--- title: "AI for Business: Where It Pays Off, and Where It Does Not" description: "Where AI actually pays off for a business and where it does not. Four questions that decide it, the patterns that work, and what it costs to start." canonical: "https://www.whatsnext-ai.com/blog/ai-for-business-where-it-pays-off" published: "2026-09-25T00:00:00.000Z" updated: "2026-09-28T07:57:39.795Z" --- # AI for Business: Where It Pays Off, and Where It Does Not Where AI actually pays off for a business and where it does not. Four questions that decide it, the patterns that work, and what it costs to start. ![Overhead view of a desk with a laptop, notebook and tablet, where a person uses a phone with one hand and a prosthetic bionic hand rests on the trackpad](https://fgzcpjbyiakhjifaciaj.supabase.co/storage/v1/object/public/media/pexels-shvetsa-5614259.jpg) AI pays off best on work that repeats often, has a demonstrably correct answer, and currently consumes expensive people's time. On one-off, judgement-heavy or undocumented work, you rarely earn it back. ### Why almost every article on this says the same thing Search for AI for business in 2026 and you get five opportunities, seven applications, ten possibilities. Productivity up, customer contact better, data analysis smarter. All true, and none of it helps you decide anything. What is missing is the other half: where it does not pay. That is the part you actually need, because the first project decides whether there is a second one. A first engagement that returns nothing costs you more than its budget. It costs you the internal appetite for the two years after it. ### Four questions that decide whether it pays Run them against every process you are considering, in this order. A process that scores well on all four is a candidate. If one fails, it usually is not. #### How often does it happen? Volume is the dominant factor, and not by a small margin. A task that occurs a thousand times a month earns back build costs that a task occurring five times a month rarely will. The time per instance matters less than people expect. Two minutes of work happening a thousand times is a better candidate than a two-day job you do once a quarter. Most teams instinctively pick the two-day job, because it feels like the bigger problem. #### Is there one demonstrably correct answer? If two experienced colleagues would answer the same question differently and both be right, the work is judgement-heavy. AI can support that, but it cannot take it over. If there is a correct answer, it is recorded somewhere, and finding it is merely laborious, you are in good territory. That distinction matters far more than how complicated the work looks from outside. #### Is it written down anywhere? AI has to draw on something. A process that lives only in one person's head can be automated, but the engagement then starts with writing it down. That is often valuable work in its own right. Just count it in the budget, because it is the line item most often forgotten. #### What does it cost today? Put a number on it before you start. Hours times rate, or revenue lost, or both. Without a baseline you cannot demonstrate afterwards that anything improved, and the second project becomes an argument instead of a decision. ### Where does AI pay off most for businesses? Four patterns that show up in almost every company: all four are high volume, have a demonstrably correct answer, and draw on a source that already exists. - Answering questions from a source too large to memorise: product catalogues, price lists, terms, manuals. - Pulling information out of documents and putting it to work elsewhere: invoices, quotes, forms, contracts. - Preparation for a human who then decides: summarising files, scoring leads, triaging requests. - Status work: establishing where something stands by checking several systems that do not talk to each other. What they share: high volume, a demonstrably correct answer, and a source that already exists. ### When is AI a bad idea? Five categories of work where AI usually does not pay off, and that most articles skip: - Work that happens rarely and carries little value per instance. Five times a year of low-stakes work does not earn back a build. Where the stakes per instance are high, low frequency can still pay. - Decisions with real consequences and no clearly correct answer. Dismissals, credit decisions, medical triage. AI can prepare these, it cannot make them. - Processes nobody can explain. If three colleagues give three versions of how it works, the process is the problem, not the absence of AI. - Work that depends on data you do not have. A model cannot know what was never recorded. - Anything your existing software will do next quarter. Do not build a summarisation feature for a platform that already has one on its roadmap. None of this is an argument against AI. It is the reason the projects that do go ahead succeed. ### What it looks like when it does fit An example with real unit economics rather than adjectives. [Borgh](https://www.whatsnext-ai.com/cases/borgh-answers-every-product-question-in-seconds-across-15000-skus) sells tools and fastening materials, more than 15,000 items. Product questions went through specialists and sometimes through an email loop with the supplier, which took time. We built an agent that answers those questions in seconds, with stock levels included. Cost per conversation came out at roughly €0.005, around 1,000 times cheaper than the same answer from a specialist. Same team, no extra advisors, and the engagement ran thirteen weeks. Note why it worked rather than the figure itself. High volume, a demonstrably correct answer, and a source that already existed as a catalogue. All four questions fell the right way. On a process that failed even one of them, the same build would have been a loss. ### What if it does not fit? Then the answer is not automatically to build. Sometimes the process itself needs replacing. Sometimes the only thing missing is that your people get better at the tools they already have. Training is the cheaper answer there. A half-day workshop costs €3,000, a company-wide programme €6,000 to €15,000. That is a fraction of a build, and it is the honest recommendation more often than you would expect from a firm that also builds. If you do build, a first automation starts at €5,000 and goes live in two to four weeks. Once something is running, maintenance sits in an ongoing partnership from €2,750 a month. What that covers, and what it does not, is set out under our [AI services](https://www.whatsnext-ai.com/ai-services). ### Starting without a pilot that stalls Pick one process that scores well on all four questions. Not the most interesting one: the dullest one with the highest volume. Measure what it costs today before anything gets built. Agree which number has to move. Put someone outside IT in charge, because they know the process and they feel the result. The full step-by-step for that first move is in our piece on [AI for small businesses](https://www.whatsnext-ai.com/blog/ai-for-small-business). And build it from day one as though it has to run in production, not as a demo. That is the difference between a project that ships and [a pilot that stalls](https://www.whatsnext-ai.com/blog/why-ai-pilots-fail). If the question is automating a whole process rather than a single task, [AI automation](https://www.whatsnext-ai.com/blog/ai-automation-for-business) is the broader picture. Want to know which of your processes scores highest? We are a Dutch AI partner that builds this, not just advises on it. [Book a free strategy call](https://www.whatsnext-ai.com/contact). We will go through your list, and tell you which ones are not worth it. ## Frequently asked questions ### How do I implement AI in my business? Start with one process that repeats often, has a demonstrably correct answer, and is already documented. Measure what it costs before you build, agree which number has to move, and put someone outside IT in charge. Build it production-ready from day one, not as a demo. ### Where does AI pay off most for businesses? On high volume with a demonstrably correct answer: answering questions from a large source such as a catalogue or price list, extracting information from documents, preparation for a human decision, and status work across systems that do not talk to each other. ### When is AI a bad idea? On work that happens rarely, on decisions with real consequences and no clearly correct answer, on processes nobody can explain, on work that depends on data you do not have, and on features your current software will ship itself before long. ### What is the difference between ChatGPT and AI? ChatGPT is one product; AI is the broader technology underneath it. For a business the more useful distinction is between generative AI, an assistant someone asks a question, and agentic AI, an agent that carries out a task on its own and is measured on the result. ### What is the best AI for business? There is not one. Which model works best depends on the task, and it shifts every quarter. We are deliberately model-agnostic: we pick per task and make sure you are not locked into a single vendor. ### What does it cost to start with AI? A half-day workshop costs €3,000 and a company-wide programme €6,000 to €15,000. A first automation starts at €5,000 and goes live in two to four weeks. Ongoing maintenance sits in a partnership from €2,750 a month.