Skip to content
All articles
Fundamentals 8 min read· by Lars Zimmermann

AI gives you a solution. How do you tell whether it fits?

A language model gives everyone the same plausible proposal. Five checks that show whether it fits your business, and what the standard requires first.

In short

A language model gives everyone the same plausible proposal, and it is usually right, just for someone else. Whether it fits your business depends on your binding constraints: legacy systems, key customers, process knowledge, certifications. Five checks take the proposal apart. ISO/IEC 42001 requires in A.6.2.2 that requirements for new AI systems are documented beforehand.

Auf Deutsch lesen: deutsche Fassung

Everyone gets an answer today. You type your problem into a language model and within seconds you receive a proposal that is neatly structured, sounds sensible and reads like a solution.

The proposal is usually correct, too. Just for someone else.

What has become scarce is the ability to recognise which of thirty sensible routes is the right one for your business. There are solid checks for that, and there is a requirement in the standard that applies at exactly this point.

Why the proposal is plausible and still does not fit

A language model produces the answer that is most probable given everything it has read. For a question that ten thousand businesses have asked before you, that is the average answer. It is correct for the average business.

Your business does not exist in that middle. What really determines your decision is almost always a single binding constraint: the 1998 machine without an interface. The customer who accounts for forty per cent of revenue and accepts no cloud. The colleague who is the only one with the release process in her head and who retires in two years.

None of these constraints is in the model. They are not even written down anywhere in your own company. And because you did not mention them, the proposal is calculated without them.

The obvious objection is: then I will simply mention them. That fails because hardly anyone knows in advance which of their peculiarities is the binding one. Every business has thirty quirks, and twenty-nine of them are irrelevant to this one decision. Which one is the thirtieth usually only becomes clear when the proposal breaks on it. That is exactly the difference between someone who knows the field and someone who has an answer: the former asks beforehand the three questions that bring the binding constraint to the surface.

That is why so many AI projects feel good and come to a halt after four months. Not because the technology failed, but because the decisive sentence was never written down.

Five checks that take a proposal apart

I put these five questions to every proposal, whether it comes from a model, a service provider or from inside the company. Together they cost twenty minutes.

  • Under what condition does this route fail? A proposal that does not name its own failure condition is advertising. Keep asking until one appears.
  • What does it cost to undo this? It is not the build that determines the risk, but the dismantling. Two days of dismantling are worth a try. Six months are not.
  • Who has to change their behaviour for this, and did that person have a say? Every proposal that presupposes the behaviour of people who were not at the table is a bet.
  • What does the failure case cost, not the normal case? Work through the day on which it goes wrong and nobody notices. That figure decides, not the saving when things go well.
  • Does the route withstand your worst day? Year-end closing, two people off sick, the big customer on the phone. A procedure that only works on Tuesdays is not a procedure.

The fourth question deserves a small calculation, because it almost always turns out differently than expected. A transaction saves one minute and runs thirty times a day, which is around a hundred hours a year. Now the other side: the transaction goes wrong in two out of a hundred cases, is noticed after six weeks on average and then costs half a day of rework plus an unpleasant phone call. At thirty transactions a day, two per cent is around twelve cases a month, so almost six hundred hours of rework a year. The calculation tips over, and not by a narrow margin.

The second question is the underrated one. In practice, dismantling costs are almost never discussed, and that is exactly where the dependency arises that becomes expensive later.

A case in which the thirtieth quirk decided

A supplier with around 90 employees wanted to speed up quotation preparation. The task was cleanly described: the enquiry arrives by email, the line items are captured, prices are pulled from costing, the quotation goes out. Four vendors, four plausible routes, plus several proposals from a language model. All came down to the same thing: text recognition on the enquiry, matching against the item master, pre-filling the quotation.

None of the routes held. The reason was in no concept paper: around a third of the enquiries arrived as photographed hand sketches from the construction site, some with dimensions in the margin. There was no text in them. The line items were nowhere in a form that could be matched. The clerk had been reading them for years because she knows the customers.

Once this one constraint was on the table, the matter sorted itself out in twenty minutes. Two routes dropped out immediately. One route was limited to the two thirds of clean enquiries, with a deliberate cut before it. The rest stayed with the clerk, and her knowledge was documented instead of trying to automate it.

The information had been available in the company the whole time. It was just that nobody had asked for it, because the task description said how things should be, not how they are.

The question that comes before the proposal

All five checks presuppose something that is missing in most businesses: an idea of what a good solution has to deliver in the first place. Without that yardstick you compare thirty proposals against your gut feeling, and then the best-worded one wins.

