Demo versus system: why most things built with AI never get used

What a demo actually is

A demo is a working answer to the question "could this exist". It is real software, it runs, and it is genuinely useful - as evidence. It is not evidence that the thing is ready to be used by anyone other than the person demonstrating it.

This is not a criticism of demos. Building one is the correct first move, and it is the move this platform makes cheap. Half an hour of describing gets you something on a screen that either makes people in the room lean forward or does not, and finding that out for half an hour is a bargain against finding it out after a quote for fifteen thousand euro.

The problem is only that the demo and the finished thing look the same. Both are a page with buttons on it. Nobody watching can see which one they are looking at, and quite often neither can the person who built it.

The five things a demo does not have

A login with roles, data that survives, sensible behaviour when something goes wrong, a person who owns it, and a way to change it next month. Each one is invisible in a demonstration and each one is the reason a demo stops being used.

Notice that none of these are visible in a demonstration. A demonstration is one person, doing the right thing, in the right order, on data they created themselves five minutes ago. Every one of the five failures needs a second person to appear.

The moment it breaks is the second user

Software that one person uses correctly is a document. Software that two people use at once is a system, and everything that makes systems difficult starts at the second person.

The second person opens the same record while the first is editing it. The second person types a phone number with spaces in it. The second person is a colleague who should not see salaries. The second person does the steps in a different order, because they were trained differently, or because they were never trained at all.

This is why a demo that impressed everybody in the meeting is quietly unused six weeks later. Nothing dramatic happened. It just never survived the first week of real people doing real things at the same time as each other.

Why this is happening more now, not less

The cost of the first eighty per cent of a piece of software collapsed. The cost of the last twenty per cent did not move at all. So the number of half-finished things being built has gone up sharply, and the number reaching daily use has not kept pace.

It used to be that getting to a working screen was itself expensive, which meant almost nothing got started that was not going to be finished. That filter is gone, and it was doing more work than anyone realised.

The figures around this in Bulgaria are worth sitting with. Eurostat has AI use among Bulgarian companies at 8.6%, and basic digital intensity among small firms at 38.3% against an EU average of 71.4%. The gap between those and the number of firms that have now tried something with AI is the pile of demos. Trying is cheap and common. Using is neither.

What crossing over actually takes

Roughly: describe who the users are and what each of them may see, put in real data rather than invented data, run it for a week alongside the old way, and fix what the week exposes. That is days of work, not months, but it is not zero.

• Name the roles before building, not after. Two roles is usually enough, and choosing them is a decision only you can make
• Load real records, including the ugly ones - the client with no VAT number, the order that was cancelled halfway
• Run it in parallel with whatever the process is now, for one full week including a Monday and a month end
• Write down every moment somebody hesitated. Those are the changes, and they are usually five or six sentences
• Decide who owns it. If the answer is "we all do", it is nobody

The honest part: some demos should stay demos

A good share of what gets built should be looked at once and abandoned. That is the demo doing its job, not failing at it.

If a demo shows you that the process underneath is not agreed, the correct response is to go and agree the process, not to finish the software. If it shows you the problem is smaller than you thought, the correct response is a shared document and no software at all. We have talked people out of building things on this platform for exactly that reason, and it is a better outcome than a system nobody opens.

The failure is not building a demo that dies. The failure is not knowing which one you have, and paying to keep something running that no one has opened since March.

How to tell which one you are being shown

Ask three questions: who is the second user, what happens when they enter something wrong, and what does it cost to change this in November. Anything that cannot answer all three is a demo, whoever built it.

These questions work on your own work, on a supplier's pitch and on something a colleague built over a weekend. They do not require you to understand any of the technology, which is the point of them.

On this platform the crossing is deliberately cheap: the same description that produced the demo is what you extend, so adding roles and real data is more describing rather than a rebuild. It is still a week of your attention, and no tool removes that.

QUESTIONS