<?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 Precision AI OS on Medium]]></title>
        <description><![CDATA[Stories by Precision AI OS on Medium]]></description>
        <link>https://medium.com/@precision_ai_os?source=rss-391a30238fce------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*jMU7QgXShyrVoSQyKdfSSg.png</url>
            <title>Stories by Precision AI OS on Medium</title>
            <link>https://medium.com/@precision_ai_os?source=rss-391a30238fce------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 09:31:26 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@precision_ai_os/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[How much does an AI consultant cost? Price the hours you’re losing first]]></title>
            <link>https://medium.com/@precision_ai_os/how-much-does-an-ai-consultant-cost-price-the-hours-youre-losing-first-cdc3f96ffbb2?source=rss-391a30238fce------2</link>
            <guid isPermaLink="false">https://medium.com/p/cdc3f96ffbb2</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[consulting]]></category>
            <category><![CDATA[small-business]]></category>
            <category><![CDATA[accounting]]></category>
            <dc:creator><![CDATA[Precision AI OS]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 20:29:34 GMT</pubDate>
            <atom:updated>2026-11-18T20:29:34.482Z</atom:updated>
            <content:encoded><![CDATA[<p>When owners of small accounting, law and insurance firms ask how much an AI consultant costs, they usually mean the invoice. That’s the smaller number. The bigger one is already on your books: the hours your staff spend every week retyping data between systems, chasing clients for documents, and writing the same few emails with different names in them. And there’s a second cost people don’t mention until it’s happened to them, which is paying for a strategy deck and ending up with nothing running.</p><p>So before you compare quotes, it helps to know how consultants price this work, what pushes the price up, and what to get in writing so you’re paying for something that runs.</p><h4>How consultants charge</h4><p>There are four common setups, and most firms will run into some mix of them.</p><p>Hourly is the simplest and the hardest to budget. It suits a short assessment, or a firm that wants an expert on call while its own people do the work. The risk is that discovery stretches, because there’s no reason for it not to.</p><p>A fixed-scope project is the most common for a first engagement. The consultant agrees to deliver specific things for a set price. This is only as good as the scope document. “Implement AI for client communications” is not a scope. “Set up drafting of missing-document follow-ups from the request list in our practice management system, with preparer approval before sending” is.</p><p>A monthly retainer usually covers ongoing support, tuning and new requests after the first build. It makes sense once something is running and you want it maintained. It makes less sense as the first thing you sign.</p><p>Some consultants now price per workflow built: a set fee for each job automated, sometimes with a monthly charge per workflow after that. This ties what you pay to what you get, which helps, as long as the contract defines “workflow.”</p><h4>What pushes the price up</h4><p>The same handful of things drive the cost almost every time.</p><p>Messy data is the big one. If client files are named differently by every staff member, or the practice management system’s request lists are out of date, someone has to clean that up before anything can run on top of it. That someone is either the consultant, billing for it, or your staff, losing hours to it.</p><p>The number of systems involved matters too. A job that touches only email is simpler than one that reads the portal, checks the practice management system and writes back to both. Each connection is setup, testing and something that can break later.</p><p>How many jobs you automate is obvious, but firms still underestimate it. Three jobs are not three times one job. They interact, and the testing grows.</p><p>Then there’s who maintains it after launch. Software updates, vendor changes, a new staff member who does things differently. If the consultant built it and you can’t maintain it yourselves, you’re signing up for ongoing cost whether the quote says so or not.</p><p>Confidentiality work adds time as well. Lawyers have professional duties around client information, and US tax preparers face section 7216 limits on using return information without consent, so a careful consultant will spend real hours on where data goes and who can see it. If nobody asks about that, it’s a gap in the quote, not a saving.</p><p>Training is the last one. Staff need to learn how to review what the system drafts and when to override it. Budget their time for that, not just the consultant’s.</p><h4>The costs outside the quote</h4><p>Even a clear quote leaves things out. Your staff’s time in discovery calls and testing is real time away from client work, often during the weeks you can least spare it. Review time doesn’t go to zero once things are running; someone still checks the drafts, and that should be planned. Software subscriptions and AI usage fees often sit outside the consultant’s price, billed to you directly. And there’s the exit question: if you stop paying the consultant, does what they built keep running, and do you own the setup, the instructions and your data?