<?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 Jan Gajowiak on Medium]]></title>
        <description><![CDATA[Stories by Jan Gajowiak on Medium]]></description>
        <link>https://medium.com/@jan.gajowiak?source=rss-b98bb43fc7c6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*IrQCqwPrGmFdj8xiTmLfBA.png</url>
            <title>Stories by Jan Gajowiak on Medium</title>
            <link>https://medium.com/@jan.gajowiak?source=rss-b98bb43fc7c6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 22:09:28 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jan.gajowiak/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[The AI System “Works”. But How Often Is It Right?]]></title>
            <link>https://medium.com/codino-pl/the-ai-system-works-but-how-often-is-it-right-f11f5f5d6f17?source=rss-b98bb43fc7c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/f11f5f5d6f17</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-testing]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Jan Gajowiak]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 16:41:01 GMT</pubDate>
            <atom:updated>2026-11-13T16:41:01.619Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bBbrVyHOHwC-7aMoEO2Urw.jpeg" /></figure><h3>A 95% accuracy score can hide critical failures. Here’s how to define correct results, measure mistakes and evaluate an LLM.</h3><p>If a program adds up the lines on an invoice, checking it is easy. Enter the amounts, run the function, compare the result with the total you expected. The test passes or it fails.</p><p>But what does a passing test mean when a language model is doing a big part of the job?</p><p>Say a company wants an LLM to spot urgent customer requests. The model gets ten different messages right. That’s a good result for those ten messages. It says nothing about the next ten, written differently. And it doesn’t tell you whether the same request, sent again, would land in the same queue.</p><p><strong>Ten correct answers prove ten correct answers. They don’t tell you how often the system will be wrong, or in what way.</strong></p><h3>Why testing an LLM is a different job</h3><p>The classic tests still apply to an application with AI inside it. Calculations, integrations, permissions, data writes and side effects all need checking as before. But once a <a href="https://codino.pl/blog/do-you-really-need-llm">large language model</a> is part of the application, a new difficulty appears. The same text can be understood in more than one way, and the same answer can be phrased in more than one way. Some of those differences are harmless. Others change the outcome of the whole process. Even with identical input and identical instructions, you can’t assume the model will answer identically every time.</p><p>Ordinary software can misbehave too, through bugs or because its output depends on something like the current date. <strong>What’s new is judging the quality of a component whose correct answer doesn’t always come in one fixed form.</strong></p><p>Which raises the obvious question: how many successful runs are enough? One? Ten? A hundred?</p><p>The count on its own settles nothing. Ten right answers to ten nearly identical, easy requests prove nothing about the messy variety a company actually receives. What you need is an agreed definition of “correct” and measurements taken on the real task.</p><h3>First decide what “correct” means</h3><p>In our example, the system labels requests as urgent or routine. Before anyone judges its answers, the company has to decide <strong>what makes a request urgent</strong>.</p><p>Is it only when the customer’s operations have stopped? What about a safety issue, a deadline that can’t be met or a clause in the contract? What about a message that says “urgent” but is really just asking for a document? Without those decisions, <a href="https://codino.pl/blog/why-data-does-not-improve-decisions">two people can look at the same case and disagree</a>. Then there’s no way to tell whether the model got it wrong or the rules were never set.</p><p>Here, the correct result is the right label under the agreed criteria. The model might write “needs immediate attention” or “urgent request”. If both lead to the same correct decision, the difference in wording isn’t a mistake.</p><p>It’s a different story when an identical request is routed to urgent handling once and dropped into the routine queue the next time. Now the business outcome has changed, not the phrasing.</p><p>Systems that generate longer text add one more risk. The model may invent a detail that isn’t in the documents or the data, like promising the customer a repair date nobody confirmed. That’s not the same as a misclassification. Both need controlling, but they’re different problems.</p><h3>What 95% accuracy can hide</h3><p>Take a simple, <strong>hypothetical</strong> set of 1,000 customer requests. Fifty genuinely need an urgent response. The other 950 are routine.</p><p>The system labels every single one as routine.</p><p>It’s right 950 times out of 1,000. That’s <strong>95% overall accuracy</strong>. Respectable, until you remember what the system was for.</p><p>It caught <strong>none of the 50 urgent requests</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pr4xH8pc8rD9PHc-mUpUvA.png" /></figure><p>This isn’t a real model’s result, just arithmetic that exposes the weakness of a single headline number. A system can post a high accuracy score because most cases fall into the easy, common category, and still fail completely at the one job the company bought it for.</p><p>That’s why a classification like this needs at least two measures.</p><p><strong>How many of the genuinely urgent requests did it catch?</strong> That’s <em>recall</em>, also called sensitivity. If it missed every urgent case, recall for that category is zero.</p><p><strong>How many of its alarms were justified?</strong> That’s <em>precision</em>. A system can flag too many routine requests as urgent and bury the team in false alarms. In our extreme example it raised no alarms at all, so there’s nothing to calculate precision from.</p><p>These are two different mistakes: a <strong>miss</strong>, where a case that needed action was overlooked, and a <strong>false alarm</strong>, where a routine case was escalated. A single accuracy percentage hides the balance between them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wSUCUFzDdfF9Gm0Nmyi14g.png" /></figure><p>One more thing. A score across a thousand <em>different</em> requests still says nothing about whether <strong>the same</strong> request gets the same label twice in a row. That’s a separate property, and it needs its own check.</p><h3>Can hallucinations and inconsistency be eliminated?</h3><p>There’s a lot you can do in an LLM-based system: constrain the output format, add rules, check claims against sources, validate results and route unclear cases to a person. Sometimes one plain rule in the application prevents a specific error better than yet another paragraph of instructions to the model.</p><p>These measures work. They can cut particular kinds of mistakes dramatically. But reducing risk is not the same as proving that <strong>no future answer will ever be wrong</strong>. A long streak of correct results is still not that proof.</p><p>Variation between answers can be reduced too, through how the model is configured and how the whole process is built around it. What nobody should promise is perfect repeatability just because of a single configuration setting. In most real applications, what matters anyway is that the <strong>decision</strong> stays consistent, not that every generated sentence is word-for-word identical.</p><p>If a vendor tells you their solution “doesn’t hallucinate”, ask what exactly they mean. Is the output checked against an agreed source? Can the system say it doesn’t know? What does it do when the information isn’t there? And which errors did they still see in testing?</p><h3>A newer model isn’t automatically better for you</h3><p>A clear definition of “correct” pays off again when a new model hits the market. It may be faster, cheaper per call or higher on the public leaderboards. None of that means it will label your customers’ requests more accurately.</p><p>The only comparison that counts is on <strong>your tasks, against your quality criteria</strong>. After that check you can talk about the new model’s advantages. Swapping the model name in a config file is easy. The change may still require adjusting integrations and instructions, and it certainly requires measuring quality again. Price per call alone shouldn’t decide the switch.</p><h3>What to ask before you sign off</h3><p>You don’t have to compute metrics or design test suites yourself. You should, however, expect five things from your team or your vendor:</p><ol><li><strong>What exactly does the system do?</strong> One clear action, such as labeling a request urgent or routine</li><li><strong>What counts as a correct result?</strong> Agreed criteria, with examples of correct, incorrect and borderline answers</li><li><strong>Which mistakes do we distinguish?</strong> Misses, false alarms, missing details and invented details, whichever apply to the task</li><li><strong>What is the claimed accuracy based on?</strong> How many cases were tested and of what kind, how many errors of each type showed up, and what the percentage actually measures</li><li><strong>Was the stability of the key decision checked?</strong> Can the same task, run again, produce a different outcome, even when both answers look convincing?</li></ol><p>Without these, nobody in the company can say what the test results actually prove. And that’s too thin a basis for signing off on a system.</p><h3>A test result is not a business decision</h3><p>For software built on an LLM, “does it work?” is the wrong question. The useful one is: <strong>in which tasks does it answer correctly, how often is it wrong, and what kind of wrong?</strong> Answering that takes agreed criteria and honest measurements, not one impressive example from a demo.</p><p>Even “99.8% of tested tasks correct” doesn’t close the subject. It describes a result under specific conditions, not a guarantee for the future. The real question is what happens in the cases where the system does get it wrong, and whether the company can live with the consequences.</p><p>And when the team asks for budget to push accuracy from 95% to 99%? Whether that’s worth paying for is the subject of the <a href="https://codino.pl/blog/is-99-percent-ai-accuracy-enough">next part</a>.</p><p><em>This article is </em><strong><em>Part 7 of 10</em></strong><em> in </em><strong><em>From Decision to Deployment</em></strong><em>, a Codino series on making better decisions about AI, data and process automation.</em></p><p><strong>Explore the series:</strong></p><p><em>Previous (Part 6): </em><a href="https://codino.pl/blog/why-data-does-not-improve-decisions"><em>You Have Data and Reports. Why Don’t They Lead to Better Decisions?</em></a></p><p><em>Next (Part 8): </em><a href="https://codino.pl/blog/is-99-percent-ai-accuracy-enough"><em>Is 99% AI Accuracy Enough?</em></a></p><p><em>Start from Part 1: </em><a href="https://medium.com/codino-pl/ai-wont-fix-a-chaotic-process-first-check-you-have-something-to-automate-254d47753f7d"><em>AI Won’t Fix a Chaotic Process. First Check You Have Something to Automate</em></a></p><p><em>Originally published on </em><a href="https://codino.pl/blog/how-often-is-ai-correct"><em>Codino</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f11f5f5d6f17" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codino-pl/the-ai-system-works-but-how-often-is-it-right-f11f5f5d6f17">The AI System “Works”. But How Often Is It Right?</a> was originally published in <a href="https://medium.com/codino-pl">Codino.pl</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[You Have Data and Reports. Why Don’t They Lead to Better Decisions?]]></title>
