You have decided to do something with AI this year. You have read the articles, sat through two vendor demos, and signed up for a tool you have opened twice. What you do not have is a plan, a developer, or a clear idea of what to do on Monday morning.
This is an AI roadmap for small business owners in exactly that position: profitable, busy, not technical enough to build anything. It covers ninety days and it is deliberately modest, because in our experience the businesses that get real value from AI did one small thing properly before doing anything ambitious. By the end you will know what to do in each month, what to buy, and the point at which a developer becomes the cheaper option.
What a roadmap actually is here
An AI roadmap for small business is not a list of tools. It is a sequence of small commitments, each producing something you can measure and each making the next one easier. The order matters more than the contents.
The analogy is hiring your first office manager. You do not hand them the whole company on day one. You give them one job, you agree how you will know it is going well, and you check in weekly. When that works, you add the next job. The roadmap treats AI the same way: one process at a time, a named person responsible, a number that says whether it worked.
Two terms you will meet. Automation means software that does a fixed sequence of steps when something happens: a form is submitted, so a row is added to a spreadsheet and an email goes out. AI means a language model (the kind of system behind Claude, GPT-class models and Gemini) making a judgment inside that sequence: deciding whether the form is a sales lead or a complaint. The roadmap adds automation first and AI second, because automation is predictable and AI is not, and you want to learn on the predictable one.
The 90 days, step by step
1. Month one, week one: pick one process that is boring, frequent and annoying
Choose one process. Not the most valuable, not the most exciting: the one that happens many times a week, follows the same steps every time, and that somebody on your team quietly resents. The quote follow-up. Copying the web form into the CRM. Chasing the invoice at day 30.
The test for a good first process: you can write its steps on an index card, it happens at least ten times a week, and nothing terrible happens if it runs wrong once. If you cannot write the steps down, the process does not exist yet and no software can run it. AI Readiness Checklist for Small Business, 20 Points to Pass covers the other things that need to be true before you start.
Resist the big one. “Answer all customer emails with AI” is a year of work disguised as a sentence. “Send the day-30 reminder automatically” is a Tuesday afternoon, and it teaches you everything the big one will need.
2. Month one, week one: name the owner and the number
One person owns the process: whoever does it by hand today, because they know where it breaks. Their job is not to build anything. It is to say whether the automated version does what the manual version did, and to be the person who gets the alert when it fails.
Then the number. Before anything is built, write down the one measure that will say it worked: hours saved per week, days from job closed to invoice sent, the share of leads called back within an hour. Measure it by hand for two weeks. Without the before number, you will never know whether the after number is good.
3. Month one, weeks two to four: build the first automation with a no-code tool, and keep it dumb
Pick a no-code automation platform: Zapier, Make, or n8n’s hosted plan. n8n vs Make vs Zapier for a Business Without a Developer compares them; for a first project any of the three is fine. Connect the trigger (the thing that starts it) to the action (the thing that happens), and nothing else.
Keep the first version dumb: no branching, no AI, no clever conditions. The owner watches it run for two weeks, checks every output against what they would have done, and writes down every case where it was wrong. Those cases are the specification for version two.
The point of month one is not the time saved. It is that your business now has one process that runs without a person, one person who understands it, and one number that went from a before to an after. Everything that follows stands on it.
4. Month two: add the second automation, and the logs
Choose and build the second process the same way, but add what was missing from the first: a record of every run, and an alert when a run fails.
Every automation platform can log each run and send a message when something breaks. Turn that on, and send the alert to the owner, not a shared inbox nobody reads. Once a week, the owner reads the log: how many runs, how many failed, why. This habit separates automations that survive from ones that silently stop in March and are discovered in June. Automation Error Handling for Businesses Tired of Silent Failures covers the failure modes; the short version is that something upstream always changes, and the log is how you find out the same day.
Add the same log and alert to the first automation. Then write both processes down in one shared document: trigger, steps, owner, where the log is, what to do if it fails. Two pages. This is the beginning of your business owning its automation instead of one person carrying it in their head.
5. Month two: fix the data the automations revealed
Your first two automations will expose something: customer records with three spellings, jobs marked complete twice, an “email” field holding a phone number. A person working by hand silently corrects these. Software does not.
Spend part of month two cleaning what the automations touch, and only that. Agree one format for the fields involved, make them required where they should be, delete the duplicates the automation tripped over. Do not clean the whole business’s data; clean the strip the two live processes run across. This is where the AI step in month three either works or does not, because a model given messy input produces confident, messy output.
6. Month three: add the first AI step, inside an existing automation
Now the AI. Not a new project: a judgment step inside one of the two automations that already run and have logs. The best first candidates are classification (is this form a lead, a complaint or spam), extraction (the amount, date and vendor from an emailed invoice), summary (a call transcript into three lines for the CRM), and drafting (a first reply for a person to edit). 25 AI Workflow Automation Examples for a Real Business has more by department.