</p><h4>What to get in writing</h4><p>Before you sign, get a document that names the specific jobs being automated, one by one. For each job, ask for an acceptance test, meaning how you’ll both know it works, ideally tied to a real week of your firm’s work. It should say who owns the setup, the configuration and the data, and what happens to all three if the engagement ends. It should state who reviews output before it reaches a client, because that answer protects you in a way no feature does. And it should spell out the exit terms: notice period, handover, and what you keep.</p><p>If a consultant won’t put that in writing, ask why before you sign.</p><h4>A simple test</h4><p>Here’s a way to make the decision concrete. Pick one job your team repeats, the kind that takes a few minutes each time and happens many times a week. Write down how many hours it took last month. Put a rough value on those hours. Now look at the quote for automating that one job, plus the outside costs above.</p><p>If the numbers make sense for one job, you have a real starting point. If you can’t name the job, you’re not ready to hire anyone yet, and that’s worth knowing before the first invoice rather than after.</p><p>Precision AI OS builds an Agent OS for professional service firms, where agents prepare the work and anything that goes to a client waits for a person’s approval by default.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cdc3f96ffbb2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Will AI replace accountants? Here’s the part of the job it takes first]]></title>
            <link>https://medium.com/@precision_ai_os/will-ai-replace-accountants-heres-the-part-of-the-job-it-takes-first-abae6e20bb50?source=rss-391a30238fce------2</link>
            <guid isPermaLink="false">https://medium.com/p/abae6e20bb50</guid>
            <category><![CDATA[taxes]]></category>
            <category><![CDATA[cpa]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[accounting]]></category>
            <category><![CDATA[small-business]]></category>
            <dc:creator><![CDATA[Precision AI OS]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 17:43:38 GMT</pubDate>
            <atom:updated>2026-11-18T17:43:38.797Z</atom:updated>
            <content:encoded><![CDATA[<p>Every January the same work eats a tax firm’s week. A return cannot move until a client sends the last statement, the missing K-1, the payroll report they swear they uploaded. Someone sends the reminder, checks the portal, sends another reminder, calls the bookkeeper, updates the tracking sheet. Nobody bills those hours. They get logged, show up again in April, and get written off. So here is the question a firm owner actually asks: will AI replace accountants? The honest answer starts with the chasing, not the judgment.</p><p>The cost shows up in three places. Hours go to work that never produces a billable line. Files arrive late, so the work compresses into a shorter window and the review lands in a pile. And the write-offs compound quietly, because time we never billed is time we never collected. None of that shows up as a dramatic loss. It shows up as a busy season that feels harder than the numbers justify.</p><p>By March the review backlog runs the office. Partners review at night. Staff wait on notes. Clients who went quiet in February call in April wanting to know where their return is. The return is sitting in a queue behind work that could not start until a document arrived. Every firm owner knows this and plans around it every year.</p><p>Clients go quiet for ordinary reasons. They are busy, they lost the email, they are waiting on their own bank. The silence is not personal, and it still stops the work. Collections follow the same path. A return that finishes late gets billed late, the invoice arrives after the value has faded, and the follow-up calls begin. None of that is an AI problem. It is a capacity problem with a paper trail.</p><h3>The work around the work</h3><p>Strip the job down and a surprising share of it is preparation. Chasing missing documents. Sorting a shared inbox. Writing the first version of a client email. Matching transactions. Pulling prior-year items into the current file. Answering the same question about a portal password for the fourth time. None of it requires a license. All of it has to be right before the real work can start, and almost all of it ends in someone else reviewing it.</p><p>That shape of task is where AI fits first. The input is defined. The output is defined. A person checks the result before it goes anywhere. Bookkeeping shows the pattern well. Matching a bank feed to the right account is a sorting problem, and sorting problems are the kind software handles without much drama. The judgment sits in the exceptions, the split, the accrual, the transaction that looks like a duplicate but is not. The exceptions are where a person earns their keep.