            <link>https://medium.com/codino-pl/you-have-data-and-reports-why-dont-they-lead-to-better-decisions-db3ea36be468?source=rss-b98bb43fc7c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/db3ea36be468</guid>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[business-intelligence]]></category>
            <category><![CDATA[data-driven-decisions]]></category>
            <category><![CDATA[business-strategy]]></category>
            <dc:creator><![CDATA[Jan Gajowiak]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 13:26:01 GMT</pubDate>
            <atom:updated>2026-11-06T13:26:01.588Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why having more data, reports and dashboards doesn’t automatically lead to better business decisions.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D4t0cSJhGyqa5bh7TS84JQ.jpeg" /></figure><p>The sales and cost reports are ready. And yet the company can’t approve a discount for a big customer without a fight.</p><p>The salesperson prices the deal on product margin. Finance adds transport and order-handling costs. Both calculations are correct. They just answer different questions. Until the company decides which one counts when a discount is approved, each side will defend its own number, and the customer will keep waiting.</p><p>You could produce another report, but a new table isn’t what’s missing. What’s missing is a shared rule for pricing the deal.</p><p>That’s one of several reasons why a company’s information doesn’t turn into decisions. Sometimes the data <a href="https://codino.pl/blog/from-excel-to-ai">arrives too late</a>. Sometimes preparing it depends on one person. Sometimes the report is finished and nobody uses it.</p><h3>Don’t grade the whole company. Pick one decision</h3><p>“Are we good at using data?” is too broad a question to answer with a single verdict. Approving discounts, planning shifts and handling complaints can each look completely different.</p><p>Pick one decision that comes back regularly and that someone in the company knows from the inside. Ideally one where the team already feels the friction: delays, arguments, numbers that need double-checking.</p><p>Then take <strong>the last three to five comparable cases</strong>. For each one, reconstruct what information was needed, when it arrived, who prepared it, what rules were used to read it and what was finally decided. This isn’t an audit. A simple table and a conversation with the people who were in the room will do.</p><p>That exercise separates four problems that, from the boardroom, all look like the same thing: “the decision is stuck again”.</p><h3>Four places where data stops helping</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PwurNAbpJc0Vj5S-KGoaew.png" /></figure><h3>1. Access: did the information arrive in time?</h3><p>In the discount case, the missing piece might be as simple as current transport costs or the extra delivery conditions. The team knows exactly what it wants to calculate. It just has to wait for someone to hand over the missing number.</p><p><strong>What to check:</strong> for each recent deal, was the full set of data available before the price was approved? If not, what was missing, and who could have supplied it?</p><p><strong>First move:</strong> agree where that information comes from, who owns it and by when it’s needed. For costs that can’t be known precisely yet, agree on an estimate instead of waiting for the final invoice.</p><h3>2. Preparation: how much work does an answer take?</h3><p>The data may all be there, but each time the same person has to gather it, check it and run the numbers. When that person is out, the decision waits. Or the calculation is produced regularly, just too late to help.</p><p><strong>What to check:</strong> how long did the calculation take for the recent deals, which steps were repeated every time, and could anyone else have done it?</p><p><strong>First move:</strong> write down how the calculation is actually produced and remove the most tedious repeated step. If the work is fast, reliable and doesn’t delay anything, leave it alone. Automation for its own sake isn’t the goal.</p><h3>3. Shared meaning: is everyone counting the same thing?</h3><p>Back to the two margins. If sales and finance include different costs, they get different, equally correct numbers. The trouble begins when both sides present theirs as the answer to the same question.</p><p><strong>What to check:</strong> ask both sides to recalculate the same handful of deals independently. Then compare which costs and assumptions each one includes. Are the differences deliberate, or did nobody ever agree on a definition?</p><p><strong>First move:</strong> set one basis for the calculation <em>for this type of discount decision</em>. Write down which costs are in, where the data comes from and how exceptions are handled. That doesn’t mean one margin definition has to serve every purpose across the company. It also won’t fix errors in the source data by itself.</p><h3>4. Use: did the numbers actually shape the decision?</h3><p>A team can have a complete, current, agreed calculation and still not use it. Maybe the report doesn’t answer the question the decision-maker is actually asking. Maybe nobody has said when a result needs a second approval. Or maybe the numbers are just one factor among several, along with customer size and the state of the relationship.</p><p><strong>What to check:</strong> did the person approving the discount see the calculation before deciding, and what difference did it make? If they skipped it, why?</p><p><strong>First move:</strong> decide what role the calculation is supposed to play for this kind of deal. Don’t jump straight to a hard threshold. First find out whether the numbers genuinely help or are just an attachment nobody opens.</p><p>Check these four in order, because a gap in access or preparation will undermine everything downstream. But treat it as <strong>an order of investigation, not an automatic diagnosis</strong>. The first problem you hit isn’t necessarily the most expensive or the most important one. And when you don’t know, ask the person doing the work rather than guessing.</p><h3>One problem, one change, one small experiment</h3><p>Suppose the review of recent deals confirms the margin problem: two definitions, two numbers. What can you do without launching an IT project?</p><p><strong>Start by measuring where you are.</strong> Take three to five comparable deals and count how many times sales and finance had to argue over how margin was calculated. Record how long it took to approve the price once the data was available. Don’t cherry-pick the worst cases.</p><p><strong>Then agree on one small change.</strong> Sales and finance settle which margin applies to this type of discount, which costs go into it and who rules on exceptions. One page of agreed rules. Not a new system.</p><p><strong>Test it in practice.</strong> For the next two to four weeks, or one full representative cycle if the decision is rarer than that, work by the agreed rules. Keep tracking the same two things: the number of arguments about the calculation and the time it takes to reach a decision. Compare like with like, because an unusual order will take longer no matter how good the data is.</p><p>Watch for side effects too. Are approved deals being corrected more often? Do the discounts still meet the profitability criteria you agreed, and are any exceptions approved deliberately? Faster isn’t a win if the decisions get worse.</p><p><strong>Then decide.</strong> If the arguments disappear, approvals move faster and profitability holds, keep the change. If the arguments were rare to begin with, or nothing got faster, don’t build it out any further. The real obstacle may be a multi-step approval chain, a vague discount policy or something the first pass simply didn’t reveal.</p><p>That way the company does more than tidy up a margin calculation. It finds out whether this problem was really what slowed the decision down. <strong>A hypothesis that doesn’t survive the test is a useful result too.</strong> It keeps you from investing in the wrong fix.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*03hp2uk4n6ms0-I02g5YhQ.png" /></figure><h3>Do you also need to predict what happens next?</h3><p>Not every decision needs a forecast or an early warning. Those only earn their place when they let you do something differently. For discount approvals, a simple flag that a deal needs a second look may be all that’s required. In other processes a forecast can be genuinely useful, but only if someone has the time and the authority to act on it.</p><p>An alert firing doesn’t mean a better decision got made. What matters is who gets it, when, and what they do next. Not every signal like this <a href="https://codino.pl/blog/do-you-really-need-llm">needs AI</a>.</p><h3>The report is not the end of the process</h3><p>A good report makes a situation easier to understand. It doesn’t replace agreeing on what the key numbers mean, who makes the call and what they do with the information once they have it.</p><p>So before you order the next dashboard, pick one decision that keeps coming back and trace the last few times it was made. You may find that all it takes is agreeing on one definition, setting a deadline for one data set or changing how an existing report is read. And if it all works well enough already, there’s nothing to fix.</p><p>And when part of the process is handled by AI, one more question comes up: <a href="https://codino.pl/blog/how-often-is-ai-correct">how often does the system actually get it right</a>?</p><p><em>This article is </em><strong><em>Part 6 of 10</em></strong><em> in </em><strong><em>From Decision to Deployment</em></strong><em>, a Codino series on making better decisions about AI, data and process automation.</em></p><p><strong>Explore the series:</strong></p><p><em>Previous (Part 5): </em><a href="https://codino.pl/blog/from-excel-to-ai"><em>From Excel to AI</em></a></p><p><em>Next (Part 7): </em><a href="https://codino.pl/blog/how-often-is-ai-correct"><em>How Often Is AI Correct?</em></a></p><p><em>Start from Part 1: </em><a href="https://medium.com/codino-pl/ai-wont-fix-a-chaotic-process-first-check-you-have-something-to-automate-254d47753f7d"><em>AI Won’t Fix a Chaotic Process. First Check You Have Something to Automate</em></a></p><p><em>Originally published on </em><a href="https://codino.pl/blog/why-data-does-not-improve-decisions"><em>Codino</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db3ea36be468" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codino-pl/you-have-data-and-reports-why-dont-they-lead-to-better-decisions-db3ea36be468">You Have Data and Reports. Why Don’t They Lead to Better Decisions?</a> was originally published in <a href="https://medium.com/codino-pl">Codino.pl</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Excel to AI. Six Ways Companies Work With Data]]></title>
            <link>https://medium.com/codino-pl/from-excel-to-ai-six-ways-companies-work-with-data-f0875aea5892?source=rss-b98bb43fc7c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/f0875aea5892</guid>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[data-engineering-101]]></category>
            <dc:creator><![CDATA[Jan Gajowiak]]></dc:creator>
            <pubDate>Fri, 30 Oct 2026 12:26:01 GMT</pubDate>
            <atom:updated>2026-10-30T12:26:01.505Z</atom:updated>
