AI agents: automating a process without breaking what already works
4 min read
Key takeaways
- Don't start with the technology: start with a measurable, repetitive, time-consuming process.
- A good first use case is measured in hours saved per week, not in an impressive demo.
- Aim for 90 days between idea and real production, not a six-month POC.
- A well-scoped internal assistant costs far less than you might imagine.
- Keep a human in the loop on any decision that commits the company.
For two years, we've been getting the same request phrased ten different ways: "we'd like to do AI, but we don't know where to start". And we see the same failures repeated: a POC that's impressive in a demo, never put into production, forgotten six months later.
The problem is almost never technical. It lies in the choice of starting point.
The three-filter rule
A good first use case passes three filters. If it fails even one, change the subject.
Filter 1: the process is repetitive and time-consuming
Generative AI excels at what is done often, in a more or less standardised way, by people whose time is expensive. Writing a standard reply, summarising a long document, extracting information from a PDF, rewording a technical note for a customer, qualifying an inbound request.
If the process is done three times a year, the saving will be nil whatever the unit gain.
Filter 2: an error isn't catastrophic
On a first project, choose a use where an error is visible, recoverable and without serious consequences. A draft reviewed by a human: yes. A bank transfer triggered automatically: no, not to begin with.
Filter 3: you can measure the result
"It saves us time" isn't a measurement. "The average time to process a case went from 22 to 9 minutes over 340 cases" is. Choose the indicator before developing, and measure it before deployment so you have a point of comparison.
90 daysThe target lead time between idea and real production
Controlled budgetFor a well-scoped first internal assistant
Just 1Number of use cases to tackle to begin with
The four use cases that work almost every time
On the projects we deliver, four families come up constantly because they pass the three filters.
- The internal document assistant. Your procedures, contracts, product sheets and meeting minutes become searchable in natural language. It's the most profitable entry point; we detail how it's built in our article on RAG assistants.
- Document pre-filling. Tender responses, visit reports, summary sheets: the AI produces an 80% draft, the human finishes and approves.
- Qualifying inbound requests. Classifying, routing, summarising and prioritising emails or tickets. Immediate gain, low risk, easy to measure.
- Structured extraction. Turning unstructured documents — invoices, CVs, purchase orders, reports — into data usable by your existing tools.
The public chatbot trap. Many companies want to start with an assistant on their showcase website. It's the most visible use case and the least profitable: low traffic, maximum quality requirement, high reputational risk. Start internally, learn, then expose.
The realistic 90-day timeline
| Period | What happens | Deliverable |
|---|---|---|
| Weeks 1–2 | Choice of use case, measurement of the starting situation, data scoping | Baseline indicator, written scope |
| Weeks 3–6 | Development of the first version, testing on real data | Version usable by 3 to 5 people |
| Weeks 7–10 | Supervised use, fixes, tuning of the instructions | Stabilised version, user guide |
| Weeks 11–13 | Rollout to the full team, measurement, decision on next steps | Quantified result, 6-month plan |
What distinguishes this timeline from a conventional POC: there are real users from week 6. A POC with no users proves nothing, other than that the technology works — which nobody doubts any more.
What to lock down before you start
- What happens to your data. Hosting in Europe, no retention, no reuse for training: to be obtained in writing from the provider before the first API call.
- The usage budget. API costs are low but not zero, and they grow with adoption. We explain how to control them in this article on LLM API costs.
- The supervision rule. Write down in black and white what the AI can decide alone and what requires human validation. That sentence is worth more than ten pages of ethics charter.
- The internal owner. An AI project without a business lead available two hours a week fails, whichever provider you use.
Generative AI doesn't solve organisational problems: it makes them visible faster. A poorly defined process that is automated remains a poorly defined process, just executed more quickly.
Where to start, concretely
Take a sheet of paper. List the five tasks that consume the most hours on your team and that resemble one another from one time to the next. Run them through the three filters. One or two will remain.
That's your first project. Not the one that makes the best demo — the one that saves hours from the first month.
If you'd like us to go through your list with you, the scoping conversation is free and lasts thirty minutes. You leave with a prioritised use case and an order of magnitude for the budget, even if we don't work together afterwards.
Sources
- IA : comment être en conformité avec le RGPD ? (in French) — CNIL (France’s data protection authority)
- Data controls in the OpenAI platform — OpenAI