</p><p>The same is true of client email. A first draft that says we still need the December statement is not a hard thing to write. It is hard to write again and again in the same week while also reviewing returns. The partner still decides what the client is told. The draft just removes the blank page.</p><h3>A test you can run this week</h3><p>Before anyone buys a tool, list the hours each person spends on work that ends in someone else reviewing it. Be specific about the categories:</p><ul><li>Reminders, portal checks, and follow-up on clients who have gone quiet</li><li>Inbox sorting, document naming, and pulling prior-year items</li><li>First drafts, transaction matching, and cleanup before a return reaches review</li></ul><p>Add it up. That total is the pool of hours you could hand off. If the number is large, the next question is who is doing the work. In a lot of firms it is the most experienced person, because they know the client, they know the history, and handing it off feels slower than doing it. That is a reasonable choice in February and an expensive one in June, when the unbilled time is already written off and the collection calls are still ahead.</p><p>The timesheet is the most revealing document in the firm and the least examined. Look for the lines that repeat. Admin. Follow up. Portal. Misc. A senior person’s week fills with entries like those, and none of them end with a client paying for judgment. They end with a review, a nod, and a write-off in April. If you want to know what to hand off first, the timesheet will tell you before any software demo will.</p><p>This is where the replacement worry points the wrong way. The firm that gets hurt is rarely the one whose software prepares a return. It is the one that keeps paying senior rates for document retrieval while another firm clears its review backlog earlier and takes on the next client. AI doing the job is not the danger. The danger is quieter: we keep doing the prep ourselves and call it client service.</p><h3>What stays human</h3><p>Signing returns. Advice. Deciding what a client is told and how. Anything that moves money. Taking on a new client. Judgment with consequences stays with a person, and so does the signature that carries the responsibility. A machine can prepare a file. It cannot sit across from a client and decide whether they should take the deduction, sell the property, or change how they pay themselves. Those conversations are the job. The prep is the work around it.</p><p>Two surveys covered on November 13, 2026, are worth reading together. Accounting Today reported on a Financial Cents survey of 486 accounting and bookkeeping professionals, mostly at firms under 30 people. Ninety-five percent said their firms had adopted AI in some form. Only eleven percent said it was embedded in daily work across much of the firm. Buying a tool is easy. Changing how work moves through the firm is not. The same day, an Accounting Today opinion piece cited a CPA career satisfaction survey where 56 percent of respondents were very concerned about the AI learning curve and 31 percent said their job scope had expanded. Both reactions make sense. Our read: a firm that adds several tools and no review process has added work, not removed it. The learning curve lands hardest on the people already carrying busy season.</p><p>The rule we’d suggest is boring. Pick one prep task that ends in review. Measure the hours for a month. Give that task to an agent, keep the review with a person, and check the number again. If the hours fall and the review holds, add another task. If the review slips, stop. That is what embedded adoption looks like: one task at a time, with a person at the end of the line.</p><p>Precision AI OS builds an Agent OS for CPA and professional service firms where agents prepare the work and anything that goes to a client waits for a person’s approval by default. That is the arrangement we would defend.</p><p>https://getprecision.ai/?utm_source=medium&amp;utm_medium=social&amp;utm_campaign=will-ai-replace-accountants</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=abae6e20bb50" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Will AI replace lawyers? Look at the hours before the headlines]]></title>
            <link>https://medium.com/@precision_ai_os/will-ai-replace-lawyers-look-at-the-hours-before-the-headlines-04b6a615162c?source=rss-391a30238fce------2</link>
            <guid isPermaLink="false">https://medium.com/p/04b6a615162c</guid>
            <dc:creator><![CDATA[Precision AI OS]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 20:29:45 GMT</pubDate>
            <atom:updated>2026-11-17T20:29:45.592Z</atom:updated>