            <content:encoded><![CDATA[<h3>Moving beyond spreadsheets isn’t just about adding AI. It’s about making business data reliable, accessible and useful for decisions.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AnO9EvStPMTuOUKVjlcS_g.jpeg" /></figure><p>A company is trying to figure out which stage of order fulfillment causes the biggest delays. The system shows the current status of every order, but it never records when a status changed. The team digs through logs and other sources and comes up empty. There is no complete history anywhere. Management can see the backlog. What it can’t see is where the backlog forms. So it may well add people in exactly the wrong place.</p><p>Another company has every piece of data it needs. Once a week, someone pulls sales, stock levels and open orders from three different systems and stitches them together in Excel. The spreadsheet is fine. It just lands on the team’s desk after the deadline for approving purchases has passed.</p><p>Neither company is getting what it could out of its data, each for a different reason. The first is missing part of the record. The second has the record but can’t get it ready in time. A new report won’t fix both, and neither will AI. <strong>Before you decide what to change, you need an honest picture of how the company actually works with data today.</strong></p><h3>Six ways of working with data</h3><p>What follows are six situations you’ll recognize, not a maturity scale for grading the whole company. Sales might run on polished reporting while production still lives on summaries built by hand. So look at one process or one area at a time, not at the business as a whole.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oGlNgb85OJKKJmYOgGkfuQ.png" /></figure><h3>1. The current state is visible, the history isn’t</h3><p>The system says a complaint was closed. It doesn’t say how long the customer waited for a first reply. It shows where an order is right now, but not how long it sat at each stage along the way.</p><p>This isn’t a reporting problem. The information the analysis needs simply doesn’t exist. The first thing to do is check whether a trail survives somewhere else: in logs, in documents, in another system. If it doesn’t, you can start recording the relevant events and their timestamps from now on. From that point, each run of the process can be compared with the last.</p><p>That’s worth doing when the history feeds a decision that keeps coming back. Just don’t expect it to recover the months you didn’t record.</p><h3>2. The data exists, but hardly anyone looks at it</h3><p>The systems hold years of sales, orders and support tickets. Yet when someone asks why complaints are climbing or margin is slipping, answering means a one-off investigation. Nobody returns to the data on a regular basis.</p><p>You don’t need a data warehouse to fix this. Start with the one question that comes up in every meeting and build a simple view of how that number has moved over time. If the view helps someone decide something, make sure it gets refreshed regularly.</p><p>That’s the key difference from the first situation: <strong>the information is already there. It’s just not being used.</strong></p><h3>3. Exports from several systems, merged by hand in Excel</h3><p>Say the company plans purchasing every week. It needs sales from one system, stock from another and open orders from a third. Someone downloads the three files, combines them and chases down the discrepancies.</p><p>That spreadsheet shows something no single system can, so Excel is doing real work here. The problem starts when preparing it takes so long that the picture arrives after the deadline for placing orders. The company decides blind, even though all the data existed days earlier.</p><p>In that case, speeding up the refresh is worth far more than the hours it saves. It means the data is there <strong>before the decision instead of after it</strong>. Often an automated import or a self-refreshing workbook is enough. Nobody has to give up Excel.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ePw-pHif8sNfusO5akQFmA.png" /></figure><p>If the report is only needed now and then, and it’s quick and reliable to produce by hand, the manual version may still be the sensible choice.</p><h3>4. Data in one place, but every analysis still made to order</h3><p>The company has a data warehouse, or at least a shared data set. The analyst no longer collects files from each system by hand. But when the board asks about a particular metric, the analyst still has to write a query, build a spreadsheet or put together a separate analysis.</p><p>That’s not the same as regular reporting. Shared data shortens the path to an answer. It doesn’t mean the answers are ready and waiting.</p><p>For unusual questions, analysis on demand makes sense. If the same metrics come back every Monday, though, a repeatable report serves everyone better. What tips the balance is how often the question recurs, not whether a warehouse happens to exist.</p><h3>5. Regular reports and BI</h3><p>Managers open a ready-made report on sales, costs or plan versus actuals. A BI tool lays it out in tables and charts and refreshes it on a schedule.</p><p>Most of the effort of rebuilding the same summary every time disappears. Results are easier to compare and deviations easier to spot. The question that remains is whether the report is fresh enough for the decision it’s meant to support. Data refreshed once a day is fine for a weekly review and too stale for some day-to-day operational decisions.</p><p>Some companies have a dozen dashboards and still commission one more report before every important meeting. Adding another view at that point doesn’t touch the underlying problem.</p><h3>6. Forecasts, alerts and AI</h3><p>Analysis can also look forward. The company gets a warning about a likely stock shortage, a demand forecast or a suggestion for which ticket to pick up first.</p><p>It helps to keep three things apart. An <strong>alert</strong> can be a plain rule, like a threshold being crossed. A <strong>forecast</strong> can come from standard statistics. <strong>AI</strong> earns its place where the patterns are too complex for simpler methods. Not every warning and not every forecast needs artificial intelligence (we’ve written separately about <a href="https://codino.pl/blog/do-you-really-need-llm">when an LLM is really needed</a>).</p><p>What makes any of this valuable is whether someone can act on it. A shortage warning helps when there’s still time to change the order or the delivery date. If it arrives too late, even a perfectly accurate model changes nothing.</p><p>AI doesn’t automatically mean building a warehouse or replacing your systems. It does mean having reliable access to data you can trust, the right permissions and a <a href="https://codino.pl/blog/ai-automation-business-case">business case</a> that holds up. Depending on where you start, that’s either a small extension of an existing process or a serious integration project.</p><h3>Why this isn’t a ladder</h3><p>A company that makes one monthly decision from a single spreadsheet, refreshed correctly and on time, may already have everything that decision requires. Swapping the spreadsheet for a BI platform will raise costs without necessarily improving the outcome.</p><p>Another company may have a warehouse and dozens of dashboards and still not get an answer to its most important question when it matters.</p><p>So there’s no obligation to climb from the first situation to the sixth, and no obligation to end up at AI. <strong>What justifies an investment is the constraint it removes, not how modern the tool is.</strong></p><h3>When a small change makes a big difference</h3><p>Back to the weekly purchasing plan. If the information already exists and the only thing delaying the decision is a manual merge, fixing that one summary can get the decision made on time, and no major IT project is required. How much that’s worth depends on how often the problem recurs and how important the delayed decision is.</p><p>That’s a possibility, not a rule. When the data was never captured, is full of gaps or can only be combined at great expense, even a change that looks simple can turn out to be hard. And automating a report nobody reads helps no one.</p><p>Companies that make little use of their data <strong>may</strong> have easy wins waiting. But not <strong>necessarily</strong>. Before you invest, check two things: will the new information arrive in time, and will someone actually do something with it?</p><h3>What a CEO should do</h3><p>Rather than planning a new platform, pick one business question that comes back regularly. Find out where the data for it comes from, how long the answer takes to produce and whether it’s ready before the decision has to be made. From there, three outcomes are all perfectly reasonable.</p><p><strong>Leave it as it is.</strong> If a spreadsheet that can be refreshed quickly and correctly serves a monthly decision, and the result is used on time, you don’t need a new system just to retire Excel.</p><p><strong>Fix the repeated work.</strong> If someone merges the same exports by hand every week and the report still arrives late, look at automating the pull or the refresh. Start with the one summary people actually use, not with rebuilding all of reporting.</p><p><strong>Start recording what’s missing.</strong> If nobody can say how long each stage of an order took, check what event trails already exist. If the history genuinely isn’t there, decide which status changes and dates to capture from now on.</p><p>Each of these answers a different problem. None of them requires moving up to a more advanced technology just for the sake of it.</p><p>Before you plan the next investment, look at the last important decision in the process you picked. <strong>Was the information you needed on hand while there was still time to act on it?</strong></p><p>And if it was, and the decision still stalled? That’s what the <a href="https://codino.pl/blog/why-data-does-not-improve-decisions">next part</a> is about.</p><p><em>This article is </em><strong><em>Part 5 of 10</em></strong><em> in </em><strong><em>From Decision to Deployment</em></strong><em>, a Codino series on making better decisions about AI, data and process automation.</em></p><p><strong>Explore the series:</strong></p><p><em>Previous (Part 4): </em><a href="https://codino.pl/blog/do-you-really-need-llm"><em>Do You Really Need an LLM? Picking the Right Technology for Process Automation</em></a></p><p><em>Start from Part 1: </em><a href="https://medium.com/codino-pl/ai-wont-fix-a-chaotic-process-first-check-you-have-something-to-automate-254d47753f7d"><em>AI Won’t Fix a Chaotic Process. First Check You Have Something to Automate</em></a></p><p><em>Browse more articles: </em><a href="https://codino.pl/blog"><em>Codino Blog</em></a></p><p><em>Originally published on </em><a href="https://codino.pl/blog/from-excel-to-ai"><em>Codino</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0875aea5892" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codino-pl/from-excel-to-ai-six-ways-companies-work-with-data-f0875aea5892">From Excel to AI. Six Ways Companies Work With Data</a> was originally published in <a href="https://medium.com/codino-pl">Codino.pl</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Do You Really Need an LLM? Picking the Right Technology for Process Automation]]></title>
            <link>https://medium.com/codino-pl/do-you-really-need-an-llm-picking-the-right-technology-for-process-automation-c5217ee0703f?source=rss-b98bb43fc7c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/c5217ee0703f</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[automation]]></category>
            <dc:creator><![CDATA[Jan Gajowiak]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 10:51:45 GMT</pubDate>
            <atom:updated>2026-10-27T10:51:45.970Z</atom:updated>
