GET STARTED
NUMEZISADVISORY &
ENGINEERING
Contents
Direct answer
Keep a use case only if it passes three tests. Gain: it can be measured in hours per week on real work. Effort: it can start with the assistant on its own or with batch processing, without an IT project. Risk: a mistake stays in-house or is reviewed before it goes out. Three use cases that pass these tests are worth more than ten attractive ideas.
An SME that takes an interest in AI rarely has no ideas: it has twenty. The problem is not finding them, it is keeping three that will hold up. The grid below is the one we use during the AI Workshop and for every decision that follows. It has three criteria, and it is deliberately strict.
Gain: how many hours a week?
A use case only counts if the gain can be measured: hours of work, mistakes avoided, a shorter turnaround. The question fits on one line: how many hours a week does this task take today, and for how many people?
Answer it with the people who do the task, using last week's figures. If nobody can answer in hours, the gain is just a feeling. This first filter rules out most ideas.
Effort: what needs to be put in place?
We rank each use case by what it takes to get started:
- Level 1, the assistant on its own. The team uses a business assistant (ChatGPT, Claude, Copilot, Gemini) on documents it already has: letters, summaries, quotes. Possible from the following week.
- Level 2, batch processing. AI works through a queue of documents or messages (sorting requests, extracting data from invoices, drafting standard replies), without a deep connection to your software. Possible within a few weeks.
- Level 3, a connection to your software. AI reads from and writes to your email, accounting software or CRM, often in the form of an agent. Several weeks to a few months, with our AI Engineering offer.
The first use cases should be at level 1 or 2. Level 3 comes once a first result has been measured.
Risk: where does a mistake end up?
AI will get things wrong. The useful question is: where does the mistake land before anyone notices it?
- It stays in-house: meeting notes, a first draft of a quote. Low risk, you can start.
- It goes to a client or a third party: a reply, a reminder. Medium risk: someone reviews every output before it is sent.
- It involves protected data or a decision about a person: a medical file, candidate data, legal advice. High risk: not as a first use case.
These three levels match the three data categories in our security guide: internal, client and protected.
The grid filled in: an example
| Use case | Gain per week | Effort | Risk | Decision |
|---|---|---|---|---|
| Reminders for unpaid invoices | 3 h | Level 2 | Client, reviewed before sending | First pilot |
| Quotes entered twice (CRM and accounting) | 2 h | Level 2 | Internal | Second |
| Meeting notes | 1 h | Level 1 | Internal | Third |
| Agent that answers complaints on its own | Unknown | Level 3 | Client, no review | Break down and wait |
| "Transform the company with AI" | Not measurable | Undefined | Undefined | Rejected |
The first three rows make up the plan. The fourth is a project, not a use case: its first step (sorting complaints and preparing a reviewed draft) can come back to the grid later. The last one is an intention.
Why three, not ten
Every use case takes time: choosing it, measuring it, refining the instructions, following up on how it is used. Ten use cases in parallel take ten times that, and the team's attention gets spread thin. Three use cases done properly build a habit; ten done by halves build discouragement.
What happens after the first three
The first use case becomes a 30-day pilot. Once all three are running, we look at what works (and extend it) and what does not (and stop it rather than keep it on life support). The next use cases come from what you observe in practice, not from the original list of ideas. If the question is whether a task calls for an assistant or for automation, our analysis of team assistant or automation sets out the rule.
Questions
Where do we find use cases if we cannot see any?
In the tasks that come back every week: data entry, reminders, letters, meeting notes, answers to the same questions. The people who do these tasks know how long they take.
How many use cases should we aim for at the start?
Three, all at level 1 or 2. A level 3 use case, connected to your software, comes after a first measured result, not before.
What about spectacular ideas, such as an agent that runs the whole customer service desk?
Keep them on the list and break them down. Their first step, for example sorting incoming requests with a reviewed draft, often passes the grid.
Does a marketing use case count?
Yes, if the gain can be measured: hours of writing, formatting or translation. Otherwise it is a gadget, and that soon shows.
Two use cases look alike, how should we count them?
As one use case with two variants. Each one takes time to scope and follow up, so a longer list does not mean a bigger gain.
Discuss this decision
hello@numezis.com ↗