            <content:encoded><![CDATA[<p>The phone rings at 6 PM. Nobody picks up. The caller wants a will, or a contract reviewed, or an answer about a letter from a landlord. By morning the matter is gone, and the only trace is a sticky note on someone’s desk. That lost call is the first cost, and it never appears on a time sheet. When people ask will AI replace lawyers, the honest starting point is the work that walks away because nobody was there to catch it.</p><p>The second cost does appear on a time sheet, though often at a discount. An associate spends an evening pulling a first draft of a routine letter from a template, checking a conflict list by hand, and retyping the same intake details into three places. The work is real. The client may not pay for all of it. Some of it gets written off because the firm can’t justify the hours, or the client pushes back, or nobody wants to admit how long the simple parts took.</p><p>Two costs, same root. A great deal of what a small firm does around the legal judgment is preparation. Preparation can be done differently.</p><h3>Where the hours go</h3><p>Start with intake. A prospective client fills in a form at night, and the form sits in an inbox until someone has a spare hour. First calls wait behind other first calls. Conflict checks get built by hand from a list that lives in a document nobody quite trusts. Then come the first drafts: the routine letter, the standard demand, the engagement terms, the cover note to a court. Then document review, where a junior reads a stack to find the three pages that matter. Then the chasing, which is its own job: signatures, identification, the document the client promised on Tuesday. Then the time entries, written days later from memory, which is how billable time quietly shrinks before it ever reaches an invoice.</p><p>None of that is the part of practicing law that keeps a lawyer up at night. It’s the preparation around it.</p><p>The pattern repeats across the week. A partner reads a document on a phone at night and dictates a note that someone transcribes the next afternoon. Two people handle the same file because nobody can see where it stopped. A retainer letter waits for a signature that waits for a client who was never sent a reminder. Nobody owns these problems, so they don’t get fixed.</p><h3>What the opinion asks</h3><p>In July 2024 the American Bar Association issued Formal Opinion 512 on lawyers’ use of generative AI tools. It is worth reading in full, but four things in it matter to a small firm.</p><ul><li>Understand the tool well enough to use it competently.</li><li>Protect client confidentiality, and get informed client consent before putting confidential information into some self-learning tools.</li><li>Supervise the work the tool produces.</li><li>Charge a reasonable fee, meaning do not bill for time saved as if it had been spent, and do not charge a client for the time it takes to learn a tool you would use generally.</li></ul><p>The fee point deserves a longer look. If a tool cuts an afternoon of assembly to twenty minutes, a lawyer billing hourly bills the twenty minutes plus the time spent reviewing the result, not the afternoon it used to take. An entry padded to match old habits is hard to defend if a client ever asks. The opinion also says that, in most cases, a lawyer can’t charge a client for learning how to use a tool. That’s the firm’s cost of doing business.</p><p>What the opinion doesn’t do is hand the work to the software. Lawyers remain responsible for the work product. The judgment, the advice, the court appearance and the signature that commits the client stay with the lawyer. So does supervision of whatever the tool produced.</p><p>That is the useful split. The tool can prepare. A person decides, signs, and answers for it.</p><p>In practice, supervision means someone with a license reads the thing before it leaves. It means the file shows who checked what, when, and what they changed. It means a client question comes back to a person who can answer for the answer. A firm that cannot describe that loop in a sentence is not ready to put a tool anywhere near a client.</p><h3>The test for this week</h3><p>Here’s a test that costs nothing. Take last week’s time entries and list the hours spent on work that ended with a lawyer reviewing it. Not the advice. Not the strategy call. The drafting, the checking, the chasing, the retyping, the summarizing, the assembling of a file so someone else could make a decision.</p><p>Add up those hours. Then ask two questions. How much of that work could have been prepared before anyone arrived in the morning? Of the matters that walked away, how many walked because the first reply came a day late?</p><p>Compare the total with the hours you actually billed last week. For many small firms the gap is bigger than expected, and much of it sits in work that software can prepare and a person can check. Sort the list into two piles: work that needs a person to decide, and work that needs a person to check. If the second pile is the bigger one, that’s where to start. Checking a good draft is usually faster than writing one from a blank page, though a bad draft can take longer to fix, so judge the tool on the drafts it actually produces. A lawyer who reviews a draft in ten minutes hasn’t given up the judgment. They have moved it to the front of the process, where it belongs.</p><p>What does not change. The client still needs someone to tell them what to do. Someone to sit with them and say the hard thing. Someone whose name is on the letter. A tool can’t carry that duty. Under the opinion, the lawyer stays responsible and has to supervise. That’s the job, not a formality.</p><p>So will AI replace lawyers? Not the ones who organize their firms around what only a person can do. The ones at risk are the ones who keep paying for preparation at the old price, and who let a call at 6 PM ring out.</p><p>Start with the list. It takes an hour, and it’s free.</p><p>Precision AI OS builds an Agent OS for professional service firms. Agents prepare the work, and anything that goes to a client waits for a person’s approval by default.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=04b6a615162c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI agents vs. RPA for accounting firms]]></title>
