will-ai-replace-accountants/"> 93% of accountants already using AI to support advisory work</a>.</p><p>Read those together, and a pattern emerges. When AI absorbs the repetitive layer, the humans don’t vanish. They move upstream into analysis, exception judgment, and advisory work that the business values more. The productivity gain becomes a promotion, not a pink slip. That’s the human-in-the-loop model in practice, and it’s already the dominant one.</p><h3>The plot twist: AP’s real problem is a shortage, not a surplus</h3><p>Here’s the fact that reframes the whole anxiety. The binding constraint in finance right now isn’t too many people. It’s too few.</p><p>Between 2020 and 2025,<a href="https://www.hirecruiting.com/newsroom/solving-the-accounting-and-finance-talent-shortage/"> roughly 340,000 accountants and auditors left the U.S. workforce</a>. The earlier stretch alone represented<a href="https://www.accountingtoday.com/news/accounting-talent-shortage-worsens"> a 17% drop in the profession</a>. The pipeline behind them is thin: CPA exam candidates fell nearly 30% between 2016 and 2024. Unemployment among accounting professionals now<a href="https://www.roberthalf.com/us/en/insights/hiring-help/accountant-shortage-how-to-compete-for-skilled-talent"> hovers near a historic 1% to 2%</a>, and 62% of finance leaders report real difficulty hiring and keeping staff. The consequences are showing up in the numbers that matter:<a href="https://www.auxis.com/theres-a-shortage-of-accountants-what-can-you-do-about-it/"> nearly 640 U.S.-listed companies reported material weaknesses tied to the accounting talent shortage</a> in a single year.</p><p>So picture the actual situation on the ground. AP teams are chronically short-staffed, the experienced people are retiring, and the work keeps growing. In that world, automating the keystroke layer isn’t taking anyone’s job. It’s covering a gap employers can’t hire their way out of. It lets a lean team clear the volume so its scarce, expensive humans spend their hours on the exceptions, the controls, and the analysis that actually need a person. The technology is arriving precisely because there aren’t enough people. That’s a very different story than “the robots are here to fire you.”</p><h3>What the AP job actually becomes</h3><p>None of this means “relax, nothing changes.” The clerk numbers are declining for a reason, and the person most exposed is the one whose entire role is keying clean invoices and nothing else. That version of the job is genuinely fading, and pretending otherwise doesn’t help anyone.</p><p>The role that replaces it looks different. It’s less <em>processor</em>, more <em>supervisor of automation</em>: someone who reviews what the system flagged, resolves the judgment calls, manages vendor relationships, watches for fraud and policy drift, and turns AP data into something the CFO can act on. The daily texture shifts from “process this invoice” to “the model handled the clean 60%; here are the 40% that need a decision, and here’s what the spend is telling us.”</p><p>If you’re in AP and want to be on the growth side of that BLS split rather than the shrinking one, the research is unusually clear about what to build:</p><ul><li><strong>Exception judgment.</strong> The interpretive, accountability-heavy work that stays human by design.</li><li><strong>AI fluency.</strong> The<a href="https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-jobs-barometer.html"> 56% wage premium is real</a>, and it goes to the people who can drive the tools, not just tolerate them.</li><li><strong>Analysis and advisory.</strong> The<a href="https://www.intuit.com/blog/innovative-thinking/will-ai-replace-accountants/"> work that’s moving upstream</a> as automation absorbs the groundwork.</li><li><strong>Controls and risk.</strong> A function built on accuracy will always need a human who owns the final call.</li></ul><h3>So, will AI replace accounts payable jobs?</h3><p>Put simply: AI is replacing accounts payable <em>tasks</em>, not accounts payable <em>people</em>, and it’s reshaping what the remaining people do. The pure data-entry job is contracting, exactly as the<a href="https://www.bls.gov/ooh/office-and-administrative-support/bookkeeping-accounting-and-auditing-clerks.htm"> BLS clerk projections show</a>. But the judgment, exception, and control work is growing more valuable; the<a href="https://www.pwc.com/gx/en/news-room/press-releases/2025/ai-linked-to-a-fourfold-increase-in-productivity-growth.html"> labor data shows augmented workers gaining, not losing</a>, and a<a href="https://www.hirecruiting.com/newsroom/solving-the-accounting-and-finance-talent-shortage/"> severe talent shortage</a> means the machines are filling a gap, not creating a glut.