Where to start with AI in a company: pick a first project, not a tool

The wrong first question

"Which tool should we buy?" arrives too early.

It leads to a demo, a trial, sometimes a subscription. It does not say which task disappears, who reviews the output, or what happens when the system is wrong. Six weeks later the team has three tools and the same copy-paste.

The useful first question is flatter: which task comes back often enough, with a start and an end clear enough, to be worth doing another way?

If you cannot write it in one sentence, you do not have a project yet. You have a wish.

Five families of first project

In real company work, a useful first step almost always falls into one of these five families.

Automate a flow you already repeat

Follow-ups, data entry, handoffs between two tools, Monday exports. The sequence already lives in someone's head. You often only need to write it down: trigger, steps, result, exceptions.

This is often where you recover the most hours, the fastest. Not because AI is magic: because the rule is already there. I laid out the detail in automation in a company: DIY or expert. The workflow automation page describes how Kirako builds that kind of flow.

Find the information without explaining it again

The same questions come back. The answers already live in emails, PDFs, a Drive, or one person's head. A focused assistant, on one limited corpus, with a review step, cuts the search. It does not need access to the whole company.

The usual failure: connect "the entire Drive" and watch the system invent as soon as a document is missing.

Prepare the files, keep the decision

Quotes, inbound requests, files that change a little every time. AI can gather, sort and prepare. A person approves. You gain prep time, not the right to sign.

The moment an action goes out on its own (a client email, an invoice, a status change), it is no longer preparation. It is a production agent, with stop rules. The AI agents page starts from that rule.

Get the team to the same level

Sometimes the bottleneck is not a flow. It is that everyone uses AI differently, or not at all. Two real cases, a review rule, a list of what never goes out: that moves faster than a prompt catalogue.

Kirako's AI training is built that way. Buying a tool before this step mostly produces sprawl.

Write the problem down before you build

If the need is still "we should do AI", the first gain is not an agent. It is following one real file for a week and noting where it sticks. Only then do you decide whether it is worth a test.

That is not a retreat. It is how you avoid plugging a tool into a topic nobody owns.

What must stay human

Whatever the project, three things stay with a person:

• the decision that commits the customer, the price, the contract or a sensitive record;
• the review when the system prepares a text or a file;
• the stop, when the case leaves the rule.

If you do not yet know who reviews, you do not have an AI project. You have a demo.

When it is too early

Do not force a first project if:

• nobody has 30 minutes to own it;
• the business rules change on every file and nobody can write them down;
• the topic is sensitive (hiring, health, credit, customer data) and you have no frame yet.

In those cases the useful deliverable is a scoping note, not a tool. That is what AI consulting is for.

See an agent before you scope

Before you pick a first project, you can try an agent on your site. It reads your public pages and answers like a front desk. It is not an audit, and it is not connected to your tools yet.

Then two useful exits:

• the trial is enough, you see what an agent can already say;
• you want to scope a real first project, you pick a slot.

The useful first project is rarely the most impressive one. It is the one you can describe, measure, and stop.