Preparing an AI agent project: access, examples and a decision-maker
Access to tools, real examples, special cases, a contact, a decision-maker: the list of what to gather before starting an AI agent project.
Most delays in an AI agent project do not come from technology. They come from an access right that does not arrive, examples that are too clean, a decision that does not get made. Here is the list of what to gather before starting.
1. Access to the tools concerned
For each tool the agent has to read or change — email, CRM, accounting software, ERP, shared folders:
- who can grant access;
- what level: read only, or write;
- a test environment, if the agent has to write in a production tool.
Granting access sometimes takes days — a publisher, an IT department, a provider to ask. Start it first.
2. Real and representative examples
Not three perfect invoices: a sample that looks like everyday work.
- routine cases, in numbers;
- difficult cases: a badly scanned document, an unknown customer, an amount that does not match;
- rare but costly exceptions.
A process is won or lost on its exceptions, not on its routine case.
3. The rules, even unwritten ones
What the person applies without thinking: deadlines, thresholds, tolerances, the tone used with a given type of customer. Scoping is there to write them down. If they only exist in one person’s head, plan time with them.
4. An available contact
About two hours a week, during the project: the person who knows the work and its exceptions. They answer questions, approve behaviours, test on their cases.
5. A single person who decides
Trade-offs come quickly: does the agent handle this case or pass it on? Does this reminder go out on its own or after approval? A single person settles it, quickly. A decision committee lengthens every deadline.
6. The indicator and its starting point
Before starting: what will be measured — time spent, processing time, number of corrections — and today’s figure. See How to calculate the return on investment of an AI agent.
What is not needed
- Fifty pages of specifications: a description of the real work is enough.
- An overhaul of your tools: the agent works in the ones you have. See Connecting an AI agent to your existing tools.
- Technical skills on your side: they are on ours.
The list, in short
| To gather | Why |
|---|---|
| Access rights, before the start | first cause of delay |
| Real examples, exceptions included | the agent is checked on them |
| The rules applied | otherwise the agent gives answers that are right for everyone, and wrong for you |
| A contact, two hours a week | they know the exceptions |
| A single person who decides | trade-offs must not wait |
| The indicator and its starting figure | without it, no gain can be demonstrated |
These are exactly the conditions written into each of our proposals: Offers and prices.
Frequently asked questions
Do we need detailed specifications?
No. You need a description of the real work, with its exceptions, and the indicator to measure. Scoping is done with the person who does the work; a document of a few pages is enough.
How many examples should we gather?
Enough to cover routine cases and the main exceptions. For a complete process, we check the agent on 15 to 30 real cases before putting it into service.
What happens if an access right is missing?
The project waits. That is why access must be obtained before the start — and it is written into our proposals: without it, the work is paused until it arrives.