</p><p>The honest headline was never “AI replaces AP.” It’s “AI replaces the part of AP that nobody actually enjoyed, and hands the interesting, judgment-heavy part back to the humans, who happen to be in short supply.” That’s not a threat to plan around. It’s a career that’s finally shedding its worst tasks.</p><p><em>Blackbee AI is an agentic Intake-to-Pay platform: the decision and control layer that sits above your ERP, automating the routine so finance teams can focus on judgment.</em><a href="https://blackbeeai.com/"><em> blackbeeai.com</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=75523c511e7e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ChatGPT vs. a System: The Honest Version of the AP Automation Debate]]></title>
            <link>https://medium.com/@blackbeeai/chatgpt-vs-a-system-the-honest-version-of-the-ap-automation-debate-66419981e9c6?source=rss-4cb0b92797f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/66419981e9c6</guid>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[ap-automation]]></category>
            <dc:creator><![CDATA[Blackbee AI]]></dc:creator>
            <pubDate>Tue, 25 Aug 2026 13:35:59 GMT</pubDate>
            <atom:updated>2026-08-25T13:35:59.661Z</atom:updated>
            <content:encoded><![CDATA[<h4>ChatGPT is genuinely useful for accounts payable, and it deserves more credit than most software vendors give it. But useful isn’t the same as sufficient. Here’s where the line actually is.</h4><p>Almost every “ChatGPT vs. [our product]” article is written by the company selling the product, and you can tell, because ChatGPT always comes off looking like a toy. That’s not honest, and it’s not useful to the finance team actually trying to decide what to do.</p><p>So let’s do the honest version. ChatGPT is genuinely good at a lot of accounts payable work. If your team is using it to pull data off invoices, analyse spend, draft vendor emails, or reason through a messy exception, you are not doing anything wrong; you’re getting real value from a remarkably capable tool. The interesting question isn’t whether ChatGPT is good. It’s whether it’s the right <em>kind</em> of thing for what you need AP to do, and there’s a specific point where the answer flips. This is about finding that point.</p><h3>What ChatGPT actually is, in an AP context</h3><p>Start with a clear-eyed definition, because the whole argument rests on it.</p><p>ChatGPT is a general-purpose large language model. It was not designed for AP. It has no native connection to your ERP, no vendor master, no contract database, no payment history, and no approval workflow. When you use it for AP tasks, you are bringing the data to it (pasting invoice text, uploading a CSV, describing a vendor relationship) and asking it to reason across whatever you’ve handed it, in the moment.</p><p>That’s not a criticism. It’s a definition, and within that definition ChatGPT is impressive. It reads documents well, follows structured-output instructions precisely, spots patterns in a dataset, drafts communications, and flags anomalies, all from a single chat window with zero integration. The constraint is structural, and it’s the same constraint that applies to any closed, general-purpose model: the outputs can’t be fully audited or governed, and for regulated functions like finance, the <a href="https://medium.com/data-science-collective/open-vs-closed-llms-in-2025-strategic-tradeoffs-for-enterprise-ai-668af30bffa0">inability to control how and where data is processed is often the real limitation</a>, not the quality of the answer.</p><p>Which points to the actual comparison. It isn’t ChatGPT versus AP software. It’s a <em>general-purpose reasoning tool</em> versus a <em>purpose-built process system</em>. Those are different categories of things, and which one you need depends entirely on what you’re asking your AP operation to do, and how big it’s gotten.</p><h3>Where ChatGPT genuinely wins</h3><p>It would be dishonest to skip past the places where ChatGPT is simply the better choice, sometimes better than any dedicated system.</p><p><strong>Speed to value is unmatched.</strong> You can open a chat, upload an invoice, paste a prompt, and have clean, structured data in under a minute. No implementation, no contract, no IT ticket, no onboarding. For a team that needs to solve a specific problem <em>today</em>, that accessibility is real.</p><p><strong>The approach is portable.