Getting started with AI in your business: picking the first task

A week of notes

Before talking about tools, spend a week writing down the repetitive tasks that go through you or your team. You do not need a system: a spreadsheet with five columns is enough.

Frequency times minutes gives the time each task takes up. In the example, 40 × 3 is 120 minutes a week: two hours someone spends copying today. By the end of the week you will have a list you can sort, and there is almost always a task nobody had counted.

How to score each task

With the list ready, score each task from 1 to 3 on five criteria. The one with the highest total is your candidate.

• It repeats: 3 if it happens every day, 1 if it is occasional.
• It takes time: 3 if it adds up to hours a week, 1 if it is a few minutes here and there.
• There are examples: 3 if you have dozens of real cases to show, 1 if every case is different.
• It is quick to check: 3 if you can see in seconds whether it came out right, 1 if it takes an expert to tell.
• A mistake is easy to fix: 3 if it gets fixed before it reaches the customer, 1 if it is expensive (money, a contract, someone’s health).

The last two criteria weigh more than the first ones. A task that saves many hours but whose mistakes are expensive is not a good first task: it is a good third or fourth, once you know how the AI behaves with your data.

Good and bad first tasks

Usually good

• Sorting incoming messages (question, order, complaint, invoice) and sending each one to the right person.
• Moving the data from a recurring type of document into a spreadsheet: supplier invoices, orders, forms.
• Drafting replies to the most common questions, which a person reviews and sends.
• Searching your documents (manuals, policies, price lists) and answering with a reference to where it came from.

Better later

• Answering complaining customers on its own, with no review.
• Deciding prices, discounts, credit or refunds.
• Any task with no examples, or where not even your team agrees on what “done well” means.
• Tasks done once a month: even if they take time, it takes a long while to know whether the AI does them well.

And some do not need AI at all: if the information already arrives structured (a form, an approved payment), a regular automation is enough, cheaper and more predictable. We explain it in AI or automation?

What to prepare before starting

1. Between 20 and 50 real cases, with what came in and what should have come out. They are the yardstick for the trial.
2. The sources the AI may use: the documents, the spreadsheet or the system, in their current version.
3. Who reviews: a specific person who checks what the AI does during the trial.
4. What it may not do: send, pay, delete or publish without approval. Write it down before starting, not after the first scare.
5. The minimum access: if it only has to read a folder, it should not be able to write to it.

If the task touches customer data, also read AI and your business data: what to ask the model provider and which permissions to give.

How to measure the trial

A two- to four-week trial, with human review of everything, is enough to know whether it is worth it. Measure three things:

• Time: how long the task took before and how long it takes now, review included.
• Hits: out of every ten results, how many came out right untouched, how many needed a tweak and how many had to be redone.
• Escalations: how many times the AI said “I don’t know” or asked for help. That is good: guessing is worse.

Write it down by hand at first. With your own numbers, the decision to expand (or not) stops being a matter of faith.

When to stop

Stopping is a result too. It makes sense to stop or switch tasks if:

• Reviewing what the AI does takes almost as long as doing it by hand.
• Mistakes do not go down after adjusting the instructions and the examples.
• To make it work, it needs access to much more than it should.

If the trial goes well, the next step is to loosen review only for the cases that proved to come out right, and to add a second task. To understand which kind of tool fits each case, read AI agents vs. chatbots. And if you want to think it through with us, tell us which task you picked.