Which processes should you hand to an AI agent first? Five criteria
Volume, accessible data, tolerance of error, measurable gain, reversibility: the five criteria for choosing the first work to hand to an AI agent.
Choosing the first process is the decision that matters most: a first success opens the way, a first failure closes it for a long time. Here are the five criteria we use, and the processes that most often meet them.
The five criteria
| Criterion | The question to ask |
|---|---|
| Volume and repetition | Does it come round at least several times a week, in a comparable form? |
| Accessible data | Is the information in a tool that can provide it, or in someone’s head? |
| Tolerance of error | If the agent gets it wrong once in twenty times, what happens? |
| Measurable gain | Can we say, before starting, what we will measure and where the figure stands today? |
| Reversibility | Can we stop the agent tomorrow morning and go back to how things were, without damage? |
Why tolerance of error matters so much
An agent can get things wrong, just like a person. The point is not to avoid this at all costs, but to know what a mistake costs. A reminder sent a day too early can be put right; a payment sent to the wrong account, much less so. That is why irreversible actions remain approved by a person. See Can an AI agent make mistakes?
Why measurement must come before the project
Without a starting figure, no gain can be demonstrated. Before choosing, record at least the time spent each week, or the processing time. See How to calculate the return on investment of an AI agent.
The processes that often tick all five boxes
In our AI agent catalogue, the first projects we recommend to a company with no experience of AI are these:
- Payment reminders and collections — the only one that brings in cash, not just time.
- Qualifying incoming requests — answering fast changes the conversion rate.
- Meeting notes — no access to a business system needed, immediate adoption.
- Following up quotes — a forgotten quote is an avoidable loss.
- Processing supplier invoices — the best-documented case, and a flow that is becoming structured with e-invoicing.
- Handling routine customer service requests — high volume, escalation simple to define.
- Processing the mailbox — the entry point for almost all work.
- Bank reconciliation — clean data, explicit rules.
This ranking is a starting point, not a truth: in your company, the right first process is the one your figures point to.
Processes to avoid at first
- Those that depend on a single person with nothing written down: their knowledge has to be gathered first.
- Those that create legal or financial commitments with no possibility of review.
- Those that change every month: the agent will spend its time being readjusted.
- Recruitment and employee evaluation: these uses are tightly regulated by the European AI regulation. See AI Act: what changes for SMEs and mid-sized companies.
From task to process
If your first choice is a single task — extracting data, sending reminders, sorting —, a Quick Start is enough. If it is a complete process, with its decisions and special cases, the Process Agent is made for it. And if you are hesitating between several, the Business Process Audit compares them, with figures to back it up.
Frequently asked questions
Should we start with the most expensive process?
Not necessarily. The most expensive is sometimes the riskiest or the least documented. A clear first success on a mid-sized process is better than a failure on the biggest one.
What if no process ticks all five criteria?
A negative answer is often a prerequisite rather than a refusal: gather the data, write down the rules, measure the starting point. That is sometimes the real first project.
How many processes should be handled at once?
One or two at the start. Beyond that, if several processes belong to the same domain, it becomes more efficient to handle the whole domain.