</strong> A good extraction or analysis prompt works regardless of your ERP, your industry, or your size. Purpose-built software has to be configured and integrated before it does anything; a prompt delivers the first time you run it correctly.</p><p><strong>It’s excellent at the unstructured and the one-off.</strong> Drafting an awkward vendor dispute, making sense of a weird invoice format, generating an ad-hoc summary from a messy export. This is exactly the sort of open-ended reasoning that dedicated AP software was never built for and never claimed to do.</p><p><strong>At low volume, it’s essentially free.</strong> Fractions of a cent per document via the API, nothing incremental for analysis on a paid plan. At the volumes where a dedicated system’s pricing would be a real budget line, ChatGPT barely registers.</p><p><strong>And it’s the best evaluation tool there is.</strong> Spending a few months manually running the processes you <em>think</em> you want to automate teaches you your real requirements: which fields matter, which exceptions eat the most time. Teams who do that make far better software decisions than teams who walk straight into vendor demos.</p><p>All of that is true, and none of it is grudging. The case for ChatGPT in AP is genuinely strong. It just has an expiry date.</p><h3>Where the ceilings are, and why they’re structural</h3><p>Here’s the part that matters, and the reason to trust it is that these limits aren’t Blackbee’s marketing or anyone else’s; they’re the well-documented structural properties of general-purpose LLMs. A 2026 catalogue of LLM constraints lists, among the core eight, <a href="https://www.promptquorum.com/prompt-engineering/ai-limitations-what-llms-cant-do">no persistent memory, no real-world action, and no ability to self-verify outputs</a>. Every one of those maps onto an AP wall.</p><p><strong>No memory across sessions → no continuity.</strong> Every chat starts from zero. ChatGPT doesn’t know what you paid this vendor last month or the exception it flagged last week, because it holds no persistent state. A dedicated AP system is a database as much as a workflow: every invoice, decision, and payment is stored and available as context for the next decision, so it gets smarter about your patterns over time. ChatGPT forgets everything when the tab closes.</p><p><strong>No real-world action → no closed loop.</strong> This is the big one at scale. ChatGPT produces outputs; it does not take actions. It can extract an invoice but not post it to your ERP; validate a match but not release the payment; flag an exception but not route it to the right approver. Every output needs a human to carry it into another system. At twenty invoices a month, fine. At five hundred, that human hand-off <em>is</em> the bottleneck, and it quietly defeats the point of using AI at all. (ChatGPT’s own scheduling and task features underline this: they’re <a href="https://suprmind.ai/hub/chatgpt/features/">explicitly not a workflow-automation platform</a>, and you outgrow them the moment you need conditional logic or error handling.)</p><p><strong>No self-verification → the hallucination tax.</strong> LLMs generate the statistically likely answer, not the verified one, and they don’t flag their own uncertainty. Hallucination is stubborn even in frontier models: one 2025 study found prompt-based mitigation cut GPT-4o’s hallucination rate from <a href="https://arxiv.org/pdf/2601.22290">53% to 23%, still far too high for unattended production use</a>. In AP, that means a human has to check the work, which is manageable at low volume and impossible at high volume. Enterprise guidance is blunt about it: <a href="https://intuitionlabs.ai/articles/chatgpt-enterprise-deployment-training-guide">never make an LLM the sole author of a critical output</a>. Keep a human in the loop.</p><p><strong>No native governance → no audit trail.</strong> When ChatGPT helps make an AP decision, the reasoning lives in a chat window. It isn’t timestamped in your system of record, linked to the invoice, or part of your controls documentation. That’s a compliance problem, not a convenience one: running LLMs in production has made <a href="https://www.getmaxim.ai/articles/llm-observability-best-practices-for-2025/">compliance-grade logging and trace-level visibility a baseline requirement</a>, precisely because the model itself provides none. When an auditor asks who approved an invoice and why, “it’s in a chat somewhere” is not an answer that survives the meeting.</p><p><strong>No systematic confidence or approval architecture → no policy enforcement.