The short answer
An AI project is finished when the system runs on real work, the team uses it without being reminded and someone on the client's side can explain it, not when the demo goes well. That takes a handover planned from the first day, real users in every sprint, a supervised first month and measures of use as well as accuracy.
A lot of AI work stops at the point where it looks finished. The system is built, the demo went well, the invoice is paid. Then it meets real work, a few things go wrong, nobody is sure who to ask, and slowly the team goes back to doing it by hand.
We think of finished differently. A system is finished when it runs on real work, the team uses it without being reminded, and someone on your side can explain how it works and what to do when it does not.
01When should handover planning start?
Handover is not a meeting at the end. It is a decision made at the start about who will own the system, what they need to know and what they should be able to change on their own. That shapes how we build: clear names, visible logs, settings in plain language and no part that only we can touch.
02When should real users get involved?
We work in short cycles with weekly demos, and the people who will use the system are in those demos from the start. Not to approve a finished product, but to try something half built and tell us what is wrong with it. The earlier they shape it, the more it feels like theirs when it goes live.
If the people it was built for do not want to open it, it is not finished.
03What does a proper handover include?
- Step 01A playbook in your wordsWhat the system does, what it does not do, how to spot a problem and what to do next. Short enough that people actually read it.
- Step 02Training on real workSessions with the team using their own cases, so the first time they see an unusual result is not the first time it matters.
- Step 03Monitoring you can readA view of how the system is doing that makes sense without us in the room: what it handled, what it passed to a person, what went wrong.
- Step 04A clear next stepWe stay on call, or step away cleanly. Either way, you know which it is and what happens if something breaks.
04What happens in the first month after launch?
The weeks right after a system goes live are when most of the useful learning happens. Real volume finds cases nobody thought of, and the team works out how the system fits into their day. We plan for that month rather than treating launch as the finish line.
- A short check-in each week to go through what the system handled and what it passed to a person.
- New edge cases added to the test set as soon as they appear.
- Small changes made quickly, while the team still remembers what bothered them.
- A clear point at which the system moves from supervised to trusted, agreed in advance.
By the end of that month, the system should feel ordinary. That is the goal. The best AI systems we build are the ones nobody talks about any more, because they are simply part of how the work gets done.
05How do you measure whether a system is used?
A system can be accurate and still fail if nobody uses it. So alongside the quality checks, we watch the simple signals: is the team opening it, approving its drafts, trusting its results? When those numbers are low, the fix is rarely the model. It is usually the place the system lives, the way it presents its work or a step that asks too much of people.
That is a higher bar than a working demo, and it takes longer to reach. But it is the only version of finished that gives you back the hours you were promised.
FAQCommon questions
- When is an AI project actually finished?
- An AI project is finished when the system runs on real work, the team uses it without reminders and someone on the client side can explain how it works and what to do when it does not. A successful demo is a milestone, not the end.
- What should an AI handover include?
- A proper AI handover includes a short playbook in plain language, training on the team's own cases, monitoring the team can read without the builder present, and a clear agreement on whether the builder stays on call or steps away.
- How do you know if an AI system is being used?
- Track simple signals alongside accuracy: whether the team opens the system, approves its drafts and trusts its results. Low use rarely points to the model. It usually means the system lives in the wrong place, presents its work badly or asks too much of people.
Written by
Wassim, design and development
Wassim shapes how everything we ship looks and feels, then writes the code behind it: websites, client panels, internal tools and the screens your team uses to work with an agent.
