The AI roadmap for a small business is one build a month

If you run a small business and you want an AI roadmap, write one page and then ship one automation every month. The twelve-month roadmap deck is the artefact that kills most AI programmes, because it commits budget to a sequence of decisions you are not yet qualified to make, and it produces nothing measurable until quarter three. A cadence produces a working system every thirty days, and each shipped system tells you what the next one should be.

That is not a philosophical preference. It is what the 2025 and 2026 data on stalled AI programmes points at, and it is how we run our own group.

Why AI roadmap decks stall before the first build

Roadmap decks stall because they buy sequencing instead of output. McKinsey reported in November 2025 that 88% of organisations use AI in at least one function, but only about a third have begun scaling it across the enterprise and just 7% describe AI as fully scaled. Planning is not the constraint in that picture. Production is.

Gartner predicted in June 2025 that more than 40% of agentic AI projects will be cancelled by the end of 2027, naming escalating costs, unclear business value, and inadequate risk controls as the causes. Two of those three are things you find out by building. You cannot cost a build accurately, or discover where it needs a guardrail, from a slide.

The IBM Institute for Business Value, with Oxford Economics, surveyed 2,000 CEOs across 33 countries in early 2025 and found that only 25% of AI initiatives had delivered the ROI expected of them, and only 16% had scaled enterprise-wide. In the same study, 61% of CEOs said they were actively adopting AI agents. Ambition at 61%, production at 16%. The roadmap is rarely the thing that is missing.

What one build a month actually looks like

A shipping cadence means: one job with a clear input and a clear output gets automated, tested, and put into production inside a calendar month, then it gets monitored while the next one is built. No phase gates. No steering committee. The month is the unit of planning, and a working system is the deliverable.

We run our own group this way. As of July 2026 our fleet is 41 registered agents, each with a named owner, a defined fence, and a run log, and in one measured 13-day stretch they ran 5,450 times with zero failures. That fleet was not designed in advance. It accumulated, one build at a time, and the governance layer got built when agent count made it necessary rather than when a deck said month four. The discipline behind that is written up in how we register, fence, and monitor every agent before it runs.

The important property of a monthly cadence is that being wrong is cheap. A build that turns out to solve the wrong problem costs you a month and teaches you something concrete about your own operation. A twelve-month roadmap that is wrong costs you the year, and you usually find out at the end of it.

Roadmap deck versus shipping cadence

	Twelve-month roadmap	One build a month
First measurable output	Quarter three, if the plan survives	Day 30
What month one produces	A prioritised list and a budget request	A system doing a job in production
Cost of being wrong	The programme	One month
What you learn	Whether people agreed with the plan	What your data, tools, and staff actually do
What compounds	Slide decay	Working systems, plus the plumbing the next build reuses
Who has to be available	A committee, repeatedly	One person who can approve in writing

The three builds most small businesses ship first

Almost every small business we look at has the same first three candidates, and none of them require a strategy exercise to find.

Intake and triage. Inquiries, orders, or applications arriving through a form, an inbox, and a phone, then sitting until someone has time. Capture, dedupe, classify, route, draft the reply. Clear input, clear output, measurable in the first week by how long a lead waits.

The number nobody trusts. Every business has one: a margin, a return on ad spend, a stock position that finance and operations report differently. Automating the reconciliation gives you a number computed the same way every day from the actual ledger, which changes decisions immediately.

Repetitive production work. Specification sheets, listings, quotes, articles, compliance paperwork. Work that follows a template and eats a day a week. The system drafts, a human approves, the approval takes ten minutes instead of four hours.

Each of these is a month. None of them needs the other two to exist first. That independence is the point: a roadmap that requires its own sequence to be correct is fragile, and most failed AI pilots die in the gap between a plan and an operation rather than on the technology.

What to write down instead of a roadmap

One page, and it holds four things.

A ranked list of candidate jobs, each with an estimate of hours currently spent on it per week. The next build, named specifically, with the month it ships. The owner: one person who can approve in writing and grant access once. Kill criteria: the condition under which you stop the current build rather than finish it out of sunk cost.

That page is a live document, re-ranked after every build, because the build teaches you things the ranking did not know. This is the actual job of an AI lead, and it is why we price the fractional AI officer engagement at $2,900 per month with one automation shipped in each of them, cancel monthly. The monthly cancel is load-bearing: it makes shipping the only thing that renews the engagement. The full scope of the role is in what a fractional AI officer owns and reports on.

The honest limit on these numbers

The failure and ROI statistics above are enterprise-weighted. IBM surveyed CEOs, McKinsey surveyed nearly 2,000 respondents skewed to large organisations, and a 12-person company does not carry the coordination overhead that sinks a 12,000-person programme. Treat them as directional.

The genuinely small-business data is thinner and more sobering. The US Census Bureau, publishing Business Trends and Outlook Survey results in May 2026, found 37% of firms with 250 or more employees reported using AI, against under 20% of firms with four or fewer employees. Adoption tracks headcount, which mostly tracks who has someone whose job it is to ship this.

On the optimistic side, Google Cloud published research in September 2025, conducted by the National Research Group across 3,466 senior leaders, reporting that 74% of executives saw ROI within the first year of a generative AI initiative. Attribute that one carefully: it is vendor-sponsored and self-reported by executives. It is still useful as a time-to-value signal, because it says value shows up in months. If value arrives in months, planning in years is the wrong shape.

Boston Consulting Group surveyed 1,250 senior executives for a September 2025 report and found 5% achieving AI value at scale against 60% with little or nothing to show, and put 70% of AI’s available value in core functions such as sales, manufacturing, supply chain, and pricing rather than in novel projects. The boring internal jobs are where the money is. Those are also the jobs you can finish in a month.

How to start on Monday

List every job in your business that a competent assistant could do from a checklist. Sort by hours per week. Take the top one that has a clear input and a clear output, and give yourself thirty days and one named owner. Ship it, run it for a fortnight, then re-rank the list with what you learned. That is the roadmap.

If you want the cadence run for you rather than by you: describe the job that should stop eating your week, and you get a written plan back within one business day. No call, no discovery session, no deck. Start async.