How to implement AI in a small business: start with one task

Before you start: which problem do you want gone?

“We want to use AI” isn't a goal. “The front desk stops asking every morning which hotel is on which route” is. The difference is that you can check the second one on any given Friday.

What gets called AI in an office today is almost always generative AI: software that writes and answers based on text. It's good at things that are already written down somewhere and that somebody has to find, summarize or rewrite. It's bad at things nobody ever wrote down.

So the first question in how to use AI in a small business isn't which tool to buy. It's this: what gets answered the same way twenty times at your company?

AI adoption in small business: where things stand

If you feel late, the numbers say you aren't. The U.S. Census Bureau asks businesses every two weeks whether they used AI in any business function. Between December 2025 and May 2026, the overall share hovered between 17% and 20%. Fewer than one in five firms with four or fewer employees reported using it. Among firms with under 20 employees, use didn't change significantly over those months (Census Bureau, Business Trends and Outlook Survey, published May 26, 2026. Checked on November 12, 2026).

Most small companies haven't started. The ones that have mostly started the way this guide suggests: with something small.

Week 1: pick one task and its owner

For three days, ask two or three people to write down the questions they get from coworkers and customers. No sorting, one sheet of paper. At Larkspur Linen, a twenty-person commercial laundry that washes for hotels, three came up: what each hotel's contract covers, what to do when an item arrives damaged, and what time each route leaves.

Pick one. A good candidate meets three conditions:

• It repeats. Several times a week, not once a quarter.
• The answer is written down in some document, even if it's scattered.
• A mistake can be fixed. Answering an internal question, yes. Deciding who gets hired or let go, no.

Then name an owner. They don't need to know IT: they need to know the work and have a fixed slot every week. This is the person who gathers the documents, logs the misses and says “not yet.” Without that person the plan doesn't hold, whatever tool you use.

Week 2: gather what's already written

An assistant answers well about your company once it has read your documents. People call that “training AI on your data,” though almost nothing gets trained: you hand it the documents so it can look things up. We go through it slowly in the guide on how to train AI on your own data.

Don't upload the whole shared drive. Start with what answers five questions: what you do and for whom, who your typical customer is, what you sell, how you talk, and what can't be said. At Larkspur that was the standard contract, the current rate sheet, the damage-claim procedure, the route schedule and one page on how emails should sound.

Two jobs nobody wants and that decide the result:

• Setting aside what's expired. If there are three rate sheets, the assistant doesn't know which one counts. You do.
• Writing down what only lives in one head. Half a page is enough: “if a hotel says an item is missing, check the pickup ticket first.”

End the week by asking it ten questions you get for real. If you want to see the full setup, it's in the guide to building an internal chatbot on your company's documents.

Week 3: test with three people and log the misses

Three people with different jobs: at Larkspur, the shift supervisor, someone in the office and a route driver. They use it in their actual work for five days and note two things: when it couldn't answer and when it answered wrong.

This week's misses are the most valuable material of the month. Almost all of them get fixed the same way: a document was missing, or two documents disagreed. What documents don't fix is the assistant making something up with a straight face. It happens, it has a name, AI hallucinations, and it's the reason for next week's most important rule.

Week 4: open it up and set ground rules

Opening it up means showing it: twenty minutes, with examples from each job, beats an email with a link. Plus one page of rules that fits on the break-room board:

• Which tool we use, and with which account.
• What never goes in: personal data about customers or employees, passwords, bank details.
• Anything going out to a customer gets read by a person first. Always.
• What it's not for: decisions about people.
• Who to ask when in doubt: the owner.

If you want that page written properly, we have a starter in the guide to an AI policy for a small business. The last task of the month is putting a monthly review on the calendar. Documents change and the assistant doesn't find out by itself.

What it costs and how to tell if it worked

There's no figure that fits every company, and be wary of anyone who quotes you one without knowing yours. What you can do is look at the four line items the cost comes from:

To measure, pick one signal in week 1 and look at it at the end:

• Do the repeat questions reach the person who used to answer them less often?
• How many times did it not know, and is that dropping week over week?
• Are the people it was set up for using it, or only the person who set it up?

If nobody opens it after four weeks, the useful conclusion isn't “AI doesn't work.” It's “that wasn't the task.”

Mistakes that keep repeating

• Starting with the tool. It gets bought, handed out, and everyone uses it for their own thing. A month later nobody can say what it was for.
• Wanting five tasks at once. None turns out well and all of them leave doubts.
• Uploading everything. Old, contradictory documents give old, contradictory answers.
• Naming nobody. What belongs to everyone gets fixed by no one.
• Taking the first answer as final. What it writes is a draft.
• Letting everyone use a personal account. Nobody knows what data has left the company. More on that in company data in ChatGPT.

When it isn't worth it

Sometimes the sensible move is to wait:

• If the task happens once a month. Setting it up costs more than it removes.
• If nothing is written down and nobody has time to write it. The document comes first.
• If the answer changes daily and lives in your business software. That means connecting that software, and that's a different project.
• If a mistake has serious consequences for a person: health, money, a job. That's not a first task.

How it works with Flipa

The plan works with any tool. With Flipa, the assistant we make, each week has its place. It works inside your team: it drafts, and a person sends.

• Gather. When you first sign in there's a guided four-step setup: your website, your documents, your connections and a last step where Flipa sorts what it has read and tells you how much it knows. A meter tells you how well it knows your company and which topics are missing. Five are must-haves, the same five from week 2.
• Test. When someone asks something that isn't in the documents, it says so and logs it in a list called “Me lo han preguntado y no lo sabía” (“I was asked and didn't know”). That's the week 3 list of misses, kept for you.
• Open. There are two roles: whoever manages the account uploads documents and invites everyone else, and everyone else asks and hands off work. The plan's owner should be the person who manages your Flipa account.
• Measure. The “Consumo” (usage) screen shows how much it's being used, by whom and for what.

What it doesn't do: it doesn't send emails or publish anything by itself, and what it writes is a draft that someone reviews. What a US team should know up front: the app's interface is in Spanish today, plans start at $99/mo, bought by card, and the Trial is free, no card needed. The details of getting started are in getting started.

Checklist before you call the month done

1. There's one chosen task and you can state it in a sentence.
2. There's an owner, by name, with a fixed weekly slot.
3. The documents loaded are current and don't contradict each other.
4. Three people have used it in their actual work and their misses are logged.
5. The ground rules fit on one page and the whole team has seen them.
6. There's a signal to measure and a date for the first review.