The rule for the first AI step: it recommends, it does not act. The classification sets a label a person sees. The extracted fields go into a “check me” queue. The draft reply sits unsent. Keep a tally: right, wrong, unsure. When it has been right for a month on the cases that matter, consider letting it act on the routine ones.
Buy, do not build, at this stage. Zapier, Make and n8n all have built-in AI steps that call OpenAI, Anthropic or Google models. A prompt (the written instructions the model works from) of ten plain sentences plus three examples of good output is enough for a first classification or summary. Keep a copy in your process document and change it one line at a time, watching the tally.
7. Month three: write the one-page rule for what the AI may and may not do
Before the AI step touches anything a customer sees, write the rule. What it may do alone (label, sort, draft, summarise). What it may do with approval (send, update a record, quote a price). What it may never do (move money, promise a delivery date, change a customer’s details).
This is a business document, not a technical one. It stops the slow creep where each convenience expands what the AI may do until it is sending under your name with nobody watching. It is also the first thing a good developer will ask for.
8. Day 90: the review, and the decision
At the end of the three months, sit down with the owner and the numbers: two automations with logs, one AI step recommending, a before and after for each, a document, a rule. Then decide one of three things. Keep going, one process a month with the tools you have. Stop, because the numbers say it was not worth it (a real outcome, learned cheaply). Or bring in a developer, because the next thing is beyond what no-code tools do well.
Picture a business like this one
The business below is a composite of the kind of company that writes to us, not a client. The numbers describe the shape of the problem, not a case study.
Picture a business like this one: a commercial cleaning company, 28 staff, two in the office, running on a job-scheduling app, QuickBooks and a shared inbox. The owner had signed up for three AI tools and used none.
Month one, they picked the quote follow-up: quotes went out, and unless someone remembered, nothing happened. The office manager owned it. The before number: 31 percent of quotes got any follow-up. A Zapier automation sent a plain follow-up three business days after each quote and a second at ten. No AI, no branching.
Month two, they added the day-30 invoice reminder from QuickBooks, turned on run logs and failure alerts for both, and cleaned the customer email field, where forty records held a phone number. They wrote the two-page process document.
Month three, they added an AI step to the inbox: every email got a label (new enquiry, existing job, invoice question, other) and new enquiries got a drafted reply, unsent. A month’s tally showed it right on nearly all, wrong mostly on emails that were two things at once. The owner wrote the rule: the AI drafts, people send.
At day 90, the follow-up number was above 90 percent, two quotes had converted that would have gone quiet, and the office manager estimated four hours a week back. The next thing they wanted, a customer portal, was clearly a developer job. They went looking for one with a process document, logs and a rule in hand, which made the scope short.
What it costs to run
The tools in an AI roadmap for small business are cheap by design. A no-code automation platform at these volumes runs from free to roughly $20 to $50 a month depending on how many steps run; check the current pricing page, because plans change and each counts usage differently. The AI step is billed by usage, and at a few hundred emails or forms a month it is typically single-digit dollars.
If you later self-host n8n to avoid per-step pricing, add a $10 to $20 a month server (Self-Host n8n for Your Business Without Being Held Hostage). The real cost across the 90 days is the owner’s time: expect a few hours a week in month one, tapering to an hour a week of log review by month three.
The mistakes we see most
Starting with the customer-facing thing. The website chatbot, the AI phone line. Highest risk, hardest to measure, first thing everyone wants. Do it fourth, not first.
No before number. Ninety days later nobody can say whether it helped, so the decision to continue becomes a feeling.
Letting the AI act on day one. The draft gets sent before anyone has counted how often it is right.
Skipping the logs. The automation stops silently. Nobody notices until a customer does.
One person carrying it in their head. They leave, and nobody understands the automations. The two-page document is the fix.
When to bring in help
Everything in this AI roadmap for small business can be done by an owner or office manager with a no-code tool. The moment a developer becomes the cheaper option is recognisable: two systems need to talk and there is no ready-made connector; you want the AI to act rather than recommend, with proper approval gates; the no-code workflow has grown to forty steps and breaks when anyone breathes on it; or the next thing is real software, a portal, a tool your staff use all day. Build vs Buy AI Tools for Your Business, the Decision Rules covers the decision, and How to Hire an AI Consultant Without Getting Burned covers how to run the first engagement without getting burned.
Arriving at that conversation with 90 days of logs, a process document and a written rule changes it. The scope is shorter, and you can judge the proposal because you have already run the small version yourself.
Levelbrook builds the next stage for businesses like this: the integrations, the AI steps that act safely, the internal tools, all in accounts you own, fixed price from a written scope. If you have done the first 90 days, or want a hand deciding what your first process should be, the form below is how the conversation starts.