</strong> ChatGPT can draft an approval-request email; it can’t enforce that the approval happens, escalate when an approver goes quiet, or apply a tiered policy (over $25,000 to the CFO, under $5,000 to the department head) consistently across every invoice. It also <a href="https://www.lyzr.ai/playbook/100-reasons-not-to-use-chatgpt-for-enterprise/">doesn’t surface systematic confidence scores or built-in human-approval checkpoints</a>, which regulated finance workflows depend on. A workflow engine enforces policy. A reasoning tool produces text. The gap between a well-reasoned recommendation and an enforced policy at scale is the gap between a smart assistant and a system.</p><p>Notice the pattern: every one of ChatGPT’s strengths (instant, flexible, no integration, reasons in the moment) is the flip side of a limit. The very statelessness that makes it frictionless is what stops it building continuity. You can’t prompt your way out of an architectural property.</p><h3>The signals you’ve crossed the line</h3><p>There’s no universal threshold, but the signals are unusually clear. Any one of these is worth taking seriously; two or three together is the answer.</p><ul><li><strong>You’re past ~200 invoices a month.</strong> Below that, the manual steps between ChatGPT’s output and your ERP are manageable. Above it, they pile into a time cost that usually exceeds what the prompts save.</li><li><strong>Your exception rate is over 10%.</strong> When more than one invoice in ten needs a human, the exceptions outpace the people running the prompts. You need exceptions managed as a workflow, not reasoned about one at a time.</li><li><strong>Your CFO wants AP visibility more often than monthly.</strong> A monthly analysis session isn’t infrastructure for weekly answers on commitments, overdue invoices, and cash exposure. That needs continuous monitoring.</li><li><strong>You’ve had an audit finding.</strong> The clearest signal of all: it means your current process isn’t producing a control environment an external reviewer accepts, and no prompt fixes that.</li><li><strong>You’ve had one duplicate payment or one fraud incident. Once.</strong> A single one typically costs more than a year of a real system. The moment it happens, the risk calculus changes.</li><li><strong>You’re spending 2+ hours a week running AP prompts.</strong> At that point the manual-AI workflow has crossed into being less efficient than a system that runs in the background.</li></ul><h3>Moving on doesn’t mean starting over</h3><p>The good news, and it’s genuinely good: outgrowing ChatGPT is additive, not a teardown.</p><p>The prompts you built have value beyond the outputs: they’ve taught you exactly what your AP operation needs, which is the single most useful input to any system decision. The sensible transition runs both in parallel for 60 to 90 days: the system takes the structured, high-volume, policy-governed work, while ChatGPT keeps doing the messy, unstructured, one-off tasks it’s genuinely best at. Over time, the structured workflows migrate to the system; ChatGPT stays useful for the open-ended reasoning no AP software is designed for, and there’s always some of that. This is also just good practice: starting tight, proving accuracy against your existing process before handing over anything consequential, is how you deploy any automation responsibly.</p><h3>What “the system” end actually looks like</h3><p>“Purpose-built” covers a wide range, and it’s worth being precise, because the category isn’t monolithic.</p><p>At the simpler end, it means invoice capture, three-way matching, approval routing, and ERP posting: the core mechanics of processing invoices without manual data entry. For many mid-market teams, that alone is a big step up from a manual process, and it addresses most of the ceilings above: persistent data, a workflow engine, an audit trail, the ability to <em>act</em> rather than just recommend.</p><p>At the more sophisticated end sits a different idea entirely: systems that don’t just process invoices faster but govern the decisions that <em>create</em> invoices in the first place. This is an agentic, intake-to-pay approach that starts at the point of spend intent and runs through contract validation, vendor risk, approval, and payment as one connected chain. The distinction that matters when you evaluate anything in this space isn’t the feature list; it’s whether the platform governs spend <em>upstream</em> of the invoice or only processes invoices after they land. That single question separates tools that process AP from systems that govern it, and it’s where the durable value sits.