How to implement AI in a business: a step-by-step roadmap

If you haven’t done much with AI in your business yet, you’re in the majority. The US Census Bureau asks businesses about it directly, and most say they aren’t using it.

This guide starts from that data and then lays out a roadmap that fits it.

Where businesses actually are

The Census Bureau’s Business Trends and Outlook Survey asks businesses whether they used AI in the past two weeks. Its May 2026 summary of the latest six months of data reports that “overall AI usage hovered between 17% and 20%.”

Size makes a big difference. From the same report: “37% of firms with at least 250 employees reported using AI in their business operations,” and “32% of firms with 100 to 249 employees said they used AI.” At the small end, AI use “didn’t change significantly among firms with fewer than 20 employees” over those six months, and “Less than 20% of firms with four or fewer employees reported using AI.”

Industry matters too. The report puts the Information sector at 39.7% and Finance and Insurance at 33.9%, against a national rate of 19.8%. Retail was lower, at around 14%.

By the government’s own count, most American businesses weren’t using AI as of May 2026. A slow, careful start is normal.

What the businesses using it actually do

A National Bureau of Economic Research paper, The Microstructure of AI Diffusion, dug into the Census data to see how firms that use AI are using it. Two findings shape the roadmap below.

The first is that use stays narrow: “Among adopters, scope remains limited: 57% use AI in three or fewer functions, most often Sales and Marketing (52%), Strategy (45%), and IT (41%).” When individual workers used it, it was “primarily for writing, document analysis, and information search.”

The second is that AI can enter a business from the bottom up. The researchers found that “worker use can occur without firm adoption, and vice versa.” Put simply, your team may already be using AI tools whether or not the business has decided anything.

A second NBER paper, Firm Data on AI, surveyed nearly 6,000 senior executives in the US, UK, Germany, and Australia. It’s a different survey asking a different question, and it found much higher use: “69% of firms actively use AI.” It also found that “executives report little own-firm impact of AI over the last 3 years, with nine-in-ten reporting no impact on employment or productivity,” and that executives’ AI use averages “only 1.5 hours a week.”

Taken together, businesses that use AI mostly apply it to specific tasks in a few parts of the business, and most of the executives surveyed haven’t seen a measurable effect yet. The roadmap below is built for that: one task at a time, with a number to judge it by.

The roadmap

1. Find out what’s already happening

Ask your team which AI tools they use, for what, and with what information. The Microstructure of AI Diffusion found worker-level AI use in 23% of firms (41% weighted by employment), and that use can happen without the business deciding to adopt AI. Your team’s answers tell you where the demand is, and they tell you what business information may already be sitting in outside tools.

2. Write a short data rule

Before anything else, decide what can and can’t go into AI tools: customer records, financials, anything covered by a contract or a confidentiality agreement. A single page is enough to start. For each tool you approve, read its own terms on whether your inputs are stored or used to train its models.

3. List the recurring work

For a week or two, write down the tasks that repeat, who does them, and roughly how long they take. You’re looking for work that’s repetitive and text-heavy, where someone can check the output quickly. Common examples are first replies to inquiries, turning call notes into a customer record, drafting proposals from past ones, and chasing follow-ups.

4. Pick one task in one part of the business

Start with one task. The NBER data shows most firms already using AI (57%) keep it to three or fewer functions, and sales and marketing is the most common function among them. Pick the task where the hours are, or where slowness costs you the most, like a lead that waits a day for a reply.

5. Measure it before you change it

Write down the number the task lives by: minutes per task, hours to first reply, errors per week, proposals sent. Without it, you’ll have no way to tell whether the AI version is better, and the decision will come down to impressions.

6. Build the simplest version that could work

Often that’s a well-written set of instructions in a tool your team already pays for. Sometimes it’s a fixed workflow that connects two systems. An AI agent that decides its own steps comes last, when the simpler options fall short. We walk through the difference, and what Anthropic, one of the companies that makes these models, advises, in what an AI employee actually is.

7. Put a person where mistakes cost money

Anything that goes to a customer should pass a person first, at least until the system has a track record you’ve checked. That protects the business while you learn what the tool gets wrong. Once that record is there, review moves to the exceptions it escalates, as covered in AI agents for small business. Anything that commits you to a price or moves money stays with a person.

8. Run it for 30 to 60 days, then decide

In our experience that’s long enough to see real volume and compare against the number from step 5. Then decide whether it stays. Dropping one is a normal result, and it still teaches you something about the next task.

9. Expand one task at a time

Only after the first task works does the second one start, with its own baseline and its own trial. Over time the tasks connect into a larger system, and each piece of it has already shown it works.

Where implementations tend to go wrong

These patterns come from our own work. We don’t know of a published study that measures them, so treat them as practitioner experience rather than data.

• Buying a platform before picking a task. The tool ends up looking for a problem, and nobody can say what it was supposed to improve.
• No owner. If no one person is responsible for the system working, it drifts, and people quietly go back to doing things by hand.
• Automating a process that doesn’t work by hand. AI makes a messy process faster without making it better. Fix the steps first.
• No baseline. Covered in step 5, and still the most common gap we see.

When to bring in help

Steps 1 through 5 need no outside help, and we’d suggest doing them yourself even if you plan to hire someone, because you know your work better than anyone you’d hire. Help starts to make sense when the task touches several systems at once, like your inbox and your CRM, or when nobody inside the business has the hours to build and watch the trial.

For more on which tasks are realistic, see what AI can actually automate in a business. If you’ve done the groundwork and want the build handled, that’s the automation work we do, sized to one task at a time.