After the demo
What happens after the demo
A demo works because a person in the room supplies what the model is missing: which project this belongs to, who owns it, what state it is in. Rollout is the moment that supply runs out. What has to be built between the two is not a better model. It is an explicit structure of how the organization works.

The demo went well. It usually does. That is worth taking seriously rather than explaining away, because the reasons a demo succeeds are real reasons and they tell you what the rollout is missing.
Why does the demo work?
It was not a trick. The data was real, pulled from real systems. The answer was right. The room was correct to be impressed.
Now name what else was in the room. Somebody chose the example, and they chose one they understood. They knew which of four systems held the true version of that record. They knew that this particular customer is invoiced differently for historical reasons nobody has written down. If the answer had been wrong, they would have seen it in half a second and quietly moved on to the next question.
That person did real work during the demo, and none of it was written down. It was not in the model, and it was not in the data.
What is the demo borrowing?
Six things, in roughly the order they break at scale.
- Object identity. Which real thing this is, across systems that name it differently. In the demo one person resolved it. In production nobody does, and the same job exists three times under three identifiers.
- Ownership. Who is accountable for this object right now, as opposed to who was listed as the owner when it was created.
- State. Where the object sits in its lifecycle, and what that permits. Closed, on hold, awaiting approval and in progress are not decorations on a record. They are the rules.
- Exceptions. The cases everybody in the room knows about and nobody has written down. Every operational business runs on a body of exceptions that lives entirely in people.
- Scope. How much was in the demo. One department, one month, one project. The behaviour of a system on a slice is not evidence about its behaviour on the whole.
- Judgement. Somebody was checking. In production the checker is the person receiving the answer, and they do not have the context to check it.
Do we need to clean our data first?
The consensus answer is yes, and it has real evidence behind it. Gartner predicted in February 2025 that through 2026 organizations would abandon 60 percent of AI projects unsupported by AI ready data, and reported that 63 percent of organizations either lack appropriate data management practices for AI or are unsure whether they have them. That survey ran with 248 respondents, and both figures are worth knowing.
It is also worth noticing who gives this advice most loudly. Data cleaning is sold by people who sell data cleaning, and the recommendation is commercially convenient for almost everyone making it, ourselves included when we are careless.
Cleaning your data has no completion criterion. Nobody can tell you when it is finished, because nothing in the exercise defines what finished means.
A company with forty years of accumulated systems that starts by cleaning them will still be cleaning them in three years, and the definition of clean will have moved twice. The work is real, but as a precondition it is a counsel of despair.
A structured model has the completion criterion that cleaning lacks. When the central operating object is named and the organization has agreed what has to be true about it, what is missing becomes a specific, finite and boring list: this system does not hold an owner, that one has no lifecycle state, these two disagree about identity. That list can be worked through and finished, and most of it turns out not to matter for the decisions anyone actually wants to make.
You do not need clean data. You need an agreed structure sitting above the data you already have, and the structure tells you which parts of the mess are load bearing.
What changes between a pilot and a system of record?
| Pilot | Production | |
|---|---|---|
| Who supplies the context | A person in the room | The model, or nobody |
| What a wrong answer costs | A moment of embarrassment | A wrong action in a real system |
| How success is defined | Somebody remarks that it helped | A definition written before it started |
| Who owns it | Whoever ran the demo | A named role in the business |
| How many exceptions appear | The ones that were chosen | All of them |
| What happens when it is wrong | Someone notices | Someone has to be able to reverse it |
How do you know it is working?
If success is defined as somebody remarking that it helped, you will get remarks. People are generous about new tools, particularly ones their executives are enthusiastic about. That is precisely the failure mode that produced the most quoted and least reliable statistic in this field.
Define success before, in the units the business already runs on, over a window somebody agreed to in advance. Days to close a period. Percentage of variations captured in the month they occurred. Hours between a site event and its appearance in the record.
That gap is the whole reason measurement matters. Regular use in one function is easy to reach and easy to mistake for progress. Scaling is the thing that does not happen by itself, and the two are far enough apart that a company can spend two years in the first and believe it is on its way to the second.
Who owns it once it is live?
The structure has to be maintained by whoever is closest to the work changing. When a new contract type appears, when a department reorganises, when a subsidiary is acquired with its own way of numbering jobs, somebody has to decide what that means for the model. If that person is a vendor, the model goes stale the day the engagement ends.
Which means somebody has to sit inside the organization until that ownership is real, rather than delivering a system and leaving. That is a service commitment before it is a technical one, and it is how we work.
What to build before the rollout, not after
- The central operating object, and one agreed identity for it across every system that holds a version.
- What the object has to connect to for the decisions people actually make, and what turns out not to matter.
- The lifecycle states, and which transitions between them are legal.
- The exceptions, written down, including the ones that only one person knows.
- The definition of success and the window it is measured over, agreed before anything starts.
- The owner, by name and by role, who decides what a change in the business means for the model.
This list does not depend on which vendor you choose, or on whether you build it yourself. It is the same list either way, which is a good sign that it is the real work. The object at the centre changes by sector and what the organization needs to know about it does not, which is the argument the model rests on.
The difference between a system that was delivered and a system that is used is whether anyone still has to supply what it does not know.
Q&A
Why does an AI demo work but the rollout fail?
Do we need to clean our data before deploying AI?
How should we measure whether an AI deployment is working?
Who should own an operational AI system after go live?
What is the difference between a pilot and a system of record?
Sources
Where does this break in your organization?
Tell us about one process you actually run. We answer with what we would look at first, not with a deck.
Related reading
Why enterprise AI pilots stall
The most quoted statistic in enterprise AI does not survive its own source. Here is what the evidence actually supports, and what it says about the gap between a pilot and production.
What an AI agent needs before it can act
Credentials decide what an agent may do. They do not decide whether the action is the right one. Agent safety is a semantics problem before it is a permissions problem.
What a vendor means when they say done
Delivery and adoption are measured by different people against different definitions, and only one of them appears in the contract.