</p><h3>The honest verdict</h3><p>So, ChatGPT or a purpose-built system? For a real and useful stretch of a finance team’s life: ChatGPT. Use it. Learn from it. Let it teach you your requirements and save you hours you’d otherwise spend keying data by hand. Dismissing it is the mistake most vendors make, and most finance teams see through.</p><p>But keep an eye on the line. The day your volume climbs, your exception queue backs up, your CFO wants real-time answers, or you have one incident you never want to repeat, that’s the day a reasoning <em>tool</em> needs to become a <em>system</em>. The skill isn’t picking a side in the abstract. It’s recognising, honestly, which side of that line your operation is on today, and noticing when it moves.</p><p><em>Written for finance teams weighing general-purpose AI against purpose-built systems for accounts payable. Limitations and figures are drawn from published research and industry sources, linked inline. Disclosure: this is published by Blackbee AI, which builds in the agentic intake-to-pay category described near the end; the analysis applies to any tool you evaluate, and the case for ChatGPT above is meant sincerely.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=66419981e9c6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Autonomy Spectrum: How Much Should You Let Finance AI Decide on Its Own?]]></title>
            <link>https://medium.com/@blackbeeai/the-autonomy-spectrum-how-much-should-you-let-finance-ai-decide-on-its-own-d7b024c155e7?source=rss-4cb0b92797f2------2</link>
            <guid isPermaLink="false">https://medium.com/p/d7b024c155e7</guid>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Blackbee AI]]></dc:creator>
            <pubDate>Mon, 17 Aug 2026 12:42:21 GMT</pubDate>
            <atom:updated>2026-08-17T12:42:21.883Z</atom:updated>
            <content:encoded><![CDATA[<h4>The question is almost always asked as a yes-or-no question. That’s the mistake. Autonomy isn’t a switch you flip; it’s a dial you set per decision, and setting it well is the entire discipline.</h4><p>Ask a finance leader whether they’d let an AI approve payments on its own, and you’ll usually get one of two answers, both delivered quickly. Some version of “absolutely not, a human signs off on everything.” Or some version of “of course, that’s the whole point of buying it.”</p><p>Both answers are wrong, because the question is wrong. “Should you let finance AI decide on its own?” treats autonomy as a single switch, on or off, trust or don’t. But autonomy was never binary. It’s a spectrum, and this is the part most teams miss: the right position on that spectrum isn’t a global setting for your whole system. It’s a property of each individual decision, set by how confident the AI is, how much money is at stake, how risky the pattern looks, and what your policy says.</p><p>Get that idea, and the scary question dissolves into a manageable one. You’re not deciding whether to trust the AI. You’re deciding, for each type of decision, how much oversight it needs, and building a system that applies the right level automatically. This piece lays out that spectrum, why both extremes fail, and why confidence-scoring paired with risk-based routing is the responsible middle where the value and the safety both live.</p><h3>The spectrum, precisely</h3><p>The research literature has settled on a useful vocabulary for degrees of oversight, and it’s worth being precise because the terms get used interchangeably in ways that cause real confusion.</p><p><strong>Human-in-the-loop (HITL):</strong> the AI pauses and requires explicit human approval before it takes a consequential action. The human is a checkpoint in the execution path.</p><p><strong>Human-on-the-loop (HOTL):</strong> the AI acts autonomously while a human monitors the flow, watching for anomalies and able to intervene, but not gating each action.</p><p><strong>Human-out-of-the-loop:</strong> the AI executes end-to-end with no human in the path at all.</p><p><strong>Human-in-command:</strong> a human retains oversight of the overall system, the authority to decide when and how the AI operates at all, even when individual decisions are autonomous.</p><p>Here’s the insight that reframes everything, articulated well in a<a href="https://www.strata.io/blog/agentic-identity/practicing-the-human-in-the-loop/"> 2026 guide to AI oversight</a>: the oversight level is a property of the <em>decision</em>, not the system. All of these models can, and should, coexist inside a single workflow, assigned dynamically by risk, context, and policy. The guide’s example is an airline rebooking agent: routine rebookings run fully autonomously, a first-class international itinerary with a fare-class complication gets routed to a human, and a supervisor monitors the whole flow for cost anomalies. Three oversight levels, one workflow, chosen per decision.