            <content:encoded><![CDATA[<h3>How to choose between LLMs, rules and hybrid solutions without creating a system that’s unnecessarily expensive to maintain.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4ZJ-AC2TZuFb8O2tYtLT9A.jpeg" /></figure><p>In one of the systems we built at Codino, the entry point looks trivial. We receive a photo of a receipt and want to answer a simple question: what did the customer actually buy?</p><p>The first instinct these days is obvious enough: feed the photo to an LLM and collect the structured data. But once you start looking at the whole process, it quickly becomes clear that “analyzing a receipt” isn’t one problem.</p><p>You have to read the text, understand the document’s structure, and pull out the date, the amounts, the store, and the individual line items. Then the abbreviated store name has to be matched to a specific retail chain, and the product as written on the receipt (often nothing like the name it’s sold under) to the actual product. After that comes enriching the data with additional attributes, and only on top of all that can you build a picture of the customer’s buying behavior.</p><p>From the business’s perspective, it’s one process. From the solution side, it’s several distinct problems, and they don’t have to share the same technological answer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6t3UHrBRs2LMZWmSPzIf_g.png" /></figure><p>Which is why, before I ask “which LLM should we use?”, I prefer to ask a different question:</p><blockquote><strong>What exactly does an LLM give us that a simpler approach doesn’t, and what will the company be left with after the rollout?</strong></blockquote><p>That distinction matters even when the automation itself already has a <a href="https://codino.pl/blog/ai-automation-business-case">solid business case</a>. Even a profitable project can be ruined by pairing a simple problem with a solution that’s expensive to maintain or needlessly hard to change.</p><h3>One process can contain several completely different problems</h3><p>Receipts share some common traits, but they can also differ wildly between stores and cases. Product names, discounts, taxes and seller details are written differently everywhere. And then there’s photo quality, which isn’t consistent either.</p><p>On this project we compared several approaches. One was as close to end-to-end as possible: the model gets the image and tries to return the required information in one shot. Another used an LLM only where the variability and ambiguity are greatest. We also tested solutions built on classic algorithms, heuristics and rules.</p><p>And this is where you hit the thing that matters more in practice than any specific model choice: <strong>no single solution wins every time</strong>.</p><p>If you’re handling a well-known range of documents and care about very low latency and a low unit cost, a more deterministic approach can make sense. If, on the other hand, the system has to adapt quickly to new formats, unusual ways of writing things, and a large number of variants, the model’s flexibility can be worth far more.</p><p>Then there’s the required quality bar and the kind of errors you’re willing to accept. Misreading a secondary field is a different problem from getting the store or the purchase amount wrong.</p><p>So I break the process into parts to find the places where the difficulty actually lives. <strong>Not to immediately assign each part its own technology.</strong></p><h3>The simplest component doesn’t always produce the simplest system</h3><p>This is where the second trap lies.</p><p>You can find a simple mechanism for every part: a parser, a handful of rules, a separate classifier, extra logic for the exceptions, and integrations between all of them. Each piece looks reasonable on its own. But after a while, changing one document variant can require fixes in several places, and bugs start appearing not inside any single algorithm, but at the seam between two components that each work correctly.</p><p>In other words, you can make a series of individually simple technical decisions and still end up with a system that’s hard to maintain.</p><p>So I don’t just look at whether a single component is cheap or easy to build. I also want to know how much of our own logic we’ll have to maintain later, how many dependencies we’re creating, how hard the system will be to change in six months, whether we’ll be able to trace the cause of an error, and what the whole thing costs to run every day.</p><p>In practice, it comes down to one simple question:</p><blockquote><strong>What will the company be left with after the rollout?</strong></blockquote><p>If the “simple” option means dozens of rules, plenty of exceptions, and several places that need touching every time a new case shows up, its simplicity may be an illusion.</p><h3>An LLM can remove the rules. In return, you have to control the model</h3><p>The same logic works in the other direction.</p><p>An LLM is not a more complicated solution by definition. If the core problem is a large number of variants in language, documents, or ways of describing the same thing, a model can cut down the need to keep hand-writing new rules and exceptions.</p><p>Receipt analysis shows this well. Reading a specific, unambiguous value can be a simple task. Matching an abbreviated product name to the actual product is a different problem entirely, especially when every store uses its own shorthand and conventions.</p><p>That doesn’t automatically mean “this needs an LLM”. It does mean this is exactly the kind of spot where comparing approaches is worth it, because the cost of hand-describing variant after variant can eventually exceed the cost of using a model.</p><p>Except that work doesn’t disappear. It just moves.</p><p>Instead of maintaining a growing list of rules, you start controlling the quality of the model’s answers. You have to check how it behaves on new data, whether a change of model or instructions degrades the results, and what to do with the cases where the answer is uncertain.</p><p>So I don’t ask which technology is “inherently simpler”. What interests me more is <strong>what kind of work and risk we’re signing up for later</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MLwfnUCS_rt53d3SfS-y1g.png" /></figure><p>In one option, that’s maintaining rules and exceptions. In another, it’s testing and controlling a more flexible model. In yet another, the best answer is a hybrid, with the LLM working only where its flexibility genuinely adds something.</p><p>That’s why on projects like this we don’t start by picking a model. First we find where the difficulty lives, then we compare different ways of solving it on real data. The technology follows from the requirements of the process. It’s not the starting point.</p><h3>Three questions I’d ask before choosing the technology</h3><p>A CEO doesn’t need to know whether a given piece of the process is best solved by a parser, a classic model, or an LLM. They should, however, expect the team to clearly justify why the proposed solution makes sense.</p><p>The first question:</p><blockquote><strong>Where in this process does the difficulty actually live?</strong></blockquote><p>Is the problem executing a known rule, moving data between systems, or interpreting cases we can’t sensibly describe up front?</p><p>The second:</p><blockquote><strong>What does an LLM give us beyond the simpler approach, or what does it let us remove?</strong></blockquote><p>The answer doesn’t have to be higher accuracy alone. The value can also be fewer hand-maintained rules, faster support for new variants, fewer separate models, or a shorter build.</p><p>And the third:</p><blockquote><strong>What will the company be left with after the rollout: what will it have to maintain, change, control, and keep paying for?</strong></blockquote><p>That last question protects you from both extremes. On one side, using an LLM just because it’s there. On the other, piling up rules and specialized mechanisms just because each one looks simpler on its own.</p><p>That’s how I actually approach technology choices in automation. <strong>I’m not looking for the simplest technology. I’m looking for the simplest solution as a whole that meets the business requirements.</strong></p><p>A good decision isn’t about using as little AI as possible, or as much AI as possible. It’s about a solution that doesn’t just work in a demo, but one you can sensibly maintain, evolve and use every single day.</p><p>And once you know what kind of solution makes sense, the next question comes up: are the company’s <a href="https://codino.pl/blog/from-excel-to-ai">data and systems in good enough shape</a> to actually build it.</p><p><em>This article concludes </em><strong><em>From Decision to Deployment</em></strong><em>, a ten-part series on making better AI automation decisions.</em></p><p><em>Previous: </em><a href="https://codino.pl/blog/ai-automation-business-case"><em>Does Automation Pay Off? How to Build the Business Case Before You Invest in AI</em></a></p><p><em>Start from Part 1: </em><a href="https://medium.com/codino-pl/ai-wont-fix-a-chaotic-process-first-check-you-have-something-to-automate-254d47753f7d"><em>AI Won’t Fix a Chaotic Process. First Check You Have Something to Automate</em></a></p><p><em>Originally published on </em><a href="https://codino.pl/blog/do-you-really-need-llm"><em>Codino</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c5217ee0703f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codino-pl/do-you-really-need-an-llm-picking-the-right-technology-for-process-automation-c5217ee0703f">Do You Really Need an LLM? Picking the Right Technology for Process Automation</a> was originally published in <a href="https://medium.com/codino-pl">Codino.pl</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Does Automation Pay Off? How to Build the Business Case Before You Invest in AI]]></title>
            <link>https://medium.com/codino-pl/does-automation-pay-off-how-to-build-the-business-case-before-you-invest-in-ai-2e3344978bd5?source=rss-b98bb43fc7c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/2e3344978bd5</guid>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[process-improvement]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Jan Gajowiak]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 15:40:26 GMT</pubDate>
            <atom:updated>2026-10-20T15:40:26.268Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why recovered hours aren’t always savings, and how to compare the real costs and benefits of an automation project before you commit.</h3><p>A process can look like a <a href="https://codino.pl/blog/ai-process-decision">sensible automation candidate</a>. There’s enough repetitive work in it, we understand the problem, and we know a system could technically take over part of the load.</p><p>At this point it’s very tempting to jump straight to simple math: the process takes people 200 hours a month, automation can absorb 70% of the work, so just price out 140 hours and compare that with the cost of the project.</p><p>I often start with an estimate like that myself. It’s quick, and it gives you a sense of scale.</p><p>The trouble starts when we treat the recovered time as money that will actually leave the company’s cost base. In practice, those are very often two different things.</p><p>So before I get to ROI, I prefer to ask a different question:</p><blockquote><em>What will actually change in the economics of this process if we automate it?</em></blockquote><h3>A good candidate isn’t automatically a good investment</h3><p>The distinction sounds obvious, but in automation conversations it blurs all the time.</p><p>A process can be a great candidate operationally: manual, repetitive, time-consuming and technically automatable. That still tells us nothing about whether it’s worth spending money on.</p><p>In one case we analyzed, we could technically have automated a large share of a document-analysis process. But once we ran the numbers on volume, the work that would remain with people, and the consequences of errors, we recommended against full automation. The system could be built. The business case didn’t hold up.</p><p>And it’s that second step that interests me here.</p><p>A good business case shouldn’t just show that the system will do the work. It should show that after the rollout the company will genuinely be better off (financially, operationally, or both), and that the improvement is large enough relative to the cost and the risk.