Four reasons AI implementations do not deliver

A few days ago I announced that YES FOR is an OpenAI Select Partner. The natural next move is to sell that status. This piece does the opposite: it explains what the status does not fix.

I am writing it because a partner status answers the question „can they handle this technology". It does not answer the question that decides the outcome of an implementation: „is this company ready to implement anything at all".

What the data says

The most solid source is RAND Corporation, which analysed enterprise AI projects in 2024 and reported that roughly 80% of them fail to deliver the expected value. That is twice the failure rate of conventional IT projects.

I went through those studies in detail, together with the MIT and S&P Global figures and the methodological caveats, in Why 80% of AI projects fail. Here I am after something else.

Those studies are global. I do not have a Polish equivalent and I am not going to pretend otherwise. The four things below are observations from conversations and projects on the Polish side, not the result of a study. Treat them as a checklist, not as a statistic.

None of them is about the model.

1. Nobody measured what the process costs today

This is the most common and the most expensive omission.

A company deploys AI in handling requests for quotes. Three months later someone asks whether it paid off. Nobody can answer, because nobody knows what the process cost before. How many hours a month it took. At which stage. How many people. What share of requests were handled late.

Without that number there is no reference point. You can show that the tool works, that people use it, that answers come out faster. You cannot show that the company made money.

This is not a reporting problem. It is a decision problem: without a baseline nobody knows whether to scale or shut down. A board that gets no answer for two quarters usually shuts down.

A baseline is collected before the start and takes a few days. After the implementation it can no longer be reconstructed, because the process has already changed.

2. The data sits in several systems that do not talk to each other

The project assumes the data is available. At the workshop it turns out that it is — in the ERP, in the CRM, in a spreadsheet maintained by one person, in an inbox, and in the warehouse system. Four of those five sources give a different version of the same fact.

That is when the part of the project nobody costed begins: agreeing which source is right. This is not a technical task. Someone in the company has to decide that customer data from the CRM overrides the spreadsheet, and own that decision.

If that decision is missing, the AI learns from contradictory data and returns contradictory answers. Within two weeks the team stops trusting the tool, and the project dies quietly, without anyone formally closing it.

Integration is the consequence here, not the cause. The cause is that nobody settled who is right.

3. The process owner did not know they were the owner

I ask in a workshop who owns the complaints process. Silence, then someone says „customer service, probably", someone else „logistics does part of it". The person who actually holds that process is in the room and finds out at that moment.

Without an owner there is nobody to make the three decisions an implementation requires: what the process should do after the change, what stops being necessary, and when we agree it works. Those decisions then fall to the vendor or to IT — people who have no mandate to make them.

The result is predictable. The tool gets added to the process instead of changing it. People do what they did before, plus operating a new system. The implementation increases the workload.

An owner is not a box on the org chart. It is the person who can say „from Monday we do this differently" and have it happen.

4. The team got a tool instead of an answer about what to stop doing

The implementation ends with training on how to use the system. The team knows where to click. It does not know what it should stop doing.

That question sounds trivial and is the hardest one in the whole project, because the answer usually means someone's work changes or disappears. Nobody wants to say it out loud, so nobody does. The old tasks stay „just in case", the new ones are added on top.

A quarter later the team is doing both, is more loaded than before the implementation, and holds a well-founded view that AI added to their workload. That view stays in the company for years and poisons every project that follows.

Adoption is not a training problem. It is a question of whether anyone dared to say what goes away.

What the partnership does not fix

Back to the start.

OpenAI Select Partner status gives me earlier and deeper insight into what the technology can do, and access to the people building it. That is real value and it is why I applied for it.

It does not give you any of the four things above. It will not measure what your process costs. It will not settle which system is right. It will not name an owner. It will not tell your team what to stop doing.

Those are decisions inside your company and nobody outside it can make them for you. A consultant, myself included, can at most make sure they get made deliberately and in the right order.

That is why I consider selling partner status as an implementation argument dishonest. It answers a question nobody asked.

Where to start

Before you pick a technology, answer four questions about the specific process you want to change:

1. What does this process cost today, in hours and in money. If you do not know, measure it for two weeks before you buy anything.
2. Where is the data and which system wins when they disagree. What is needed is a decision with a name attached, not a diagram.
3. Who owns this process and does that person know it. If there is no such person, the implementation has nothing to rest on.
4. What will the team stop doing once the tool works. If the answer is „nothing", this is not an implementation, it is an addition.

Those four answers usually take two or three weeks. An implementation without them takes several months and ends where it started.

If you want to go through that groundwork with someone from outside, we do it as process mapping and an AI audit. If you would rather do it yourselves, those four questions are the whole method. There is nothing more to it.