AI Roadmap for Small Business: A 90-Day Plan to Audit, Pilot, Measure and Expand

The short answer

A workable AI roadmap for small business fits in 90 days and has four parts: two weeks auditing where time and leads actually go, one pilot chosen for countable value rather than excitement, six weeks running it with a baseline and a kill rule, and a final fortnight deciding what to expand, fix or stop. Most small businesses do not need a strategy document. They need one pilot that pays back and the discipline to measure it, because the evidence says that is exactly where most AI efforts fall over.

Why a roadmap and not a tool

Adoption is rising fast. Statistics Canada's analysis for the second quarter of 2026 found 19.2% of Canadian businesses had used AI to produce goods or deliver services in the previous 12 months, roughly triple the 6.1% of two years earlier. Among the smallest firms, those with one to four employees, the figure was 19.9%, so this is no longer a big-company story. By the third quarter of 2026, 25.2% of businesses said they planned to use AI in the next 12 months, up from 14.5% a year earlier.

The same data shows the gap a roadmap closes. Among the smallest AI-using businesses, only 24.0% had trained existing staff and 10.7% had used outside help, compared with 68.1% and 30.2% of large firms. Small businesses are adopting tools without changing how work is done around them, and the research on outcomes is blunt about where that leads: the MIT NANDA report covered by Fortune found about 95% of corporate generative-AI pilots produced no measurable profit-and-loss impact, blaming a learning gap and poor integration rather than the models. It also found purchased tools from specialised vendors succeeded about 67% of the time, while internal builds succeeded a third as often. We unpack those reasons in why most AI pilots never reach production. The roadmap below is designed around them.

The 90-day AI roadmap for small business

Days 1 to 14: audit

Do not start with AI. Start with where the hours and the leads go. For two weeks, have each person log the recurring tasks they do, how long each takes, and how often. Separately, trace what happens to every enquiry and every call for one week: when it arrived, when a person replied, what happened next. Our one-week automation audit gives the structure, and the free AI readiness scorecard scores the things that decide whether a pilot can work at all: whether your data lives in systems with usable exports and connections, whether anyone owns the process, and whether you have a policy for staff use of AI tools.

Output of the audit: a ranked list of five candidate processes, each with hours per week, who does it, which systems it touches, and what "better" would mean in a number.

Days 15 to 21: choose one pilot

Choose on four criteria, in order:

1. Countable value. Hours saved that will be redeployed, or revenue that is currently leaking (unanswered calls, slow follow-up). If you cannot name the number, it is not a pilot, it is an experiment.
2. Low blast radius. A mistake should be visible and cheap. Drafting replies for a human to send beats sending them. Booking appointments beats quoting prices.
3. Data you already have. The pilot should run on content and records you can hand over this month, not on a database you need to build first.
4. An owner. One person who will check it daily and has the authority to switch it off.

Typical first pilots that pass: after-hours call handling, enquiry acknowledgement and qualification, invoice reminders, review requests, and internal document search. Typical first pilots that fail: anything customer-facing with pricing authority, anything that needs three systems integrated first, and anything chosen because a competitor announced it. Before buying, check what is already inside tools you pay for; our post on AI already in your CRM and tools explains why that often beats a new vendor.

Days 22 to 75: run the pilot with a baseline and a kill rule

Before switching anything on, record the baseline: the hours, response times or conversion rates from the audit. Write down two numbers: the result that means "expand" and the result that means "stop". Then:

• Week 1: run in shadow mode. The system drafts, suggests or logs; people still do the work. Compare its output with theirs every day.
• Weeks 2 to 4: go live on a slice: one location, one service line, after-hours only. Keep a human in the loop for anything irreversible.
• Weeks 5 to 8: widen the slice, reduce the approvals where the error rate justifies it, and keep a log of every failure and what caused it.

