Writing an AI Roadmap for the Next Twelve Months

Most AI roadmaps are wish lists with dates

The typical AI roadmap we are shown is a slide with four columns labelled Q1 to Q4 and a dozen coloured boxes. Chatbot in Q1. Forecasting in Q2. Agent in Q3. 'AI-first operations' in Q4. It was made in a workshop, it was exciting, and by April nobody looks at it.

It fails for predictable reasons. No outcome is stated, so nobody can tell whether a box succeeded. Dependencies are missing, so the forecasting project discovers in month two that the sales data needs cleaning. And the whole year is committed at once, when anyone who has delivered AI projects knows the second half of the year will look different once the first half has taught you something.

This post is about the document itself: how to write a roadmap for a business that already has some AI activity and wants the next year to count. If you are starting from nothing, the quarter-by-quarter adoption roadmap covers the first year's sequence.

The one-page structure

The 'not doing' list is the most underrated section. Writing down that you are not building a customer-facing agent this year, and why, prevents the idea from being quietly restarted by someone who missed the meeting.

Start from outcomes, not technologies

Good outcomes sound like business targets that happen to involve AI. 'Handle 30% more order volume without adding to the customer service team.' 'Reduce stock write-offs in chilled products.' 'Cut average quote turnaround from three days to one.'

Bad outcomes are technology deployments. 'Implement an AI assistant.' 'Launch a machine learning platform.' 'Adopt agentic AI.' These cannot fail, which means they cannot succeed either.

With outcomes set, each candidate initiative earns its place by answering one question: which outcome does this move, by roughly how much, and how will we know?

Choosing what goes in

You will have more ideas than capacity. Score each candidate quickly on four things, and be honest:

1. Value: how much it moves an outcome if it works
2. Confidence: how likely it is to work, given the data and process as they are today
3. Readiness: whether the data, systems and people it needs are available now
4. Effort: build time, internal time and running cost

High value and high readiness go into committed initiatives. High value but low readiness usually means a foundation project comes first. Low confidence but high potential belongs in options, with a small test to raise confidence. Balancing these as a portfolio of quick wins and bigger bets keeps the year from being either timid or reckless.

Then check capacity. For a 100-200 person business, two committed AI initiatives at a time is often the realistic maximum, because each needs an internal owner and expert time. If the roadmap has six running at once, it is a wish list again.

Commit to two quarters, keep the rest as options

Plans for AI work degrade quickly with distance. The tools change, the first projects reveal data problems nobody anticipated, and a business priority shifts. A roadmap that fixes the whole year pretends none of that will happen.

• Quarters one and two: committed, funded, owned, with specific deliverables
• Quarter three: likely, with a named decision point and the evidence needed
• Quarter four: options only, deliberately loose

Each quarterly review rolls the window forward. By the end of the year you have committed four quarters of work, each decided with the best information available at the time.

A roadmap is a set of decisions you have made and a list of decisions you have deliberately postponed. Both are useful.

Write the stop rules now

Every committed initiative needs a condition under which it stops or changes course, agreed before it starts. For example: 'If extraction accuracy on 300 real invoices is below 90% after the first build phase, we pause and reassess scope.' Or: 'If fewer than half the sales team use the proposal drafting tool weekly after two months, we investigate before extending licences.'

Stop rules feel pessimistic to write. They are the opposite. They let a team start ambitious work knowing that failure will be caught early and cheaply rather than defended for a year.

An illustrative example

An illustrative 140-person specialist insurance broker might end up with a roadmap like this:

• Outcomes: cut submission preparation time for brokers; reduce errors in policy data entry; answer routine client queries faster
• Committed Q1-Q2: document extraction from client submissions into the broking system; AI-assisted drafting of renewal summaries, with broker review
• Foundations: clean up client records across two systems; approved tools and policy refresh; logging standard for AI systems
• Options for Q3-Q4: a client query assistant, pending results of a four-week test on past emails; renewal risk scoring, pending data quality work
• Not doing: fully automated quoting, because errors are costly and regulatory expectations are high
• Review: first week of each quarter, management team, one hour

When we help write roadmaps like this at SpiderHunts, through our enterprise AI service, the most valuable contribution is often in the readiness and effort scores. Businesses tend to overestimate readiness, and a realistic estimate of the data work changes the order of the whole plan. For more background on the options, our AI automation guide covers where each kind of project fits.

Frequently asked questions