Short answer
Write one page, not a strategy deck. It should name the two or three business problems in scope, the first project with its measure of success, what you are deliberately excluding this year, the budget split between build and running costs, and who owns and approves. Plan the first project in detail and the next two in outline. Get something narrow into production in the first quarter, because a roadmap that produces nothing early tends to produce nothing at all — then let what you learn reorder the rest.
What the exercise is actually for
Most AI strategy work in small companies is imported from a much larger context. Maturity models, capability assessments and multi-year ambition statements exist because organizations of thousands need to coordinate. A company of ten to a hundred people does not have a coordination problem — it has a sequencing problem and a limited budget.
So the exercise has a narrower purpose: decide what to do first, decide what not to do, and set the conditions under which you would spend more. Done honestly it takes an afternoon and fits on a page. Done as a strategy program it takes a quarter, and the quarter is the cost — because the tools you assessed at the start have moved by the time you finish.
Three questions that replace a maturity assessment
- Which business problem, stated without mentioning AI? “Proposals take four days to get out and we lose deals to slower response times” is a problem. “Adopt generative AI in sales” is a technology looking for one. If the sentence needs the word AI to make sense, it is not yet a business problem.
- What is genuinely repetitive here? Enough to name one credible candidate. The criteria that rank candidates against each other belong to our process audit method — four of them — and are not worth re-deriving at roadmap level.
- What are we prepared to spend before requiring evidence? Deciding this up front converts an open-ended commitment into an experiment with a budget. It is also the question that makes stopping acceptable, which is what allows a first project to be genuinely small.
Answer those three and you have the substance of a roadmap. A maturity score would have added a number and no decision.
The four places value shows up
Across LYVIA’s own engagements with companies in this range, what actually gets deployed falls into four groups — a working classification from practice rather than a published framework. Each behaves differently on cost and on time to value.
- Individual productivity. Per-seat assistants that improve how work is done. Fast to start, recurring cost, and rarely visible in the accounts — the honest picture is in our guide to AI productivity tools.
- Process automation. Work removed from the company rather than done faster. Slower to start, higher effort, and the category where a structural saving actually appears. Candidates are catalogued in our list of use cases that survive the pilot.
- Better decisions. Forecasts, recommendations, and knowing sooner. Real value, hardest to measure, and the one most dependent on your data being consistent — see our note on AI in decision making.
- Commercial visibility. Being found and cited when buyers ask an assistant rather than a search engine. A different discipline from the other three, covered in our SEO and GEO content strategy.
Most companies should start in one of the first two and touch the fourth in parallel, because it compounds slowly and there is no reason to delay it.
A twelve-month sequence that survives contact
The shape below is how LYVIA typically phases a first year with a client — offered as a working pattern rather than a benchmark, and expected to be reordered by what the first project teaches.
| Step | What ships | What runs alongside |
|---|---|---|
| Step 1 — first quarter | One narrow process in production, with a measure agreed before the start. | An approved assistant on a company plan, and the one-page policy from our guide to implementing AI in your business. |
| Step 2 — second quarter | A second automation, chosen partly for whether it reuses the plumbing of the first. | Training for the roles whose work is most text-heavy, along the lines of our guide to training your team. |
| Step 3 — third quarter | Something with more surface area: project delivery, as in our note on AI for project management, or market monitoring, as in our guide to competitive intelligence. | Nothing new — this is the quarter to resist adding. |
| Step 4 — fourth quarter | Review what is running and retire what is not used. | Decide, on evidence rather than ambition, whether the next year needs an internal hire or an external partner. |
The load-bearing part is the first quarter. Everything after it should be treated as an outline that the first project has the right to rewrite.
Plan the first project in detail, the next two in outline, and nothing beyond. A twelve-month plan specified to the week in a field that moves this fast is a document you will either ignore or, worse, follow.
The budget split that matters more than the total
The total depends on what you are automating, so any figure quoted generically is noise. The split is where the recurring mistakes live, and three lines are consistently forgotten.
- Running costs. A workflow that calls a model has a bill proportional to usage. Forecast it before building and check it against reality after the first month, using the method in our ROI guide.
- Maintenance. Connected tools change their interfaces without asking. Something has to be reserved for keeping what you built alive, or it stops working in month five and nobody is funded to fix it.
