When not to build AI: 9 cases where no is the right answer

When should you not build an AI project?

Do not build when nobody can write the process down, when the master data is dirty, when no success number exists, when the volume is small, when a single error is expensive, when there is no owner, when a legal question is open, when the process is about to change, or when a plain rule would do the job. Nine cases, each with a cheaper next step.

We read through twelve Hungarian AI vendor sites in August 2026. All twelve communicate success and nothing else. Not one describes the conditions under which the thing falls over, which from a buyer’s seat is the more useful half. A no is much easier to reverse than an abandoned rollout.

There is no country-level failure statistic for our market. We looked: neither the statistical office, nor the IT trade association, nor the large consultancies publish one. What follows comes from global surveys and from the requests for proposal that reach us, and the second group is not a representative sample. Where a figure is an estimate, I say so.

Can anyone write down the process you want to automate?

If two colleagues describe the same sequence differently, you do not have a process, you have a habit. AI does not make a habit smarter, it just produces the wrong output faster. Write the process down before you automate it, otherwise there is nothing to measure the errors against.

The symptoms are easy to spot. There is no written flow chart, or there is one and nobody has opened it in three years. Every case has an exception. The answer to why do we do it this way is that we always have.

An agent needs a decision rule. If the rule lives in three heads with three different contents, you cannot assemble an evaluation set, which means you cannot say whether the system is wrong. A project like this does not crash. It drifts for months until somebody says out loud that we do not know if it works.

The cheaper step is process redesign. McKinsey’s 2026 AI survey reports that EBIT impact correlates most strongly with rebuilding the process from the ground up rather than layering AI onto the old one (McKinsey State of AI, 2026; worth checking the figures on the source page, since summaries quote them inconsistently). Two weeks of process mapping costs a fraction of a pilot that answers the wrong question.

Is the master data behind the AI clean enough?

If the same partner appears under three names, half the product master fields are empty, and somebody patches the report in a spreadsheet afterwards, AI will not hold on that ground. Retrieval returns the wrong record, the model answers from it with confidence, and the mistake is harder to see than an empty cell.

There is measured evidence here, at least in English. In Anthropic’s own testing plain embedding search had a 5.7% retrieval failure rate, which dropped to 1.9% with hybrid search plus reranking (Anthropic, 2024-09-19). That is on a clean English corpus. For smaller European languages there is no maintained public retrieval benchmark at all. The last independent test we could find for Hungarian measured substantial overlap between the similarity distributions of related and unrelated sentence pairs (Péter Harang, 2025-01-09). So the starting point outside English is worse, and without your own measurement nobody can tell you by how much.

The cheaper step is rarely AI. Often it is reworking the input form: mandatory fields, dropdowns, validation at the point where the data is created. In accounts payable, data entry accounts for 30 to 35% of the total processing cost (IOFM citation, 2026-07), while the average invoice costs 9.40 dollars to process and 2.78 dollars at best-in-class organisations (Ardent Partners, quoted by WEX, 2026-07-20).

Can you say in advance what counts as success?

If the goal of the project is to have AI, the project has already failed, it just does not know yet. Without a success number there is nothing to decide against in month six, and nothing to show the person who released the budget. You need one number, one baseline, one date.

The MIT Project NANDA study from summer 2025 reviewed more than 300 announced deployments, 52 interviews and 153 executive surveys, and found that 95% of generative AI pilots produced no measurable P&L impact (The GenAI Divide, 2025-07). According to the researchers the model was not the weak link. The tools did not learn from corrections, forgot context between sessions, and lived outside the systems where the work actually happens.

The practical test is short. Write down in one line what you are measuring, what it is today, and when you will look again. If you cannot state today’s value, then the first project is measurement, not AI. That usually takes two weeks, and at the end of it you know whether the build is worth starting. Our AI ROI calculator gives you a frame for the payback maths, but the input numbers have to be yours.

Is the volume high enough to pay back?

At a few dozen cases a month, the build and the maintenance usually cost more than the time saved. This is not a matter of opinion, it is multiplication. Monthly case count times minutes saved per case, converted with your own internal hourly cost, compared against the operating fee.