            <link>https://medium.com/@precision_ai_os/ai-agents-vs-rpa-for-accounting-firms-13c4a79e0bf0?source=rss-391a30238fce------2</link>
            <guid isPermaLink="false">https://medium.com/p/13c4a79e0bf0</guid>
            <dc:creator><![CDATA[Precision AI OS]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 17:24:17 GMT</pubDate>
            <atom:updated>2026-11-14T17:24:17.039Z</atom:updated>
            <content:encoded><![CDATA[<h4>Fixed rules still have a place. The question is which work actually follows them.</h4><p><em>I’m a co-founder of Precision AI OS. Nothing here is tax or legal advice.</em></p><p>A bank-feed rule that categorizes transactions from the same vendor the same way every month is automation. So is a macro that reformats a trial balance export. Robotic process automation, or RPA, can automate steps such as clicking through screens and moving data according to instructions written in advance. I would choose between fixed rules and an agent by looking at the job&#39;s inputs and the exceptions it needs to handle.</p><p>Rules-based automation follows fixed instructions and handles the cases it was programmed for. I&#39;d consider it when the input arrives in a consistent format and the output goes to a defined destination. I would also test what happens when the format changes. A scanned statement instead of a download might need a separate step. A document under a spouse&#39;s name could need an exception rule. The test should show whether an unfamiliar input is flagged for a person.</p><p>An agent can work toward a defined goal and choose steps and tools within the workflow. The firm sets it up to hand the file to a person when the work is ready for review or it reaches a limit. For document collection, one possible task is to compare what was received with what was requested and draft a follow-up for a staff accountant to read and release. I&#39;d test whether it identifies the missing items correctly before relying on that draft.</p><p>Because the agent chooses its next step, the workflow needs a defined stopping point and a handoff to a person. I would test that handoff by giving it an incomplete file and checking that it asks for help. I would also retest a rules-based workflow when its inputs or instructions change.</p><p>To sort the work, I&#39;d keep rules where the inputs are uniform and the cost of a quiet mistake is low, and I&#39;d look at agents where the job crosses the portal, the tax software and the inbox and the inputs vary from client to client. Either way, I&#39;d want a named reviewer and a named person who signs off before the first file runs. The reviewer also needs a standard to check against, so pick a few finished files the firm considers right and compare the drafts to those.</p><p>I&#39;d start with one workflow and test it against finished files while staff do the job the usual way. I would move to live work only after checking those test drafts, with a person reviewing each draft and handling each release. At day 30, decide whether the evidence supports a wider scope. If it doesn&#39;t, revise the workflow or stop the pilot. The test gives you examples to assess against your firm&#39;s standards.</p><p>I wouldn&#39;t retire the rules that work. I&#39;d keep the bank-feed rules and macros and test an agent on jobs those rules don&#39;t cover. The agent prepares the file, and a person still decides what the client receives and signs what gets filed.</p><p>AI agents for CPA and accounting firms and other professional service firms. Agents prepare the work. Sends wait for your team&#39;s approval by default. <a href="https://getprecision.ai/?utm_source=medium&amp;utm_medium=social&amp;utm_campaign=ai-agents-vs-rpa-accounting">https://getprecision.ai/?utm_source=medium&amp;utm_medium=social&amp;utm_campaign=ai-agents-vs-rpa-accounting</a></p><p><em>Steven de la Fe is a co-founder of Precision AI OS and a CPA with 20 years of tax and accounting experience, including seven years at PwC. For the past ten years he has helped build Taxfyle, growing it from $0 to $25 million in annual recurring revenue.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=13c4a79e0bf0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Anthropic's S-1 tells firm partners about reviewing AI agents]]></title>
            <link>https://medium.com/@precision_ai_os/what-anthropics-s-1-tells-firm-partners-about-reviewing-ai-agents-ca3a478da317?source=rss-391a30238fce------2</link>
            <guid isPermaLink="false">https://medium.com/p/ca3a478da317</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[professional-services]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[internal-control]]></category>
            <category><![CDATA[accounting]]></category>
            <dc:creator><![CDATA[Precision AI OS]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 16:40:42 GMT</pubDate>