Alongside, write a one-page acceptable-use policy for staff. Cybersecurity and privacy concerns were the most-cited barrier to AI use among Canadian businesses (13.4%) in the Statistics Canada analysis, ahead of cost (10.6%), and the fix is mostly rules: what data may go into which tools, who approves new tools, and how customers are told. Our AI policy template is a starting point.

Days 76 to 90: measure, then expand, fix or stop

Compare against the baseline with the same method you used in the audit. Count hours redeployed, not hours "freed". Count leads answered within the target time, bookings made, invoices paid sooner. Then make one of three decisions:

• Expand: the number beat the "expand" line. Widen to the whole business, reduce approvals further, and pick the next pilot from the audit list.
• Fix: the value is there but the failure log shows one recurring cause: bad data, a missing integration, unclear rules. Fix that one thing and run four more weeks.
• Stop: the number missed the "stop" line, or the failures are the kind that damage trust. Switch it off, write down why, and go to the next candidate. A stopped pilot that cost a month is a cheap education; a limping one that runs for a year is not.

Use the formula in how to calculate automation ROI for the expand decision, with the one-time cost you actually paid and the hours you actually measured.

What the second 90 days look like

If the first pilot expanded, the next cycle is shorter: the audit list already exists, the policy is written, and the team has seen one thing work. Pick the next candidate, repeat the shadow-then-slice pattern, and start connecting pilots where they touch: the call handler that books appointments feeds the CRM that triggers the follow-up that creates the invoice. That connected layer is what digital transformation means in practice for a small business, and it is where the compounding returns are. BDC's June 2026 study argues the same point at national scale: it estimates Canadian SMEs could unlock nearly $350 billion in growth if more firms reached the digital and AI maturity of the top 8%, and notes that only 30% of SMEs currently use generative AI. Its advice to the rest is to start small and integrate AI into operations, which is this roadmap in a sentence.

If the first pilot stopped, do not conclude that AI does not work for you. Conclude that one process was not ready, read the failure log, and pick the candidate with the cleanest data next. Our digital strategy work is mostly this: deciding what to build next, and why, from evidence rather than demos.

Key takeaways

• Audit hours and leads for two weeks before choosing anything; let the numbers pick the pilot.
• One pilot, one owner, a written baseline, and an expand line and a stop line decided in advance.
• Shadow mode first, then a slice, then wider; reduce approvals only as the error rate earns it.
• At day 90, expand, fix or stop. Each is a good outcome if it was measured.

Frequently asked questions

What should a small business automate with AI first?

Whatever the audit shows has countable value, low blast radius, data you already hold and a willing owner. In practice that is usually after-hours call handling, enquiry acknowledgement and qualification, invoice reminders, review requests or internal document search. Avoid starting with anything that has pricing authority or needs several systems integrated first.

How much does a 90-day AI pilot cost?

It depends on whether the pilot uses a feature already inside a tool you pay for, a vendor product priced per use, or a custom build connected to your systems. Count the one-time setup, the monthly fees and the owner's time, and set the budget against the hours or revenue in your audit. A fixed-price quote for the setup makes the ROI calculation a real number.

How do I know if an AI pilot is working?

Compare against the baseline you recorded before launch, using the same method: hours redeployed, response times, bookings, invoices paid sooner. Decide the "expand" and "stop" thresholds before you start, keep a failure log, and make the call at day 90. A pilot without a baseline cannot be judged, which is how limping pilots survive for years.

Sources

1. Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026 — Statistics Canada, 11 June 2026
2. Canadian Survey on Business Conditions, third quarter 2026 — Statistics Canada, The Daily, 31 August 2026
3. MIT report: 95% of generative AI pilots at companies are failing — Fortune, 18 August 2025
4. A $350B opportunity: Canada's next phase of growth to be driven by AI and digital technologies — Business Development Bank of Canada, 3 June 2026

How Flowmatix can help

We put this into practice for businesses in Toronto and around the world through our Digital Transformation Consulting, ai agents & ai solutions services. Want to know what it would take for your business? Get a personalised quote in about 3 minutes.