</p><p>Finance is exactly the same. A $40 recurring SaaS invoice matched at 99% confidence, and a first-time vendor requesting a $90,000 payment to a just-changed bank account are not the same decision, and no sane design would apply the same oversight to both. The skill isn’t picking one autonomy level. It’s building a system that assigns the right one, transaction by transaction.</p><h3>A necessary aside: autonomy is not explainability</h3><p>These two get conflated constantly, and keeping them separate sharpens the whole discussion.</p><p><strong>Explainability</strong> is about <em>why</em>: can you see the reasoning behind a decision after it’s made? <strong>Autonomy</strong> is about <em>who</em>: did a human sit in the decision path, or did the AI act alone?</p><p>They’re different axes. You can have a fully autonomous decision that’s perfectly explainable: the AI paid the invoice on its own and can show you exactly which PO, contract, and receipt it matched against. You can also have a human-reviewed decision that’s a total black box: a person clicked approve on a recommendation they didn’t understand. Responsible finance AI needs both properties, but solving one doesn’t solve the other. This piece is about the autonomy axis; explainability is its essential companion, not its substitute. Watch for vendors who answer an autonomy question with an explainability answer, or vice versa; it usually means one of the two is weak.</p><h3>Why full autonomy fails (and increasingly, breaks the law)</h3><p>Start with the tempting extreme: let the agents run. The trouble is that finance is precisely the domain where unsupervised autonomy is most dangerous, and regulators have noticed.</p><p>The risk compounds with what the agent can <em>do</em>. As one analysis puts it, an agent that reads a document is one category of risk; an agent that reads a document and then<a href="https://www.elementum.ai/blog/human-in-the-loop-agentic-ai"> executes a transaction</a> against your ERP is an entirely different one. Tool access, the ability to move money, is what turns a helpful model into a material exposure. The European Systemic Risk Board has warned that autonomous agents can execute financial transactions independently,<a href="https://www.elementum.ai/blog/human-in-the-loop-agentic-ai"> compressing timelines and increasing the speed</a> of potential fraud or money laundering. Speed, the whole selling point, is also the danger: an autonomous system can make a thousand wrong or fraudulent payments before a human notices one.</p><p>And this is no longer only a prudence argument. Regulation is hardening human oversight from best practice into legal requirement. The EU AI Act mandates human oversight, audit trails, and explainability for high-risk systems; the<a href="https://www.trantorinc.com/blog/human-in-the-loop-vs-fully-autonomous-ai-agents"> Colorado AI Act took effect in February 2026</a>, California’s SB-833 adds requirements from mid-2026, and in financial services, automated decisions already require<a href="https://www.trantorinc.com/blog/human-in-the-loop-vs-fully-autonomous-ai-agents"> </a>complete audit trails and <a href="https://www.trantorinc.com/blog/human-in-the-loop-vs-fully-autonomous-ai-agents">explainability under SOX, GLBA, and anti-money-laundering rules</a>. Fully autonomous agents acting on consequential financial decisions aren’t just risky; for a growing set of cases, they’re non-compliant.</p><h3>Why full human review fails too</h3><p>So swing to the other extreme: a human reviews everything. Safe, surely?</p><p>Safe, and pointless. If a person has to touch every invoice, you’ve bought expensive AI to keep doing manual work, and you haven’t even solved the problem you had. Most teams are already drowning:<a href="https://www.mindsprint.com/resources/blogs/ai-accounts-payable-2026"> 66% of finance teams </a>still manually enter invoice data into their ERP, and as volume and vendor diversity grow, the exception queue grows faster than the team can clear it. Routing everything to a human doesn’t add control; it just recreates the bottleneck you were trying to escape, now with a software licence attached.</p><p>There’s a subtler failure too. When a human reviews <em>everything</em>, they review nothing well, attention flattens across a thousand identical approvals, and the one transaction that actually mattered slips through in the same half-second click as the 999 that didn’t. As one AP guide puts it, the goal isn’t to review every invoice; it’s to<a href="https://pinnasys.com/blog/automate-finance-and-accounts-payable-workflows-with-ai"> review the right ones</a>. Undifferentiated review is a worse control than targeted review, not a stricter one.