</p><h3>200 recovered hours can be worth a lot. Or almost nothing</h3><p>The most common shortcut looks like this:</p><blockquote><em>hours recovered × hourly labor cost = savings.</em></blockquote><p>Sometimes that math is entirely justified. If automation cuts outsourcing, overtime, or some other real expense, the benefit genuinely hits the bottom line.</p><p>Sometimes the value sits in the future. The company is growing, and without automation it would have to hire another person within months. The project doesn’t reduce today’s costs, but it heads off an expense that was very likely coming.</p><p>And then there’s the situation I run into most often: people get time back, but the cost of the team doesn’t change. That benefit can still be very real, if it lets the company handle more volume, close cases faster, or move people to higher-value work.</p><p>Except you still have to be able to actually use that freed-up capacity.</p><p>Imagine automation saves 20 minutes a day for thirty people. Over a month, that adds up to roughly 200 hours. In a spreadsheet, it looks like more than a full FTE.</p><p>In reality, you can’t lay off 1.25 people. If those 20 minutes are scattered across the whole organization and don’t translate into higher throughput or specific work getting done, the financial value can be far lower than the sum of hours suggests.</p><p>Which is why, on projects like this, I ask two questions:</p><blockquote><em>What exactly will we do with the recovered time?</em></blockquote><p>and:</p><blockquote><em>Can we turn that recovered time into a specific business result?</em></blockquote><p>If the answer is “we’ll handle 20% more cases without growing the team”, you have a concrete value mechanism. If the answer is “everyone will have a bit less on their plate”, I’d be careful about booking the full value of those hours as savings.</p><p>One more thing: never count the same benefit twice. If the recovered time lets you avoid a new hire, that avoided cost is already the business result. I wouldn’t add the value of those same hours on top of it.</p><h3>The value of a project isn’t just lower costs</h3><p>With automation it’s easy to focus purely on reducing work. In practice, the value can come from several places.</p><p>A real cost can disappear. A future cost can be avoided. Throughput can go up, generating additional margin. And the cost of errors, rework and delays can change.</p><p>I’d be especially careful with the third case. If automation lets you handle an extra USD 1 million in sales, I wouldn’t put USD 1 million of benefit into the business case. What interests me is the additional margin that actually stays in the company thanks to the extra throughput.</p><p>It’s a small difference in wording, and a very large one in the spreadsheet.</p><h3>Don’t compare today’s process with an idealized tomorrow</h3><p>The second common problem shows up when we compare the cost of the process today with the cost of the process “after automation”.</p><p>I prefer to look at two possible futures over the same period.</p><p>The first is the future without the project. What happens to volume? Will the team have to grow? Will outsourcing increase? What will errors, rework and delays cost if the process stays as it is?</p><p>The second is the future after the rollout. Here you have to count not just what the system takes over, but also what stays with people: verification, exceptions, corrections, and the harder cases.</p><p>And then there’s the technology itself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4LHiaXE9VWq9hHuxFF3fig.png" /></figure><p>Here I apply a simple rule: compare the full cost of the process in both scenarios, not the price of an IT project against the cost of people’s work.</p><p>In practice, that means including the cost of getting to the solution and running it afterwards: implementation, integrations, your own team’s time, the technology, and the work that still remains after automation. I wouldn’t build a full multi-year TCO model at this stage, but leaving these items out can seriously flatter the result.</p><p>I look at errors the same way. The current process can produce mistakes too, so I’m not interested in “the cost of AI errors” on its own, only in how the cost of errors and rework changes between the no-project scenario and the post-rollout one.</p><h3>Five line items that are enough for a first decision</h3><p>At the start, you don’t need a financial model with twenty tabs. Five questions usually suffice:</p><ol><li><strong>Without the project:</strong> what will the process look like in 12–24 months?</li><li><strong>After the rollout:</strong> what genuinely disappears, and what stays with people and the system?</li><li><strong>The difference:</strong> which cost goes away, which one do we avoid, what additional margin can we generate, and how does the cost of errors change?</li><li><strong>The investment:</strong> what does it really cost to get there?</li><li><strong>The decision threshold:</strong> how good does the result have to be for the company to consider the project worth it?</li></ol><p>That last item matters more than it might seem. A positive result is not yet a good investment. The project has to be weighed against the risk, the payback period the company expects, and other uses for the same capital.</p><h3>A simple example: USD 60k of annual benefit. So what?</h3><p>Say a company processes about 4,000 documents a month today and expects 6,000 within a year. The current team is already close to its limit.</p><p>Without automation, the company assumes that kind of growth will require one more person, at a fully loaded annual cost of about USD 150k. It also assumes errors and rework will cost about USD 30k a year at that volume.</p><p>With automation, the extra hire isn’t needed, but other costs appear: about USD 60k a year to run the solution, USD 40k for external verification of the harder cases, while the cost of errors and rework drops to about USD 20k.</p><p><strong>Annual costs, without the project vs. with automation:</strong></p><ul><li><strong>Additional hire:</strong> USD 150k vs. USD 0</li><li><strong>Technology:</strong> USD 0 vs. USD 60k</li><li><strong>External verification:</strong> USD 0 vs. USD 40k</li><li><strong>Errors and rework:</strong> USD 30k vs. USD 20k</li><li><strong>Total:</strong> USD<strong>180k vs. USD 120k</strong></li></ul><p>The annual benefit, coming from the difference between the scenarios, is therefore about USD 60k.</p><p>If the implementation costs USD 120k, the simple payback period is about two years.</p><p>That’s payback, not ROI.</p><p>Under the same simplification, the three-year ROI would be about 50%: over three years the project would generate USD 180k of benefit, and after subtracting the USD 120k initial investment, USD 60k would remain, half the amount invested. For a larger, multi-year investment, back-of-the-envelope math like this obviously doesn’t replace a proper financial analysis.</p><p>But the more important question is: is two years a good result for this company?</p><p>There’s no universal answer.</p><p>If the company expects this type of project to pay back within 18 months at most, this one, as it stands, doesn’t make the cut. At a USD 120k investment, it would need about USD 80k of annual benefit, not 60.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fv6X2WbS4NlmUnOD_igz6g.png" /></figure><p>And this is where the business case starts helping you make the decision instead of merely justifying it.</p><p>We can ask: what would have to change to get from 60 to 80? A lower cost of exceptions? Higher volume? A larger share of cases handled without a human? Additional margin from the extra throughput?</p><p>If the biggest unknown is, say, the share of cases the system can handle on its own, you can invert the math and calculate the minimum required automation rate. Then run a pilot and check whether real data comes in above that bar.</p><p>To me, that’s far more useful than declaring “we assume 80% automation”.</p><h3>A good decision doesn’t always end in a rollout</h3><p>After running the numbers, I usually see one of three situations.</p><p>The project holds up even under conservative assumptions and clears the required return threshold. Then it’s worth moving forward.</p><p>The result hinges on one or two things we don’t know yet. Then we don’t need a full rollout, just a test designed specifically to measure those unknowns.</p><p>Or the project only looks good under very optimistic assumptions. Then it’s better to shrink the scope, or not do it for now.</p><p>In conversations with companies, that is sometimes the most valuable outcome of the analysis: not finding a way to implement AI, but quickly reaching the conclusion that in this particular spot the investment doesn’t hold up.</p><p>A good business case isn’t there to justify the project. It’s there to help decide whether the company should actually spend the money.</p><p>And if the answer is “yes”, only then is it worth moving to the next question: what’s the simplest solution that will deliver this result, and do we really need AI for it.</p><p><em>This article is part of </em><strong><em>From Decision to Deployment</em></strong><em>, a ten-part series on making better AI automation decisions.</em></p><p><em>Previous: </em><a href="https://codino.pl/blog/ai-process-decision"><em>Don’t Automate Your Most Important Process First. Pick One the Company Can Learn On</em></a></p><p><em>Next: </em><a href="https://codino.pl/blog/do-you-really-need-llm"><em>Do You Really Need an LLM? Picking the Right Technology for Process Automation</em></a></p><p><em>Originally published on </em><a href="https://codino.pl/blog/ai-automation-business-case"><em>Codino</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2e3344978bd5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codino-pl/does-automation-pay-off-how-to-build-the-business-case-before-you-invest-in-ai-2e3344978bd5">Does Automation Pay Off? How to Build the Business Case Before You Invest in AI</a> was originally published in <a href="https://medium.com/codino-pl">Codino.pl</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Best First AI Automation Project Isn’t Your Biggest One]]></title>
            <link>https://medium.com/codino-pl/the-best-first-ai-automation-project-isnt-your-biggest-one-ab95653c34e8?source=rss-b98bb43fc7c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/ab95653c34e8</guid>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[process-improvement]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[business-strategy]]></category>
            <dc:creator><![CDATA[Jan Gajowiak]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 12:03:09 GMT</pubDate>
            <atom:updated>2026-10-16T12:03:09.891Z</atom:updated>
