Where Should a Small Business Start With AI? (Not With a Tool.)

You keep hearing that your competitors are "doing AI." Somebody on your team already pays for a chatbot subscription on a personal card. A vendor wants a demo. You feel behind, and the obvious move seems to be buying something.

Don't. Most AI rollouts stall because they start with a subscription and then go looking for a problem. The honest answer to "where should we start?" is boring: start with the repetitive, text-heavy work your team already hates, pick one workflow, measure it, and only then pick a tool.

You are also probably less behind than you think. The U.S. Census Bureau's business survey found overall AI use hovering between 17% and 20% of firms from December 2025 to May 2026, with fewer than 20% of firms of four or fewer employees reporting use. Plenty of your peers are still at the starting line. The ones who do well from here are the ones who start in the right place.

Why does starting with a tool go wrong?

A tool-first rollout usually looks like this. Someone buys seats. There is a lunch-and-learn. A few people use it to rewrite emails. Three months later nobody can say what changed, and the renewal notice arrives.

The problem is not the tool. It is that nobody defined the job. Without a specific workflow and a "before" number, you have no way to tell whether AI helped, hurt, or just added a browser tab. Starting with the work fixes that.

How do you find the work worth automating?

Run a one-week workflow inventory. It costs nothing but attention.

The one-week inventory

1. Pick the teams. Two or three is plenty: say, operations, sales and whoever answers the shared inbox.
2. Hand everyone a simple log. A shared spreadsheet or a paper sheet with five columns: task, who does it, how often, minutes per instance, and what it touches (customer data, money, contracts, nothing sensitive).
3. Log for five working days. Only recurring tasks. Ignore the one-off fire drills.
4. Flag the text. Mark tasks that are mostly reading, writing, summarizing, sorting or copying information between systems. That is where today's AI tools are most useful.
5. Ask one extra question. "Which of these would you happily never do again?" The answers are your shortlist, and they also tell you where you will get buy-in instead of resistance.

At the end of the week, multiply frequency by minutes for each task. You now have a rough weekly time cost for every recurring job, from the people who actually do the work. (People who manage a task always underestimate its exceptions. People who do it every day can list them from memory.)

How do you pick the first AI workflow?

Score each shortlisted task from 1 to 5 on four things. Higher is better for a first project.

Add the four scores. Your first project is the highest total, with one rule: if it scores 1 or 2 on risk or data sensitivity, it is not your first project, no matter how much time it costs. Start where mistakes are cheap. You can graduate to harder work once your team knows how these tools behave.

What does this look like in practice?

Three common first projects, one per department. These illustrate the method. They are not promises of results.

Operations: inbox triage

The work: someone reads every message to the shared info@ inbox, decides what it is (quote request, support issue, invoice, spam, job applicant) and forwards it.

The AI version: an assistant reads each new message and suggests a category and a route, and the person confirms or corrects it.

Why it scores well: high frequency, mostly text, and a wrong label is caught by a human before anything goes out. Watch the data column: if customers send account numbers or medical details to that inbox, check your tool's data terms first (more on that below).

Sales: proposal first drafts

The work: a salesperson rebuilds each proposal from an old one, swapping in the client name, scope and timeline.

The AI version: the salesperson pastes in call notes and your approved service descriptions, the tool produces a first draft in your template, and the salesperson edits and owns the final version.

Why it scores well: it is text-heavy and repetitive, and nothing goes to the client without a person reading it. Keep pricing and legal terms out of the generated part. Those stay in your locked template.

Account management: meeting notes to CRM

The work: after a client call, someone types up notes, pulls out action items, and updates the CRM record. Often late. Sometimes never.

The AI version: a transcription or notes feature (many video meeting and CRM tools now include one) produces a summary and action list. The account owner checks it and saves the approved version to the CRM.

Why it scores well: it is frequent and the "before" is usually bad, so even a modest improvement shows. Tell people when a call is being recorded, and check what recording consent rules apply where you and your clients are.

What should you measure before you start?

If you skip this, you will be arguing from vibes in a month. Before you change anything, capture a baseline for your chosen workflow over a week or two:

• Time per instance. Minutes to triage one email, draft one proposal, log one meeting.
• Volume. How many per week.
• Turnaround. Time from the trigger (email arrives, call ends) to the task being done.
• Error or rework rate. How often something is misrouted, corrected or sent back.
• One quality signal. Proposal win rate, CRM records updated within a day, customer response time. Pick one that matters to the business.

Then run the AI version for the same length of time and measure the same things the same way. Include the time spent reviewing and correcting AI output, because that is real time. If the numbers do not move, or review time eats the savings, that is a valid result. Stop, adjust, or move to the next workflow on your list.

What should you never paste into a consumer AI tool?

This is where small businesses get into trouble, usually through someone trying to be helpful on a personal account.

Consumer and business plans often come with different data terms. Anthropic, for example, says it uses consumer Claude chats to improve its models if the user allows it in their privacy settings, while for its commercial products it says it will not use inputs or outputs for training by default. Other vendors have their own settings and plan tiers. Read the terms for the exact plan your team uses, not the marketing page.

Until you have an approved business account with terms you have actually read, keep these out of every AI tool:

• Customer personal information: names with contact details, account numbers, anything a customer would be upset to see leaked
• Health, financial or legal records, and anything covered by a regulation you already follow
• Passwords, API keys and login details
• Contracts, pricing sheets and anything under an NDA
• Employee records, reviews and payroll
• Unreleased financials or plans you would not post publicly

The fix is not banning AI. People will use it anyway, just quietly. The fix is a short written policy, an approved tool, and a little training so everyone knows the rules. That is exactly what we cover in AI training for teams.

When do you finally pick a tool?

Now. You know the workflow, the baseline and the data involved, so the questions are concrete:

• Does software you already pay for (email, CRM, office suite, meeting tool) have a built-in AI feature that does this?
• Do the data terms for this plan fit the sensitivity score you gave the task?
• Does it connect to the systems the workflow touches, or will someone be copying and pasting all day?
• How does it handle your real examples, including the messy ones?

Test two options on the same set of real tasks and compare both against your baseline. There is no single best tool. There is the one that fits this workflow. (For a wider view of where AI is showing up in marketing specifically, see how AI is transforming marketing.)

What should you do next?

Start the one-week inventory on Monday. By Friday you will have a ranked list and a first candidate, which is further than a lot of companies get after a year of demos.

If you would rather have someone run it with you, our AI workflow audit maps your workflows, scores the opportunities and tells you honestly what to build first, including when the answer is "nothing yet." See the rest of our approach on the AI integration page, or book a free call and we will talk it through.