</p><p>Both extremes, then, fail, one on risk and compliance, the other on value and, quietly, on control itself. The answer is in between, and it has a specific shape.</p><h3>The responsible middle: confidence-scoring plus risk-based routing</h3><p>The design that actually works has two moving parts, and they have to work together.</p><p><strong>Part one: confidence-scoring.</strong> Every decision the AI makes carries a confidence score, and that score determines how much oversight it gets. The pattern is now well established across finance AI. In a typical tiered setup, transactions matched at<a href="https://chatfin.ai/blog/finance-ai-agents-the-complete-guide-to-autonomous-ai-for-accounting-budgeting-reconciliation/"> very high confidence auto-process, a middle band gets a quick human review, and low-confidence cases are escalated with full context</a> for investigation; for example, auto-clear above 98%, quick-review between 80 and 98%, escalate below 80%. The AI, in effect, knows what it doesn’t know, and asks for help exactly when it’s unsure. Confidence-scoring every field is the foundation; it’s the core job of an<a href="https://claude.ai/products/invoice-processing-agent"> invoice-processing agent</a> worth the name, and it’s what lets routine, unambiguous work run untouched while genuine ambiguity finds a human.</p><p><strong>Part two: risk-based routing, because confidence alone is not enough.</strong> This is the part weaker systems miss, and it’s the crux of the whole framework. Confidence measures whether the AI read the invoice <em>correctly</em>. It says nothing about whether the invoice is <em>safe to pay</em>. Those are different questions, and the gap between them is where fraud lives.</p><p>The illustration is worth holding onto. Picture an invoice whose data the AI extracted at 99% confidence, but the supplier’s bank details differ from the master record. High confidence, high risk. A confidence-only system waves it through; a well-designed one<a href="https://getorcha.com/en/produkt/kreditoren"> flags it for mandatory review a</a>nd routes it to a controller regardless of the confidence score, precisely because a changed bank account is the signature of payment-redirect fraud. The oversight was triggered by <em>risk</em>, not uncertainty.</p><p>So the responsible middle layers risk-based overrides on top of confidence:</p><ul><li><strong>Amount limits:</strong> above a set value, always human-in-the-loop, regardless of confidence. A perfectly read $250,000 payment still gets eyes on it.</li><li><strong>Anomaly triggers:</strong> new bank details, an unusual price spike, a possible duplicate or split invoice, <a href="https://getorcha.com/en/produkt/kreditoren">mandatory review </a>even at high confidence.</li><li><strong>Hard policy rules:</strong> non-negotiable guardrails, such as never exceeding a PO amount without approval, that no confidence level can override.</li><li><strong>First-time-vendor gates:</strong> unknown counterparties get more scrutiny than trusted ones,<a href="https://pinnasys.com/blog/automate-finance-and-accounts-payable-workflows-with-ai"> independent of how clean the data looks</a>.</li></ul><p>This is why routing has to be driven by <em>risk and policy, not just dollar amount</em>. A low-value payment to a high-risk, just-onboarded vendor may deserve far more scrutiny than a large, routine payment to a supplier you’ve paid monthly for five years. Amount-based routing can’t tell the difference; risk-based routing can, which is exactly what an<a href="https://claude.ai/products/approval-orchestration-agent"> approval-orchestration layer</a> is for, and why it needs live signals like continuous<a href="https://claude.ai/blog/ai-vendor-risk-scoring"> vendor risk scoring</a> feeding into the routing decision.</p><p>Put the two parts together, and you get the behaviour you actually want: the routine majority runs straight through, the risky minority gets a human, and, crucially, the exceptions arrive pre-classified by type and confidence, a triaged queue of the decisions that genuinely need judgment rather than<a href="https://engini.ai/blog/ai-tools-accounts-receivable-automation-exceptions"> an undifferentiated pile </a>where every exception looks the same.