            <content:encoded><![CDATA[<h4>How to choose a process that delivers a visible result, keeps risks manageable, and gives your team room to learn.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6k85gtdE_Z-Z78_DWxkEsA.jpeg" /></figure><p>You have a few processes that look like good automation candidates. Each has a clear beginning, a final result, and fragments you can point to that eat up people’s time today.</p><p>Which brings up the next question:</p><blockquote><em>Which process do we start with?</em></blockquote><p>In many companies the answer comes quickly: the most important one. The one that hurts the most. The one that, after a successful rollout, would most impress the board, the owner, or the investors.</p><p>That sounds reasonable.</p><p>And that’s exactly what makes it dangerous.</p><p>The most important process in the company is not always the best first process to automate. It’s often too complex, too visible and too expensive to be the place where the organization learns automation.</p><p>A well-chosen first process doesn’t have to be the biggest one. It should, however, have a good ratio of impact to cost and risk. It should give the company a noticeable result while letting it learn a new way of working without unnecessary exposure.</p><p>This article is for CEOs, owners and boards who already have a few automation ideas and don’t want to burn the budget on a badly chosen first project.</p><h3>The first project’s job is to teach the company how to automate</h3><p>We often hear a version of this argument:</p><blockquote><em>Let’s start with the key process. If it works, it’ll be a big win. The board will see the result and unlock budget for more automation.</em></blockquote><p>I understand the thinking. The CEO wants to see business impact. The board wants the investment justified. The owner doesn’t want to fund a project that’s just an experiment of no consequence.</p><p>The problem is that the first automation is rarely just a technology project.</p><p>It’s the moment the company learns many things at once: how to describe processes, prepare data, define a correct result, talk about errors, test a solution before production, and decide when the system can act on its own versus when a human should approve the outcome.</p><p>That’s a lot of learning for one project.</p><p>Which is why it’s not always wise to gain that experience on a process you can’t afford to break.</p><p>If the first automation touches a strategic area, every mistake costs more. Every delay is more visible. Every misunderstanding between business and technology turns into frustration faster.</p><p>And the pressure for a quick win can drive bad decisions: too broad a scope, too much automation, too little testing.</p><p>A good first project should be important enough that its result matters, but not so critical that the company can’t afford to learn on it.</p><p>If you can automate a smaller, repetitive process and show a concrete result, the conversation with the board becomes much easier:</p><blockquote><em>If we could clean up this less strategic process and get measurable benefits out of it, let’s see what we can achieve in the core ones.</em></blockquote><p>That’s a safer road than starting at the hardest place in the company.</p><h3>The biggest problem isn’t always the best first project</h3><p>The most important processes usually share a few traits that make them a hard place to start.</p><p>First, they’re complex. They have exceptions, dependencies, judgment calls, informal arrangements, and people who “just know” what needs to be done. As long as everything runs manually, the company often doesn’t see this. It becomes visible the moment you have to describe the process to a system.</p><p>Second, mistakes are expensive. An error in a minor administrative task usually means a correction. An error in a financial, legal, sales or operational process can have far more serious consequences.</p><p>Third, they’re politically visible. If the first automation project touches a process the whole board is watching, there’s little margin for learning. And a first implementation almost always surfaces something the company hadn’t seen before: gaps in the data, inconsistent rules, undocumented exceptions, or teams working in different ways.</p><p>None of this means core processes shouldn’t be automated. That’s often exactly where the biggest value sits.</p><p>It’s a question of sequence.</p><p>First build the ability to automate on a safer process. Then move into strategic areas with more confidence.</p><h3>A good first process delivers a result and contains the risk</h3><p>Don’t go looking for a perfect process. Those rarely exist.</p><p>Look for a process that meets a few simple conditions.</p><p>First, it happens often enough. If something occurs once a quarter, even a well-built automation may not produce a result that convinces the organization.</p><p>Second, it’s repeatable. It doesn’t have to be simple, but it should have parts that can be described and verified.</p><p>Third, a mistake shouldn’t immediately cause serious damage. Early on, it’s better to choose processes where a system error can be caught and corrected quickly.</p><p>Fourth, it helps if the process allows a hybrid setup. In practice, a first project rarely means the system doing everything on its own. More often the system takes over part of the work while a human keeps the harder cases.</p><p>And fifth: the first result should show up fast. Not a full production system in two weeks, but within 30–60 days you should be able to show the direction makes sense: less manual work, faster handling, fewer errors, or better-prepared data.</p><h3>The impact, cost and risk matrix</h3><p>To avoid choosing the first process based on gut feeling or internal pressure, use a simple matrix.</p><p>It doesn’t have to be an elaborate financial model. To start, it’s enough to score each process across six areas.</p><ul><li><strong>Business impact.</strong> Will automation cut cost, save time, improve quality, increase throughput, or improve the customer experience? A high score means high value for the company.</li><li><strong>Volume and repeatability.</strong> Does the process happen often, and in a similar way each time? A high score means many similar cases.</li><li><strong>Speed of first results.</strong> Can you show a result within 30–60 days, even on a fragment of the process? A high score means quick proof the direction makes sense.</li><li><strong>Implementation and integration cost.</strong> Does it need difficult integrations, data cleanup, big system changes, or many departments at the table? A high score means it’s hard to implement.</li><li><strong>Cost of running the technology.</strong> Will the solution be expensive to run day to day, e.g. LLMs, OCR, image analysis, infrastructure, or exception handling? A high score means high operating cost.</li><li><strong>Risk of errors.</strong> What happens if the system gets it wrong? A high score means serious consequences of a mistake.</li></ul><p>The first three criteria show the potential. The higher, the better.</p><p>The last three show cost and risk. The higher, the more carefully you should treat the process as a first project.</p><p>The point is not to turn the matrix into math for the board. It’s something simpler: to stop basing the automation conversation on whoever talks loudest about their problem.</p><h3>The matrix in action</h3><p>Say the company is weighing three processes.</p><p>Score each one from 1 (low) to 5 (high). The first three scores describe the potential. The last three describe the cost and risk.</p><p><strong>1. Classifying and pre-checking documents</strong></p><ul><li>Business impact: 4</li><li>Volume and repeatability: 5</li><li>Speed of first results: 4</li><li>Implementation cost: 3</li><li>Running cost: 2</li><li>Error risk: 2</li></ul><p><strong>Verdict: Good first candidate.</strong></p><p><strong>2. A core financial process</strong></p><ul><li>Business impact: 5</li><li>Volume and repeatability: 4</li><li>Speed of first results: 2</li><li>Implementation cost: 5</li><li>Running cost: 4</li><li>Error risk: 5</li></ul><p><strong>Verdict: Important, but better later.</strong></p><p><strong>3. Analyzing rare, long documents</strong></p><ul><li>Business impact: 3</li><li>Volume and repeatability: 1</li><li>Speed of first results: 2</li><li>Implementation cost: 4</li><li>Running cost: 4</li><li>Error risk: 5</li></ul><p><strong>Verdict: Bad first candidate.</strong></p><p>This comparison isn’t meant to replace a proper analysis. It’s meant to help with the first decision.</p><p>It makes one thing plain: the financial process may have the biggest impact, but it also has a high implementation cost and a high risk of errors. That doesn’t mean it isn’t worth automating. It means it may not be the best place to learn.</p><p>Document classification, on the other hand, may not be the most strategic process in the company, but it has good volume, quick results and contained risk. That’s often the better first project.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UrCN3pAnhhcx--mn21xJHA.png" /></figure><h3>How to read the result</h3><p>After scoring a few processes, three groups of candidates usually emerge.</p><h4>The best first candidate</h4><p>A process with high or medium impact, high repeatability, quick first results, a reasonable implementation cost and contained error risk.</p><p>These are often processes built around document classification, initial triage of requests, reading data out of repetitive documents, drafting responses for a human, or cleaning up data before a decision.</p><p>They’re not always the most strategic, but they’re good for learning, building trust, and showing a result.</p><h4>A good candidate, just not first</h4><p>A process with high impact but also a high implementation cost or a high risk of errors.</p><p>It may well be worth automating. Just come back to it later: once the company has experience, test data, monitoring, better-described exceptions and more confidence in how to work with AI.</p><h4>A bad first candidate</h4><p>A process with low volume, expensive errors, difficult integration and a result that’s hard to measure.</p><p>Even if it sounds attractive, it can be a poor choice for a first automation. The risk is too high and the chance of a quick, measurable win too small.</p><h3>What the examples teach</h3><h3>Document verification as a hybrid process</h3><p>Manual document verification is a good example.</p><p>A company was receiving thousands of documents, all checked by hand. The natural idea was: let’s automate document verification.</p><p>Analysis showed, however, that not all documents were equally suited to automation. Some were standard and repetitive. Others were more complicated, less predictable, or needed interpretation.</p><p>Fully automating everything would have been expensive and risky. A hybrid setup turned out to be the better answer.</p><p>The scheme was simple:</p><ol><li>A document enters the system</li><li>The system classifies it</li><li>A simple, standard case is verified automatically</li><li>A harder case gets a preliminary classification from the system</li><li>A human gets it for a quick review and can confirm or change the call</li></ol><p>The result was practical: human work dropped sharply where documents were repetitive, and nobody tried to force automation onto the hard cases.</p><p>That’s an important lesson for a first automation project:</p><blockquote><em>You don’t have to automate 100% of cases for the project to make business sense.</em></blockquote><p>Sometimes the best outcome comes from a split: the system handles the standard cases, the hard ones go to a human.</p><h3>AI doesn’t always mean an LLM</h3><p>Choosing the process is one thing. Choosing the technology is another.</p><p>In one of the processes we analyzed, the documents looked like a natural fit for language models. They varied, some data came from photos, and the content needed analysis.</p><p>A closer look showed that despite the differences, the documents had a structure repetitive enough that cheaper algorithms, rules and heuristics could deliver the expected result more cost-effectively than a solution built mainly on an LLM.</p><p>The takeaway is simple:</p><blockquote><em>Don’t start by asking how to use an LLM. Ask what the simplest solution is that will deliver the expected result.</em></blockquote><p>We cover that in its own article: <a href="https://codino.pl/blog/do-you-really-need-llm">when you really need an LLM</a>, and when simpler automation beats language models.</p><h3>Not every accuracy point is worth its price</h3><p>In another process, automating document classification paid off at around 95% accuracy.</p><p>The score could have been pushed higher, but the cost of getting there was greater than the cost of handling the remaining errors.</p><p>That matters, because the goal of automation is not always maximum possible accuracy. The goal is the best business outcome.