Our own price list puts a simple one to three step automation at 100,000 to 500,000 Ft, a multi-system automation with error handling at 500,000 to 2,500,000 Ft, an AI-driven complex process at 2,500,000 to 10,000,000 Ft, and operations at 30,000 to 150,000 Ft a month. Converted at a rounded 385 HUF per euro, that is roughly EUR 260 to 1,300, EUR 1,300 to 6,500, EUR 6,500 to 26,000, and EUR 78 to 390 a month. If the monthly saving does not clear the operating fee, you never need to price the build.

At low volume even the most expensive tool is cheap. Zapier’s 19.99 dollar Pro plan includes 750 tasks, which in a six-step workflow is 150 runs, about 13 cents per run (Zapier pricing, retrieved 2026-08-14). At 150 cases a month that is entirely reasonable. The same step count at 40,000 runs is the range where moving into code starts to pay. Our AI agent and chatbot pricing breakdown works through the line items.

How much does a single bad decision cost?

If a wrong output moves money, reaches a customer, or deletes something in a live system, the question is not model accuracy but who signs it off. Language model output is probabilistic. You can drive the error rate down. You cannot drive it to zero.

There is a useful rule from the security side too. Simon Willison’s lethal trifecta says an agent becomes exploitable when it has access to private data, meets untrusted content, and can communicate externally at the same time (Willison, 2025-06-16). If all three are present in the system you are planning, and the cost of an error is high on top of that, the project cannot start in that shape.

The answer is not to put a human in front of everything and then act surprised about the payback. If every output needs full review, you have bought back drafting time rather than decision time, and that is how it should be priced. It often still pays. It just has to be said at the start of the project instead of in month four.

Is there a named person who checks it weekly?

If the owner is IT as an abstraction, there is no owner. You need somebody whose weekly work includes reviewing failed outputs, and who has the authority to change the rule when the rule turns out to be wrong. Without that the system degrades quietly, and nobody notices for months.

The same MIT NANDA study measured its sharpest difference exactly here: 67% success rate with an internal expert paired with an external partner, 22% for projects built by internal IT alone. The gap between those two is not a nuance.

For us this has moved into the first few scoping questions. If the answer is that somebody will look at it, the pilot starts, the first wave of exceptions arrives in month three, and there is nobody to decide what the correct behaviour is. If you are choosing a vendor right now, our criteria for picking an AI development partner cover the same question from the other side of the table.

Is there a legal or compliance blocker?

Two things are worth settling before development starts: whether the data can leave its current environment at all, and who is responsible for the system in the language of the regulation. If either answer is uncertain, a classification and documentation round comes before the code, not after it.

The timing in August 2026 misleads a lot of people. The transparency obligations under Article 50 of the EU AI Act apply from 2 August 2026, so a chatbot has to disclose that the user is talking to a machine and synthetic content has to be marked in a machine-readable way. The high-risk deadlines, however, moved with the Digital Omnibus regulation to 2 December 2027, and to 2 August 2028 for systems embedded in products ((EU) 2026/1744, FPF timeline, 2026-07-28). Plenty of articles still quote the old date.

The penalty ceiling alone justifies clearing this up. Breaching provider, deployer or Article 50 obligations can reach 15,000,000 euro or 3% of global turnover, prohibited practices 35,000,000 euro or 7%, whichever is higher. For SMEs the cap is the lower of the two figures (AI Act Article 99). A four to eight week classification and documentation round is cheap insurance against that.

One limit worth stating plainly: there is no mutual recognition between compliance regimes. A NIS2 audit or an ISO certificate does not cover the AI Act. Overlapping controls can be reused. The compliance itself stays a separate exercise.

Will the process change within a few months?

If the process is in the middle of an ERP migration, a reorganisation or a regulatory change, the agent you build now will have to be written twice. The cost of the first version is not an investment, it is tuition. The right move is to wait and clean data in the meantime.

Our own market gives a good example of what that looks like in practice. Hungarian receipt data reporting becomes mandatory on 1 September 2026 for businesses using manual receipt books and computer-generated receipts, within three calendar days of issue (NAV). The legacy filing client shuts down on 31 December 2026 and eVAT becomes mandatory on 1 January 2027 (Andersen, 2026-02-19). The eVAT machine-to-machine production environment has run on the 2.0 data structure only since 3 August 2026, with no parallel operation (NAV, 2026-08-03). Anyone wiring AI into their current invoicing flow this month has a fair chance of starting over in January.