            <atom:updated>2026-10-27T16:42:49.695Z</atom:updated>
            <content:encoded><![CDATA[<p>We&#39;d point a firm partner to one part of Anthropic&#39;s S-1 in particular. The filing went to the SEC on Monday, October 26, and it&#39;s a preliminary prospectus, so details can still change. The risk factors section, though, already reads like a briefing on what happens when software starts doing client work.</p><p>We should say where we stand. Precision builds AI agents for professional service firms and uses Claude to do it. We have no affiliation with Anthropic, and none of this is investment commentary.</p><p>Start with what the filing says agents can do. On page 47: &quot;agentic AI systems are capable of autonomously taking actions on behalf of users, interacting directly with external systems, executing code, making API calls, sending communications, or completing transactions with limited human oversight.&quot; For a firm, &quot;sending communications&quot; could mean the client email about a missing tax document, and &quot;completing transactions&quot; could mean the vendor payment run. These used to be the steps a staff member did by hand after a manager looked at the work.</p><p>The filing also warns about a quieter failure. On page 48 it says &quot;agentic systems may pursue stated objectives in ways that are technically consistent with their instructions but contrary to the user’s actual intent.&quot; Every partner has seen the human version of this, the staff accountant who did exactly what the email said and missed the point of the engagement. With an agent the same mistake can happen faster, and across a whole client list at once.</p><p>Then there&#39;s the review itself. The filing names &quot;organizational pressures within customer environments to maximize agent autonomy in order to achieve productivity gains.&quot; That will sound familiar to anyone who has worked a busy season. The reason for adopting agents is to save time, so the temptation is to approve faster, check less and let more go out on its own. In the same passage on page 48, the filing warns that reviewers facing that volume may not keep up.</p><p>Here&#39;s how we&#39;d turn that into a conversation with your team, or with whoever sells you the software.</p><p>First, list the actions that need approval before they take effect: client emails, payments, e-filings, changes to the general ledger, uploads to a client portal. Even the ones that can be reversed take work to clean up. Those are the steps where a person signs off, by name, every time, at least until the firm has evidence that a category is safe to loosen.</p><p>Second, ask what the reviewer sees when they approve. If the answer is a short summary and an approve button, the review is a formality. A useful approval screen shows the draft, the source documents it came from and anything the agent flagged as missing or conflicting. Reviewers do better when the context comes to them.</p><p>Third, ask what record is kept. The filing says on page 49 that &quot;maintaining comprehensive records of the actions, decisions, and reasoning of our agentic systems across complex, multi-step workflows is technically challenging&quot;. The firm still owns the engagement and the client relationship, and the professional judgment stays with its people. So the firm needs its own record: what the agent prepared, who approved it, what they changed and what went out. Don&#39;t count on rebuilding that from a vendor&#39;s logs after a client complaint.</p><p>Fourth, ask how the review load will be managed. If every routine reminder needs the same attention as a payment, people will stop reading both. Sort the work. Routine, low-risk items can be grouped for a quick pass. Anything unusual goes to a person with time to look at it.</p><p>Fifth, keep an eye on the rules. The filing says on page 69 that &quot;In the United States, the regulatory framework for AI remains fragmented and uncertain&quot;. Professional standards haven&#39;t gone anywhere, and they already expect review, supervision and documentation.</p><p>None of this is an argument to slow down. Agents are useful because so many of a firm&#39;s hours go to preparation: chasing documents, matching what arrived against what was requested, drafting the follow-up, assembling the workpaper. Hand that work over. Keep the release decision with your people.</p><p>The filing doesn&#39;t tell firms what to do. But it&#39;s the vendor&#39;s own risk disclosure, and it says plainly that review may not keep up with the speed of preparation.</p><p>That&#39;s the design we chose. AI agents for CPA and accounting firms and other professional service firms. Agents prepare the work. Sends wait for your team&#39;s approval by default. If you want to see how the approval step works, it&#39;s here: <a href="https://getprecision.ai/">https://getprecision.ai</a></p><p>The filing itself is public on EDGAR: https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm. The passages quoted above are in the Risk Factors section, pages 47 to 49 and page 69.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ca3a478da317" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>