</p><p>In one process, 95% can be an excellent result if the remaining errors are cheap to catch and fix. In another, even 99% can be too little if a single mistake carries heavy consequences.</p><p>That’s another topic worth its own discussion: testing AI, the cost of errors, and what accuracy level is actually acceptable.</p><h3>When not to start with automation, even if it’s technically possible</h3><p>Not every process that can be automated should be the first project.</p><p>We analyzed a process built around long text documents. Based on their content, specific events mentioned in the documents had to be recorded.</p><p>Technically, automation was possible. The problem was elsewhere.</p><p>There were relatively few documents, and the cost of a single error was very high. A mistake could lead to consequences bigger than the potential savings from automation. Full automation made no sense as a first step.</p><p>It’s a good example of a process that can look attractive technologically but scores badly on the impact, cost and risk matrix.</p><p>Low volume, expensive errors and difficulty judging quality are warning signs.</p><p>In a case like this, it’s better to start by supporting the human: summarizing the document, flagging relevant passages, drafting a recommendation, or checking completeness. Fully automating the decision can wait.</p><h3>When can you start with a core process?</h3><p>The rule isn’t “never start with an important process”.</p><p>There are situations where starting with a core process makes sense, but specific conditions have to hold.</p><p>The process should be well described. The company should know when it starts, when it ends, what data it needs, and what a correct result means.</p><p>It should have high volume or very clear business impact. If a process is strategic but rare, it will be hard to show results quickly.</p><p>The automation should be possible in stages. Better to start with classification, recommendations, data preparation or human support than with fully automated decisions.</p><p>A human should stay at the critical points of the process, at least at the beginning. That limits the risk and generates data about how the system behaves.</p><p>And the board has to accept a pilot approach. If the expectation is full automation of a strategic process from day one, the risk of disappointment is high.</p><p>Under those conditions a core process can be a good candidate. Not because it’s the most important, but because despite its importance it meets the conditions for a safe start.</p><h3>Processes I wouldn’t pick first</h3><p>Some processes may be important but are rarely a good first automation project.</p><p>Be especially careful with ones that combine several of these traits:</p><ul><li>low volume</li><li>a high cost per error</li><li>lots of exceptions</li><li>no test data</li><li>unclear criteria for a correct result</li><li>difficult integrations across many systems</li><li>a requirement to fully automate the decision</li><li>results that only show after many months</li><li>serious legal, financial or reputational risk</li></ul><p>That doesn’t mean such processes should never be automated. They’re often worth tackling, but usually once the organization has its first implementations behind it, understands its data better, and knows how to judge the quality of a system’s work.</p><h3>What to do before choosing the first process</h3><p>The simplest exercise looks like this: pick three processes that passed the <a href="https://medium.com/codino-pl/ai-wont-fix-a-chaotic-process-first-check-you-have-something-to-automate-254d47753f7d">initial readiness test</a>. For each of them, answer six questions:</p><ol><li>What real business impact would automation deliver?</li><li>How often does the process repeat?</li><li>Can you show a first result within 30–60 days?</li><li>How hard will implementation and integration be?</li><li>What will the technology cost to run?</li><li>What happens if the system gets it wrong?</li></ol><p>Then drop the processes that combine a high cost of errors, low volume and difficult integration.</p><p>From the rest, choose one that delivers a visible result while letting the organization learn without excessive risk.</p><p>That’s usually your best first candidate.</p><h3>Build the ability to automate first</h3><p>The first automation shouldn’t be a show of force. It should be a well-chosen project that builds trust, skills, and data for the implementations that follow.</p><p>If the first project is too big, too risky or too hard, the company may draw the wrong conclusion: that automation “doesn’t work here”. Meanwhile the problem may not be the technology at all. It may be a badly chosen first process.</p><p>So don’t start with the biggest problem just because solving it would make the biggest impression.</p><p>Start with a process that has a good ratio of impact to cost and risk.</p><p>If a smaller, repetitive process can be automated and delivers concrete benefits, that’s a far stronger argument for automating the core processes than any slide deck full of promises.</p><p>Only once you’ve chosen a good first candidate is it worth moving to the next question:</p><blockquote><em>Does automating this process actually pay off?</em></blockquote><p>And that calls for <a href="https://codino.pl/blog/ai-automation-business-case">simple ROI math</a>: the cost of the work, the cost of errors, the cost of implementation, maintenance, and running the technology.</p><p><em>This article is part of </em><strong><em>From Decision to Deployment</em></strong><em>, a four-part series on making better AI automation decisions.</em></p><p><em>Previous: </em><a href="https://medium.com/codino-pl/ai-wont-fix-a-chaotic-process-first-check-you-have-something-to-automate-254d47753f7d"><em>AI Won’t Fix a Chaotic Process. First Check You Have Something to Automate</em></a></p><p><em>Next: </em><a href="https://codino.pl/blog/ai-automation-business-case"><em>Does Automation Pay Off? How to Build the Business Case Before You Invest in AI</em></a></p><p><em>Originally published on </em><a href="https://codino.pl/blog/ai-process-decision"><em>Codino</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ab95653c34e8" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codino-pl/the-best-first-ai-automation-project-isnt-your-biggest-one-ab95653c34e8">The Best First AI Automation Project Isn’t Your Biggest One</a> was originally published in <a href="https://medium.com/codino-pl">Codino.pl</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Won’t Fix a Chaotic Process. First Check You Have Something to Automate]]></title>
            <link>https://medium.com/codino-pl/ai-wont-fix-a-chaotic-process-first-check-you-have-something-to-automate-254d47753f7d?source=rss-b98bb43fc7c6------2</link>
            <guid isPermaLink="false">https://medium.com/p/254d47753f7d</guid>
            <category><![CDATA[business-process]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[business-strategy]]></category>
            <dc:creator><![CDATA[Jan Gajowiak]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 20:28:39 GMT</pubDate>
            <atom:updated>2026-10-12T20:28:39.179Z</atom:updated>
            <content:encoded><![CDATA[<h3>Before investing in AI, find out whether your workflow is actually ready to be automated.</h3><p>In many companies, the AI conversation starts the same way.</p><p>Someone on the board notices that competitors are experimenting with automation. Employees are already using ChatGPT. LinkedIn keeps serving up stories about companies “saving dozens of hours a month”.</p><p>Sooner or later, the question lands on the table:</p><blockquote>Where could we use AI?</blockquote><p>It’s a reasonable question. The problem is that it usually arrives one step too early.</p><p>Before a company starts choosing a tool, a model, an integration or a vendor, it should answer a less glamorous but more important question:</p><blockquote>Do we actually have a process that’s fit for automation?</blockquote><p>Because AI won’t tidy up your company for you.</p><p>It can speed up work, help analyze data, draft a recommendation or move information between systems. But if nobody in the company can clearly explain how a process works today, automation very often doesn’t solve the problem. It just moves the chaos around faster.</p><p>In practice, the first problem is rarely “we don’t have AI”. It usually looks like this:</p><ul><li>the process exists only in a few people’s heads</li><li>everyone runs it slightly differently</li><li>the company knows the expected outcome, but can’t describe the path that leads to it</li></ul><p>That’s exactly why so many implementations start too early.</p><h3>Start with the process, not the technology</h3><p>“Where could we use AI?” sounds like the right question, but it easily leads you astray. You end up hunting for a use case for the technology instead of solving a specific business problem.</p><p>The result is predictable. The company deploys AI where all it really needed was:</p><ul><li>a better form</li><li>a simple integration</li><li>cleaner data</li><li>or a plain business rule</li></ul><p>A few months later, nobody can answer:</p><ul><li>what exactly improved</li><li>how much time was saved</li><li>whether the project made sense at all</li></ul><p>The question that should come first is a different one:</p><blockquote>Which process in the company is repetitive, wears people down, produces errors, or eats up time on work nobody should still be doing by hand?</blockquote><p>Less fashionable, far more useful.</p><p>In practice, the place to start is not choosing an AI model. It’s mapping the process:</p><ul><li>what comes in</li><li>what should come out</li><li>what the main steps are</li><li>where the exceptions show up</li><li>where a human genuinely has to make a call</li></ul><p>Only then can you see what the company actually needs.</p><p>Sometimes it will be an LLM. Sometimes OCR. Sometimes an integration with the CRM or ERP. And sometimes the answer turns out to be much simpler: first, get everyone working the same way.</p><p>The best automation often isn’t in the most obvious process.</p><h4>The process doesn’t have to be simple. It has to be repeatable</h4><p>A common mistake is assuming that only simple, linear tasks are fit for automation. That’s just not true.</p><p>A process can be complex. It can have plenty of exceptions, span several systems, and include moments where a human has to weigh the situation. It can still be a good candidate. The one condition: you have to be able to describe it.</p><p>The trouble starts when:</p><ul><li>every employee does the task their own way</li><li>the decision criteria are fuzzy</li><li>the outcome depends mostly on one person’s experience</li></ul><p>A well-described process doesn’t have to be perfect. It’s enough to know:</p><ul><li>when it starts</li><li>when it ends</li><li>what data it needs as input</li><li>what result it should produce</li><li>which steps repeat almost every time</li><li>where a human decision is required</li></ul><p>If you don’t know these things, you don’t have a technology problem. You have an operations problem, and you need to see it first.</p><h4>Three signs the company doesn’t have a process yet</h4><p>Not every recurring problem is a process ready for automation. Sometimes a company just has an area of work that people handle on experience, intuition and personal workarounds.</p><p>There are three warning signs.</p><h4>1. Everyone describes the task differently</h4><p>If three people from the same team explain the same task in three different ways, the company doesn’t have a shared process yet. It has several private ways of working.</p><p>That doesn’t rule out automation, but it means you first need to decide:</p><ul><li>which way of working is the right one</li><li>or design a new, shared process</li></ul><p>In many companies, the first sensible process description only appears when someone from outside starts asking simple questions:</p><ul><li>where do you get the data from?