ISO/IEC 42001 has a control of its own for this, and it deliberately sits at the start of the life cycle. A.6.2.2 requires the requirements for a new AI system to be specified and documented. The objective of the section is to define criteria and requirements for each stage of the life cycle. In practice this means: objectives and boundary conditions are in place before development starts.

The implementation guidance in Annex B (B.6.2.2) names the rationale for the project, objectives, data requirements and requirements for all life cycle stages. Acceptance criteria and sign-off by the business owner are not mentioned there. I ask about them in the audit anyway, because without them nobody can say when the system is finished.

Whoever has this can sort thirty proposals in an hour. Whoever does not, discusses proposals.

One point that the standard does not prescribe and that I consider the most important: the specification is signed off by the business owner. Not by IT, not by the service provider, not by the person who will later build the system. Otherwise whoever builds writes the requirements so that they fit their solution. That is not ill will, it is the normal case. Separating requirement and solution is half the benefit of the whole control.

The same applies to comparing several vendors. If the specification exists, everyone receives the same document and responds to the same points. Without it, each vendor describes the task the way its product solves it best, and in the end you are comparing three different tasks with each other. That is the most common reason why comparing offers in the AI field is so laborious and why price then decides as the only remaining distinguishing feature.

You can only judge a proposal once you have written down beforehand what it has to be able to do. Before that you are comparing wording.

What I see at this point in the audit

Three questions from A.6.2.2 get the room moving in every audit:

  • Show me the requirements specification for this system. Why was it built, and what does that follow from?
  • Do the requirements cover the entire life cycle, including operation, maintenance and decommissioning?
  • When and under what circumstances do you review the requirements again?

The typical nonconformity almost always looks the same. There is a functional description, often even a good one. Requirements for operation, maintenance and decommissioning are missing. The consequence only shows later: the controls for deployment and operation cannot draw on any specified criteria, because there are none.

In plain terms: nobody can say when the system is good enough, when it has to be maintained and when it ought to be switched off. It simply keeps running until someone notices that it has been producing nonsense for months.

The question about decommissioning is the most uncomfortable and the most useful. Almost nobody has answered it. Whoever answers it has automatically also answered the question of how to recognise that the system is no longer worth it.

What you can do on Monday

Take the AI project that is currently furthest ahead in your company. One page is enough, handwritten is fine too.

  • Why are we doing this? One sentence that contains a number. Without a number it is a wish.
  • How do we recognise that it works? These are your acceptance criteria. At least one quantity that can be measured.
  • What must it not do? The binding constraints that nobody at the top knows. This is where the old machine, the customer without cloud and the colleague with the knowledge in her head belong.
  • Who signs this off? A name, not a department.
  • When do we switch it off? The condition under which the project is ended.

Take this page into the next conversation, whether with a vendor, a consultant or a language model. You will find that half of the proposals die at point three, and in two minutes instead of four months.

And if nothing occurs to you for point two, you already have the most important result in your hand. If you cannot find a quantity that would show success, the project is not yet ready for a decision. Then the next step is not technology but a measurement over two or three weeks. That feels like a delay and is the shortcut.

And you will notice how to recognise a good counterpart: they ask about points three and five before proposing anything. Whoever offers a solution immediately has the average answer on offer. That is no reason to turn them away. It is a reason to put the page on the table before they start.

Share: LinkedIn E-Mail

Frequently asked questions

Are proposals from a language model any use at all?+

Yes, as a collection of material. Within minutes a model gives you the common routes to a problem, including the variants that nobody in your own company thinks of. What it cannot provide is the selection, because for that it would have to know your binding constraints. Use it for breadth and decide on depth yourself.

What is a binding constraint?+

The one circumstance that decides between success and failure, regardless of how well the rest is solved. Typical in SMEs: a legacy system without an interface, a key customer with its own requirements, a person with irreplaceable process knowledge, a certification that demands particular evidence. These constraints are rarely written down anywhere, and that is exactly why every external proposal is calculated without them.

Does ISO/IEC 42001 require a specification for every small AI tool?+

A.6.2.2 applies to new AI systems and to material enhancements of existing systems. What counts as a material enhancement in your company is for you to define, and that is exactly what the audit asks about. For a tool that cannot harm anyone, a brief entry is enough. For a system that prepares decisions, I expect the full specification including acceptance criteria.

What belongs in the requirements that is regularly forgotten?+

Operation, maintenance and decommissioning. The functional description usually exists, the rest of the life cycle is missing. Without these requirements there are later no criteria for when to maintain, retrain or switch off. Added to this is the question of when the requirements themselves are reviewed again, for instance when the system does not work as intended.

Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)

Last updated: 10 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.

Questions about your own case?

In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.

Continue reading