</p><p>Done well, agents can autonomously resolve<a href="https://peakflo.co/blog/ai-agents-exception-management-accounts-payable-guide"> 70 to 80% of AP exceptions</a> while the genuinely consequential ones are escalated with full context. That’s not full autonomy, and it’s not full review. It’s the right oversight on the right decision, every time.</p><h3>Autonomy should be earned, not assumed</h3><p>One more principle separates responsible deployments from reckless ones: autonomy is a dial you turn up as evidence accumulates, not a setting you commit to on day one.</p><p>The disciplined path is to start tight and prove it. Run the AI in<a href="https://chatfin.ai/blog/finance-ai-agents-the-complete-guide-to-autonomous-ai-for-accounting-budgeting-reconciliation/"> </a>parallel with your existing process for <a href="https://chatfin.ai/blog/finance-ai-agents-the-complete-guide-to-autonomous-ai-for-accounting-budgeting-reconciliation/">60 to 90 days</a>, comparing its decisions against human ones across a full cycle, before it touches anything live. Begin autonomy with the safest slice, high-confidence, PO-backed invoices from top suppliers, and widen coverage only as measured accuracy justifies it. Raise the auto-processing thresholds based on observed performance, not optimism. Every human correction feeds back to improve the model, so the straight-through rate climbs over time on evidence rather than faith.</p><p>This reframes the whole autonomy question from a leap into a ratchet. You’re never betting the close on an unproven system; you’re granting it more independence in proportion to the trust it has demonstrably earned. And you keep a human-in-command over the whole arrangement, the authority to widen, narrow, or halt the AI’s remit, even as individual decisions become autonomous.</p><h3>A framework you can actually use</h3><p>If you’re evaluating any finance AI, or auditing one you already run, the autonomy question reduces to six checks:</p><ol><li><strong>Does every decision carry a confidence score, and can you set the thresholds?</strong> <br>No score, no principled way to decide what’s autonomous.</li><li><strong>Can risk, amount, and policy override confidence?</strong> <br>A 99%-confident payment to a changed bank account must still be able to be stopped. If confidence is the only gate, the system is unsafe.</li><li><strong>Is routing driven by risk and policy, or just dollar amount?</strong> <br>Amount-only routing misses the risk that actually matters.</li><li><strong>Can you start in parallel/shadow mode and expand autonomy with evidence?</strong> <br>Earned autonomy beats assumed autonomy.</li><li><strong>Is every autonomous decision explainable and logged for audit?</strong> <br>The autonomy axis and the explainability axis both have to be covered.</li><li><strong>Is there always a human-in-command?</strong> <br>Someone must own the authority to change the system’s remit, regardless of how autonomous the transactions are.</li></ol><p>A system that passes all six isn’t “autonomous” or “supervised.” It’s <em>calibrated</em>, which is the only setting that’s both safe and worth paying for.</p><h3>So, how much should you let finance AI decide on its own?</h3><p>The honest answer is the one the binary framing can’t accommodate: <em>it depends on the decision, and a well-built system should make that determination for you, transparently, on every transaction.</em></p><p>Let it decide the thousand routine, high-confidence, low-risk calls on its own; that’s where the hours go, and the error rate is lowest. Make it stop and ask on the consequential, ambiguous, or risky ones; that’s where judgment and accountability belong. The line between the two isn’t drawn once by a nervous committee; it’s drawn continuously, per decision, by confidence-scoring and risk-based routing working together.</p><p>Which points to the real measure of trustworthy finance AI. It isn’t how much the system can do on its own. It’s how reliably it knows when <em>not</em> to, and confidence-scoring paired with risk-aware routing is how a system earns, decision by decision, the autonomy it takes.</p><p><em>Written for finance leaders weighing how much decision authority to delegate to AI. Frameworks and figures are drawn from published research and industry sources, linked inline.</em></p><p><em>Disclosure: This is published by </em><a href="https://www.blackbeeai.com/"><em>Blackbee AI</em></a><em>, whose Intake-to-Pay platform is built around the confidence-scoring and risk-based routing described here; the framework stands on its own evidence and applies to any finance AI you evaluate, this one included.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d7b024c155e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>