- A second project. Spending the entire allocation on one ambitious build leaves nothing with which to act on what you learned — which was the main return on doing the first one.
Whether the money sits centrally or with teams is a separate question, and both work provided someone can see the total.
What to leave out on purpose
The exclusions section is the most useful part of the page and the one that is always missing. It converts a wish list into a plan, and it gives the owner something to point at when a new idea arrives in month three.
Common and sensible exclusions for a first year: building your own models, hiring an internal specialist before there is a stream of work to justify one, a company-wide rollout of anything, and any project whose value depends on data you do not currently have. Each may be right eventually. Writing down that they are out this year is what stops them from consuming the attention that the first project needs.
Two decisions sit adjacent to the roadmap and are worth taking deliberately rather than by drift. Whether a given need is met by a subscription or by something built for you is worked through in custom internal tools versus off the shelf. And if the ambition is to turn what you do into a product other companies pay for, that is a different business rather than a line on this plan — building a SaaS with AI sets out why. Sector-specific starting points, from sales to the pipeline after the lead, are catalogued in our list of use cases.
Reviewing it without rewriting it
Quarterly, against three questions rather than a full replan: what did we put into production, what did it change, and what did we learn that makes the next item wrong?
That third question is the one with teeth. A roadmap reviewed only against delivery becomes a schedule, and a schedule in this field is a promise to keep doing something after you found out it was not the right thing. The most useful outcome of a quarterly review is often a reordering, and it should be an easy thing to do rather than an admission of failure.
How these roadmaps die
- Nothing ships in the first quarter. The sponsor’s attention moves, the assumptions age, and the next planning cycle starts over. Across LYVIA’s own engagements this is the strongest single predictor of a roadmap that goes nowhere — an observation from client work rather than a published benchmark.
- The first project is the most important one. High stakes attract scrutiny and resistance, and a failure there ends the program. Choose the first for how much it teaches, not for how much it is worth.
- No owner. A plan belonging to everyone is executed by nobody — the operating-model question again.
- Buying a partner before knowing the problem. An agency engaged to define the strategy will define one shaped like the work it sells. The criteria for that decision are in our buyer’s guide to choosing an AI automation agency.
One constraint that belongs in the roadmap rather than at the end of it: if anything you build produces output used in the EU, the regulatory dates changed in July 2026 — what applies now for companies outside the EU.
Frequently asked questions
Does a small business really need an AI strategy?
It needs a short sequencing decision, not a strategy document. The purpose is to answer three things in writing: which problem is worth solving first, what you are deliberately not doing this year, and how much you are prepared to spend before requiring evidence. A company of ten to a hundred people that answers those three has everything a longer document would have given it, and can start this month.
What should the document actually contain?
One page is usually enough: the two or three business problems in scope, the first project with its measure of success, the list of things deliberately excluded, the budget split between running costs and build, and the names of who owns and who approves. Everything else — technology surveys, maturity scores, ambition statements — is content that reads well and changes no decision.
How much should we budget in the first year?
The useful discipline is the split rather than the total, because the total depends entirely on what you are automating. Reserve room for three things people forget: the running cost of what you put in production, the maintenance of it when a connected tool changes, and enough for a second project once the first has taught you something. Spending the whole allocation on one ambitious build is the pattern that leaves nothing to act on what you learn.
Should we start with productivity tools or with automation?
Both, because they are different budgets serving different purposes. Assistants improve individual work and cost per seat every month; automation removes work from the company and costs a build. The mistake is treating them as competing options — or expecting a subscription to produce the structural saving that only automation delivers.
How far ahead is it worth planning?
Plan the first project in detail, the next two in outline, and nothing beyond that. The pace at which tools change makes anything more specific a work of fiction, and a plan you have to defend for a year is a plan you will follow past the point where it stopped being right. Revisit the outline each quarter against what you actually learned.
What is the most common way an AI roadmap goes wrong?
It stays a document. As a rule of thumb from LYVIA's own engagements rather than a published benchmark, a roadmap that has produced nothing running in production within its first quarter usually never does — the sponsor moves on, the assumptions age, and the next planning cycle starts from the beginning. Anything that gets one narrow thing live early beats a better plan that does not.
If you want the one-page version for your company — first project, exclusions, budget split and measure — that is the shape of a first conversation. Book a call.
