Mid-sized companies: moving from AI trials to production

In many mid-sized companies, AI came in through trials: a proof of concept here, an assistant there. The demonstrations convince, then nothing goes into production. The phenomenon is well known, and it is not technical. Here is why projects stall, and how to see them through.

Why trials stall

1. The use case was chosen because it could be demonstrated, not because it was profitable.
2. No indicator was agreed before the trial: nobody can say whether it is a success.
3. The trial data was selected: in real conditions, the exceptions arrive.
4. Nobody is responsible for what comes next: who maintains it, who approves, who pays for running it?
5. The rules are not written down: access, human approval, place of data processing — each department raises its questions, and the project waits.

The five conditions for going into production

1. Choose the process for its measured value

Time spent, processing time, volume: quantified before choosing. A Business Process Audit ranks processes by value and feasibility, on your real data.

2. Agree on the indicators before the trial

Two indicators at most, with their measured starting point. That is what turns a demonstration into a decision.

3. Check on real cases, exceptions included

Not on a flattering sample: on 15 to 30 real cases for a process, with the rules for handing over to a person when the agent does not know.

4. Write the governance once, for all agents

Who decides what, which actions are approved by a person, which access rights, which records, where data is processed. Written once, these rules avoid reopening the debate with every project. See Governance.

5. Organise what happens after delivery

An agent connected to your tools lives with them. You need a lead on the business side, regular monitoring, and a planned running budget. See Why an AI agent needs ongoing monitoring.

When several agents arrive

The second and third agents raise a new question: consistency. The same rules, the same vocabulary, a shared memory — otherwise each agent answers in its own way. That is the subject of Making several AI agents work together and of the Company Brain, which remains optional.

A realistic order of march

1. One process chosen for its value, measured, in service within a few weeks.
2. A second process in the same domain.
3. The whole domain, with its knowledge written down for the agents: Business Agents.
4. Shared rules and memory, if several domains are involved.

Each step is decided on the result of the previous one.

Frequently asked questions

Why do so many proofs of concept never reach production?

Because they prove that something is possible, not that it is useful. Without an agreed indicator, without real cases and without someone appointed to take it forward, the project remains a demonstration.

Do we need a shared AI platform before deploying agents?

Not necessarily. Shared rules — access, approval, records, place of data processing — applied to every agent are better, together with a shared memory when several agents have to work together.

How long for a first agent in production?

For a well-chosen process, about three weeks from scoping to going live. Long delays rarely come from technology: they come from access rights, trade-offs and approvals.