It is also worth being realistic about how long platform changes take. Surrey County Council’s ERP replacement went from 16.6 million pounds to 27.9 million, and from 15 months to 33 (The Register, 2024-07-16). Different country, and not an AI project, but the lesson carries: while the base system is moving, the target of anything built on top of it is moving too.

Would a deterministic rule solve it?

If if-X-then-Y covers most of the cases and the input is already structured, you do not need an agent, you need a coded workflow. Cheaper, testable, and it will not run forty pointless loops overnight. AI enters at a single point in that design: where free text has to be understood.

This is not our opinion. Anthropic’s own engineering guidance says a pre-coded workflow beats an autonomous agent for most tasks, and that agents are worth choosing only when the number of steps cannot be predicted in advance (Building Effective Agents, 2024-12-19).

The difference shows up on the invoice. The same 1,000 documents cost 8.70 dollars with a single extraction call and 232 dollars through an eight-step agentic loop, on the same model. Twenty-seven times the price, and it is the number of turns that decides it rather than the model. The calculation comes from the official Claude price list, assuming 12,000 input and 500 output tokens per step.

Self-hosting is not free either, the cost just shows up somewhere else. n8n Cloud Pro is 50 euro a month, while managed n8n hosting in the Hungarian market starts at 40,000 Ft on a VPS and 80,000 Ft on your own hardware, roughly EUR 104 and EUR 208 (n8n.hu, retrieved 2026-08-14). On cost alone, self-hosting starts to make sense somewhere above 40,000 executions a month. Below that you need a data residency or compliance reason. Our comparison of chatbots, RPA, n8n and agents walks through all four approaches.

What happens in month six of a badly started project?

Nothing crashes. The system runs, and nobody is watching. The output has drifted from the demo, several people have edited the prompts, the run cost has grown, and the success number never appears because it was never agreed. This is the most common ending, far more common than a dramatic failure.

Price changes are not a theoretical risk. Google’s official price list shows the Gemini 3.6 and 3.7 Flash introductory rates holding until 31 December 2026 and doubling afterwards, from 0.75 and 3.75 dollars to 1.50 and 7.50 dollars per million tokens (Google). Any 2027 operating plan built on today’s prices is underfunded.

The token bill has surprised bigger organisations too. Uber burned through its full-year AI coding budget by April 2026, Priceline’s Cursor renewal came back at four to five times the price, and Jellyfish measured token consumption per developer growing 18.6 times over nine months (TechCrunch, 2026-06-05). A monthly budget cap does not protect you here, because it triggers after the expensive run has already completed. You need a per-run cost limit and a step limit, built in at the start.

Observability belongs in the same budget line. Self-hosted Langfuse runs under an MIT licence with no licence fee, the hosted Core plan is 29 dollars a month (Langfuse), and LangSmith Plus is 39 dollars per user with 14 days of base trace retention (LangChain). Those 14 days matter, because the question in month six is exactly what happened six weeks ago.

When is building an AI agent a good idea after all?

When the volume is high enough, the data exists in machine-readable form, errors are reversible, there is a named owner, there is one measurable number, and the process is likely to hold steady for a year. Five of those six conditions is enough for a pilot. With three of six, waiting is the better call.

On the partner question I am openly biased, so I will let the number speak: in the MIT NANDA survey the pairing of an internal expert with an external partner reached 67% success, against 22% for projects built by internal IT alone. A development company is writing this, so weigh it accordingly. The sample, at least, is not ours.

One example from our side of the good case. The locally running search system we built on the kultura.hu archive works because the conditions held: a large document set, content that already existed in machine form, reversible errors, and a source citation next to every answer so the user can check the machine. That project was part of our talk at AI Summit Budapest 2025.

And an honest limit. If what you need is a three-step form to CRM to email automation, that falls in the 100,000 to 500,000 Ft band, about EUR 260 to 1,300, and a good freelance n8n developer will put it together. We add something where several systems, error handling, permissions and compliance documentation are all in scope. If your task is the former, we tend to say so during the scoping call.

Summary and frequently asked questions

If you said yes to several of the nine, it does not mean AI would not work for you. It means the order is different: process, data, measurement, then model. If the conditions are in place, our 90-day rollout plan shows what a disciplined first round looks like.