Almost every company that decides to "use AI" starts with the same wrong question: which tool should we hire? The right question is different — which process goes first. The wrong tool you can swap in a week. The wrong process eats up months and burns the team's patience with the technology.
This piece is the criterion for choosing that first process: the ones that usually deliver quick returns, the ones that seem obvious and disappoint, and the three-question test that separates one from the other.
Why most companies start in the wrong place
A company's first AI project is almost never chosen by criteria. It's chosen by whoever spoke loudest in the meeting, or by the flashiest demo someone saw in a video. The result is predictable: the most visible process gets picked instead of the most repetitive one.
These are different things. Visible is the client presentation, the board report, the marketing material. Repetitive is the same task, with the same structure, dozens of times a week, done by someone who's already tired of it. AI pays off on the repetitive, because that's where volume covers the effort of setting it up — and where an occasional error costs little.
There's a second, more uncomfortable reason: the repetitive process is usually invisible to whoever is deciding. Whoever sits on the board doesn't see the three daily hours the team spends copying data from one system to another. That's why the best first project rarely shows up in the board meeting — it shows up when someone asks the team what eats up most of their day.
The processes that usually pay off first
1. First-level triage and response. Questions that repeat themselves — deadline, price, order status, required document, business hours. It's not about replacing customer service: it's about AI answering what's repeated and passing on to a person what's specific. Quick return because volume is high and the error is cheap.
2. Document reading that turns into a field. Invoice, contract, report, form, résumé, receipt. Someone opens the file, reads it and types the information somewhere else. That's the most expensive and most invisible manual work that exists in Brazilian companies — and it's what modern AI does best.
3. Classification and routing. An email, a ticket, a form comes in: which department does it go to, at what priority. One person does this all day, and it's the kind of decision AI gets right consistently, because the pattern repeats.
4. First draft of standardized text. Proposal, product description, quote response, report with a fixed structure. It's not AI writing in your place — it's AI delivering the 80% draft for someone to review in five minutes instead of writing it in forty.
5. Querying internal knowledge. The company has a manual, a policy, project history, an old spreadsheet — and nobody can find anything. Asking in plain language and getting the answer with its source is one of the uses with the highest team satisfaction, because it eliminates a daily frustration.
6. Data checking and cross-referencing. Matching what's on the order with what's on the invoice, finding discrepancies, flagging what fell outside the pattern. Work humans do poorly precisely because it's repetitive — attention drops after the twentieth line.
7. Meeting and conversation summaries. The cheapest to roll out, the lowest resistance, and a good first project when the company is still skeptical — because the benefit shows up in the first week.
The ones that seem obvious and disappoint
It's more useful to say where not to start than to stack up promises:
- Decisions with serious consequences and no review. Credit, diagnosis, legal, termination. AI helps prepare the decision; having it own the decision alone is a different level of project, risk and responsibility.
- Anything that depends on data the company hasn't organized. If the information is in three spreadsheets with different names and a system nobody updates, the project turns into data cleanup — and that's what the budget will pay for, not AI.
- A process that changes every week. You can't automate what hasn't stabilized yet. First the process becomes a process, then automation comes in.
- Replacing what already works well. If the team delivers well and on time, touching that creates a lot of friction for a small gain.
- The flashiest case to showcase. A pilot chosen to impress tends to be hard to measure. And a project that can't be measured doesn't survive the second quarter.
The three-question test
Take your candidates and run each one through these three. If the answers are yes, yes and yes, you've found your first project:
1. Does it happen at least every week? Frequency is what pays for the setup. A quarterly task, however annoying, isn't worth automating first.
2. Can someone explain how it's done, start to finish, in five minutes? If nobody can, the process isn't clear — and AI doesn't organize what the company doesn't understand. This is the test that eliminates the most candidates, and it's the most useful one.
3. Can the error be caught before it causes damage? Is there review, checking, someone along the way? Start with processes where the error shows up fast and costs little. Confidence for the harder cases builds afterward.
Tool or project: the difference that decides the budget
Much of what's on this list starts with a tool subscription and a bit of discipline — and it's honest to say so. Meeting summaries, text drafts and asking the manual questions fit into an off-the-shelf tool, and anyone selling a project for that is overselling.
The math changes when three things show up: the data is sensitive or can't leave the company, the process needs to talk to systems that already exist, or the result needs to be auditable — knowing who decided what, when, and based on which information. At that point it's no longer a tool, it's AI implementation for companies: integration, permission rules, a record of what was done, and someone accountable when it fails.
If your question is still about cost before scope, the article how much it costs to have AI in your business is the way to go. If the first process on your list is customer service, it's worth reading about AI chatbots for companies first.
What the first project looks like in practice
A well-chosen first project has four characteristics, and none of them is technical:
- A single process, with a clear start and end. Not "AI in customer service," but "answer the ten most common questions."
- A number measured beforehand. How many times a week, how long each one takes. Without that there's no way to prove a gain afterward — and without proof there's no phase two.
- One owner. Someone from the department, not IT, who answers questions and validates the result.
- An emergency exit. How to go back to the old way if it doesn't work. A project with no plan B scares the team, and a scared team sabotages the pilot without even realizing it.
Frequently asked questions
Where should a small company start using AI?
With the most repetitive process that's already clear in the team's head — usually triaging repeated questions, reading documents that turn into fields, or classifying tickets. Small size isn't an obstacle; a large scope is.
Do I need to organize my data before using AI?
For internal knowledge and data cross-referencing cases, yes: the quality of the answer is the quality of what's stored. For triage, drafting and summarizing, no — these work with what the company already has today.
What's the risk of AI making a mistake with a client?
Real, and that's why the first project should have human review along the way. The costly error isn't AI getting something wrong: it's AI getting something wrong without anyone noticing. Start where the error shows up fast.
Will the team resist?
They resist when the project is presented as a headcount cut. They don't resist when it starts with the task the team itself hates doing — which is exactly where the return is greatest. Choosing the first process is a people decision before it's a technical one.
Can AI be used without data leaving the company?
Yes, and that's the main reason to turn AI use into a project instead of a subscription. The methods vary by case — from a model running on your own infrastructure to agreements that prevent data from being used for training — and the choice depends on the type of information involved.
How long until results show up?
It depends on the process chosen and how organized it already is. What can be guaranteed is the opposite: a project that starts without measuring the "before" never shows a result, because there's nothing to compare it to.
The first step, today
Before looking at any tool, do one thing: ask three people from operations which task they would make disappear if they could. Write down the answers and run each one through the three-question test. What's left is your first project — and that list costs an afternoon, not a budget.
Already know which process hurts and want to know if it's worth a project?
Tell us what your team repeats every week and we'll tell you, with no commitment, whether it's a case for an off-the-shelf tool, a custom AI project, or fixing the process first.