Which task should you start with AI at your company? A priority matrix
Choose a first, useful, and verifiable use without chasing every possibility.
SqualiOnline editorial team · 2026-09-07
The problem isn't finding a task where artificial intelligence could help: you can find twenty in half an hour. The problem is that a poorly chosen first attempt burns the trust of the people who will have to use every one that follows, and it burns fast. This guide helps you choose a single one to start with, and set in advance how you'll know whether it worked.
You start from tasks, not from tools
You build the list by talking with the people who do the work, not in a meeting with the people who describe it. For each task you need five answers.
- How often it happens: daily, weekly, monthly.
- How much time it takes and whose. The time of a person who could be doing something else counts differently than dead time.
- What information is needed to do it, and where that information lives today.
- What happens when it's done badly, and who notices.
- How similar it is to itself each time.
That's five answers per task, and you can gather them in an afternoon. Whoever skips this step ends up choosing the first case based on how interesting it is, which is the worst criterion available.
Three questions that rule out half the list
- How variable is it? A task with a new context and different rules every time is hard to automate and even harder to check: you wouldn't be able to tell whether the result is right without redoing the work.
- What happens if the outcome is wrong? A wrong draft is fixed in a minute. A wrong message to a customer, a wrong price, wrong technical advice produce consequences that can't be undone.
- Does the data exist, and is it readable? If the information lives in scanned photocopies, in people's heads, or in five systems that don't talk to each other, that's where the real work is, and it isn't a job for artificial intelligence.
The matrix
Once you line the tasks up side by side, the choice usually makes itself. The three rows below show how to fill in the table: they don't describe a real company.
| Task | Frequency | Variability | Error | Data ready | Verdict |
|---|---|---|---|---|---|
| Sorting incoming requests by type | High, every day | Low: the requests are always the same four or five | Recoverable: it ends up in the wrong queue and gets retrieved | Yes, it's text and already exists | Good first case |
| Searching for an answer inside internal documents | High, but different for each person | Medium | High if someone acts on a wrong answer | Only partly: many documents are scans and some are outdated | Good case, but only after the documents are sorted out |
| Preparing the monthly summary for management | Low: once a month | High: it changes with what you want to look at | Medium: a wrong number leads to a wrong decision | Yes, it's in the business management system | Not well suited as a first case: it saves little and everything must be rechecked |
Before you say artificial intelligence
Several tasks on the list can be solved with something simpler and more stable: a rule, a required field in a form, a label, a separate mailbox, a query on the business management system. It's worth checking, because the simple solution can be explained to anyone and doesn't need supervision.
The cost you don't see
The cost is rarely in the middle. It sits at the two ends, and it's the part that optimistic quotes forget.
- Where the input information comes from, and who keeps it updated over time.
- Where the result ends up. If someone has to copy it by hand from one window to another, you've moved the work, not removed it.
- Who checks it, how often, and on what sample. Supervision is a recurring cost, not a one-time phase.
- What happens when the system doesn't respond or responds badly: who notices, and how work continues in the meantime.
- Who answers the customer when the outcome affects them.
A case that flows in and out of the tools people already use costs much less than one that requires an extra window to keep open and remember to check.
Choose, and write the criteria first
- Choose a single task: the most frequent one among those with low variability and a recoverable error.
- Write down what a person does today, step by step, including the exceptions.
- Measure the starting point — time spent, volumes, errors observed over a defined period — before touching anything. Without this measurement, no comparison will be possible.
- Define what will count as success, in an observable way, and who will declare it.
- Set the date on which the decision will be made, and the three allowed answers: extend it, fix it, or drop it.
The third point is the one that gets skipped most often, and it's the one that separates a project from a demonstration. Without the starting figure, in the end you'll be arguing over impressions.
What this guide doesn't cover
Here you choose the task to start with. How you then organize the pilot project — on what sample, with what supervision, and with what final decision — is a separate job. The direct comparison between traditional automation and artificial intelligence, when both are viable, also deserves its own discussion.
Frequently asked questions
How many tasks should you start with?
One. Two projects together double the meetings and split your attention, and when something doesn't work you can't tell whether the problem is the chosen case or the way you're working. You choose the second task using what you learned from the first.
What if the first case doesn't bring measurable savings?
That's an outcome, not a failure, as long as the starting measurement existed. It tells you that class of task isn't worth it, and it usually leaves something reusable behind: sorted documents, clean data, a process finally written down.
Does the data need to be perfect before you start?
No, but it needs to be readable and have an owner. If no one knows which version of a document is the valid one, the problem is identical with or without artificial intelligence: the difference is that with AI it spreads faster and with more apparent confidence.
Let's find a first use case suited to your company.
If you’d like to talk it through, the service that handles this is Artificial intelligence.