</li><li>what do you check?</li><li>when do you consider the case closed?</li><li>what do you do with an exception?</li></ul><h4>2. Nobody can say what “done well” means</h4><p>This comes up all the time with:</p><ul><li>document analysis</li><li>lead qualification</li><li>ticket handling</li><li>preparing quotes</li></ul><p>At first everyone says “come on, it’s obvious”. Then, the moment you try to write down the quality criteria, it turns out everyone understands them a little differently. And if you can’t say what a correct result looks like, you can’t say what the automation is supposed to get right.</p><h4>3. Exceptions outnumber the standard case</h4><p>Every process has exceptions, and that’s perfectly normal. The problem starts when exceptions become the norm. If most cases need individual judgment, automating the whole process is going to be risky.</p><p>At that point, it’s better to go one level down and automate a fragment:</p><ul><li>classifying the incoming request</li><li>gathering the data</li><li>spotting what’s missing</li><li>comparing documents</li><li>or preparing a recommendation for a human</li></ul><p>And very often that alone delivers a real result.</p><h3>Don’t automate a wish. Automate a way of working</h3><p>Take a simple example: finding new customers or suppliers.</p><p>The sentence:</p><blockquote>We want to automatically find customers online.</blockquote><p>sounds good, but it’s far too vague.</p><p>It doesn’t say:</p><ul><li>what kind of customers you’re looking for</li><li>which sources you’ll use</li><li>what criteria decide whether a result is good</li><li>what’s supposed to happen next</li></ul><p>That’s a wish, not a process.</p><p>A much better description looks like this:</p><blockquote>Every day we pull a list of companies from specific industries, check their location, size, business profile, decision makers and fit with our offer. Then we score each company against agreed criteria, save the result in the CRM, and draft a first outreach message for the sales rep.</blockquote><p>Now that’s the beginning of a process. It doesn’t have to be perfect, but it has defined:</p><ul><li>inputs</li><li>criteria</li><li>steps</li><li>an output</li></ul><p>Only a description like that can be seriously analyzed for automation.</p><p>AI works far better on a described way of working than on a vague ambition like:</p><ul><li>“we want to sell more”</li><li>“let’s improve support”</li><li>“let’s do something about the documents”</li></ul><h4>Task, recommendation, or decision?</h4><p>Early on, it’s worth being clear about what exactly the company wants to hand over to a system. This is not a detail, because the risk of the entire project depends on it.</p><h4>Automating a task</h4><p>The system performs a repetitive step:</p><ul><li>extracts data from a document</li><li>moves information between systems</li><li>classifies a request</li><li>detects missing fields</li></ul><p>This is usually the safest place to start.</p><h4>Automating a recommendation</h4><p>The system suggests a decision, but a human still signs off on it.</p><p>It can:</p><ul><li>rank a lead</li><li>draft a reply to a customer</li><li>propose a category for a complaint</li><li>summarize a contract</li></ul><p>This is the right model where AI is meant to speed up work while accountability stays with a person.</p><h4>Automating a decision</h4><p>The system, on its own:</p><ul><li>approves</li><li>rejects</li><li>blocks</li><li>escalates</li><li>or triggers the next step of the process</li></ul><p>And this is where things get serious. Automatically extracting data from an invoice is a completely different project from automatically rejecting a complaint, or making a call that affects a payment, a customer or legal exposure.</p><p>The closer you get to a business, financial or legal decision, the more these things matter:</p><ul><li>quality criteria</li><li>auditability</li><li>a full history of actions</li><li>clear rules for handing a case over to a human</li></ul><p>AI can support decisions, but the company has to know who is accountable for them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uCx5aWACGixT0r2QSFKhpQ.png" /></figure><h4>You don’t have to automate the whole process</h4><p>The second common mistake is thinking that automation means taking over the entire process end to end. It usually doesn’t, and it usually shouldn’t.</p><p>A process can be too big, too changeable, or too dependent on people to automate in full. But it can still contain fragments that are repetitive, well described and easy to measure. And that’s where the best first result usually comes from.</p><p>Fragments like:</p><ul><li>pulling data together from several sources</li><li>an initial classification of a request</li><li>checking a document for correctness</li><li>detecting missing information</li><li>comparing offers</li><li>drafting a first version of a reply</li><li>moving data between systems</li><li>preparing a recommendation for a human</li></ul><p>In practice, the bigger return often comes not from trying to automate everything, but from picking one specific fragment of a process well.</p><p>Small. Repeatable. Measurable.</p><h4>Example: customer support</h4><p>The company says:</p><blockquote>We want to automate customer support.</blockquote><p>Still too broad.</p><p>Customer support can cover:</p><ul><li>complaints</li><li>order status questions</li><li>invoice requests</li><li>data changes</li><li>technical questions</li><li>returns</li><li>escalations</li><li>cases that need a manager’s decision</li></ul><p>Trying to automate all of it at once is going to be risky. Better to start by checking what actually repeats.</p><p>In many companies, it turns out that a large share of tickets boils down to a few topics:</p><ul><li>order status</li><li>resending an invoice</li><li>updating contact details</li><li>missing information in the request</li></ul><p>Then the first step doesn’t have to be “AI for all of support”. It can be:</p><ul><li>classifying tickets</li><li>detecting missing data</li><li>drafting replies for the agent</li></ul><p>That doesn’t replace the support team, but it removes repetitive work from a fragment of the process you can clearly describe. And that’s exactly what a good first implementation is about.</p><h3>A problem that’s too broad has to be narrowed down</h3><p>Plenty of automation ideas sound great until you try to write them down. “Let’s grow sales.” “Let’s find new customers.” “Let’s improve customer support.” “Let’s do something about the documents.”</p><p>These aren’t bad goals. They’re just too broad to be a starting point for automation.</p><p>A process sounds different:</p><ul><li>classifying leads by fit with the offer</li><li>analyzing RFQs and drafting a first response</li><li>routing customer tickets to the right categories</li><li>reading data from invoices and matching it against orders</li><li>preparing a weekly sales summary from CRM data</li></ul><p>A goal says what the company wants to improve. A process says what happens, step by step. AI needs the second one.</p><h3>A quick test: process or chaos?</h3><p>Before the company starts talking tools, models and vendors, take one process and answer a few questions.</p><ol><li>Does this process repeat in a similar way each time?</li><li>Do you know when it starts?</li><li>Do you know when it ends?</li><li>Do you know what data it needs as input?</li><li>Do you know what result it should produce?</li><li>Can you describe the main steps?</li><li>Do different people run it in a similar way?</li><li>Do you know which decisions are simple and which need a human?</li><li>Do you know where the errors usually happen?</li><li>Could you start by automating just one fragment?</li></ol><p>If the answer to most of these is “yes”, the process is probably worth analyzing further.</p><p>If most answers sound like:</p><ul><li>“I don’t know”</li><li>“it depends”</li><li>“everyone does it a bit differently”</li></ul><p>then automation is premature. Not because AI isn’t up to it, but because the company doesn’t yet know what exactly is supposed to be automated.</p><h4>A 30-minute mini-audit</h4><p>The simplest readiness test needs no workshops, no slide decks and no tool selection.</p><p>Pick one process that weighs on the team today and answer four questions:</p><ol><li>How does this process start in practice?</li><li>What specific result should it produce?</li><li>Which three steps repeat almost every time?</li><li>Where is a human most often needed?</li></ol><p>If you can’t answer these in 30 minutes, the process isn’t ready for automation yet. It needs to be described first.</p><p>If the answers are clear, you can move on: point to the fragment that is the most repetitive, the least risky and the easiest to measure.</p><h4>What usually comes out of this analysis?</h4><p>Most often, one of three scenarios.</p><h4>1. The process is ready for automation</h4><p>It has:</p><ul><li>inputs</li><li>an output</li><li>repeatable steps</li><li>clear points where a human decides</li></ul><p>Then you can look for the best fragment to start with.</p><h4>2. The process needs tidying up first</h4><p>Different people run it differently. The decision criteria are fuzzy, or nobody can say what a good result is supposed to look like.</p><p>Then the work starts with straightening out the process, not with deploying AI.</p><h4>3. The process is too chaotic</h4><p>It has:</p><ul><li>no clear beginning</li><li>no end</li><li>no defined inputs</li><li>no defined output</li></ul><p>The company can’t say how the task is done today. Then AI is not the first step. The first step is understanding what is actually going on inside the company.</p><p>That may sound less exciting than a quick tool rollout, but it usually saves a lot of money and frustration.</p><h3>Process first, AI second</h3><p>AI can be a very good automation tool. It can:</p><ul><li>speed up work</li><li>reduce errors</li><li>analyze documents</li><li>clean up data</li><li>support people in making decisions</li></ul><p>But it won’t replace thinking about the process.</p><p>If the company doesn’t know:</p><ul><li>how the task is done today</li><li>who makes the decisions</li><li>what data is needed</li><li>what a good result means</li></ul><p>then automation is going to be risky. Not because AI is bad, but because nobody knows what it’s actually supposed to do.</p><p>A good first step is to pick the two or three processes that weigh on the team the most and run them through a simple checklist.</p><p>Only once you know which processes are ready for automation is it worth moving on to the next question:</p><blockquote>Which of these is worth automating first?</blockquote><p><em>This article is part of </em><a href="https://codino.pl/blog/ai-process-automation?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=jg_medium_ai_process_automation&amp;utm_content=01_fix_process"><strong><em>From Decision to Deployment</em></strong></a><em>, a ten-part series on making smarter AI automation decisions, published by Codino.</em></p><p><em>Next in the series: </em><a href="https://codino.pl/blog/ai-process-decision?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=jg_medium_ai_process_automation&amp;utm_content=01_fix_process"><strong><em>Don’t Automate Your Most Important Process First. Pick One the Company Can Learn On</em></strong></a>.</p><p><em>Originally published at </em><a href="https://codino.pl/"><em>Codino</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=254d47753f7d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codino-pl/ai-wont-fix-a-chaotic-process-first-check-you-have-something-to-automate-254d47753f7d">AI Won’t Fix a Chaotic Process. First Check You Have Something to Automate</a> was originally published in <a href="https://medium.com/codino-pl">Codino.pl</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>