AI transformation roadmap: four phases from first department to whole company.
A roadmap that is only a list of use cases rarely survives the first quarter. This one is built around four phases, each ending at a gate with something running and a named person who signs it off.
What does an AI transformation roadmap look like?
An AI transformation roadmap sets out four phases over about twelve months. Map takes four weeks: hours per process, an assist or take-over decision per task, and the rules. Prove takes 30 to 60 days: one department live with agents and tests. Scale takes 90 days or more: further departments on the same knowledge and environment. Run is monthly upkeep with quarterly reviews.
New to the term? What is AI transformation →
Christos Papadimitriou, theagency47 · Updated October 2026Why a roadmap fails when it is a list of use cases
Most AI roadmaps start with a workshop and end with a prioritised list: twenty use cases, scored for value and effort. The list is not wrong. It is just not a plan. It says nothing about which department goes first, what knowledge the agents will read, who approves their output, or what has to be true before the next step starts. Six months later the company has a few pilots that each solved one problem on their own and cannot be extended.
A working roadmap answers four different questions: what goes first, what it is built on, who signs off each step, and what happens after launch. That is what the four phases below are for. They sit above our per-project delivery method; inside each phase, individual agents are built the way our guide to building an AI agent describes.
Phase 1, Map (4 weeks)
The aim is to see where the hours go and decide what agents will and will not do, before anything is built.
- Hours per process, per department. Interviews and a short time tally put a number of hours a week against each recurring process. Estimates are fine at this stage; they are checked later against real data.
- Assist, co-work or replace, decided per task. For each task in scope: how rule-based is it, what does an error cost, who approves the output, how often does it arrive. The answers decide whether an agent assists a person, works alongside them or takes the task over. The test is described in assist or replace, role by role.
- Autonomy and approval rules. What each agent may do alone, what always needs a person, and who owns it. Written down, not implied.
Gate: a signed phase plan with a fixed price per phase, a named executive sponsor, and the first department chosen.
For mid-size companies, Map is our four-week AI transformation assessment.
Phase 2, Prove (30 to 60 days)
One department goes live. The point is to prove the model works in your company, with your data and your people, and to build the foundation every later department will reuse.
- Pick the first department on three tests: high volume of work, mostly rule-based, and a low cost of error. Sales operations, tier-1 customer service and finance operations usually pass all three.
- Company brain, stage 1. The knowledge those agents need, written once as plain files the team can read and correct. This becomes the first layer of the AI company brain.
- Eval suite before launch. Every agent passes a set of test cases with known answers before it touches live work, and again after every change. Our 20-case eval checklist shows the structure.
Gate: the agreed KPI has moved against its baseline and the acceptance tests pass. If not, fix or stop before scaling.
Phase 3, Scale (90 days and more)
- Second and third departments on the same brain. New agents read the same knowledge layer, so each department adds to it rather than starting again.
- Permissions per audience, logs per agent. Once a customer-facing agent sits next to internal ones, the brain needs rules about who may see what, and every agent's actions need a log someone reviews.
- One environment. Staff sign-in, approvals and model budgets live in one place the company controls.
Gate: steering committee sign-off on each new department, based on the previous department's results.
Phase 4, Run (monthly)
- Health checks and knowledge upkeep. Agents are monitored, test suites re-run, and stale knowledge goes back to its owner.
- Quarterly review: what to hand to agents next. The map from Phase 1 is revisited with real numbers, and the next tasks are chosen.
A 12-month timeline
A typical sequence for a company that starts with one department and adds two more in the first year. Larger companies often run Prove inside a longer scoping and contracting cycle; the order stays the same.
| Phase | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | M10 | M11 | M12 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Map | ||||||||||||
| Prove: first department | ||||||||||||
| Scale: departments 2 and 3 | ||||||||||||
| Run: upkeep and new agents | ||||||||||||
| Quarterly review |
Roles: who does what
| Role | Who | Responsible for |
|---|---|---|
| Executive sponsor | CEO, COO or a managing director | Says yes, chooses the first department, signs each gate, chairs the steering group |
| Process owner | Head of the department in scope | Brings the hours and examples, sets the KPI, proves the result |
| Agent owner | One named person per agent | Reads the logs, approves changes, can switch the agent off |
| IT and security | CIO, IT lead or an outside provider | Access, sign-in, logs and the off switch |
For a mid-size company these are often four people wearing other hats. That is fine, as long as each role has a name. Who decides what is covered in CEO vs CIO.
One-page roadmap template
Copy this into a document and fill it in during Phase 1. If a line stays empty, that is the next conversation to have.
| Line | Your answer |
|---|---|
| Executive sponsor | Name |
| First department and why | Hours a week, how rule-based, cost of an error |
| Tasks in scope | Each with: assist, co-work or take over |
| KPI and baseline | One number, measured before launch |
| Approval rules | What always needs a person |
| Agent owners | One name per agent |
| Knowledge needed | Documents, policies, prices, owners |
| Prove gate date | Day 30 to 60 |
| Next two departments | In order, with the reason |
| Quarterly review dates | Four dates in the calendar |
Mid-size company deciding where to begin? Read AI for mid-size companies: start with one department. Want help filling it in? An AI Audit is the short version of Phase 1.
Questions about AI transformation roadmaps.
How long should an AI transformation roadmap cover?
Twelve months is a practical horizon: long enough to take three departments live and short enough to plan with real numbers. Review it every quarter rather than writing a longer plan up front.
What is the difference between an AI strategy and an AI transformation roadmap?
A strategy says why and where AI matters to the business. A roadmap says which department goes first, what each phase delivers, who signs off each gate and when. A roadmap without a strategy drifts; a strategy without a roadmap stays on paper.
Which department should go first?
The one with a high volume of frequent, rule-based work and a low cost of error, where the hours are easy to measure. Sales operations, tier-1 customer service and finance operations are common first choices.
Do we need a data platform before we start?
No. Agents read the documents and systems you already have. What you do need is the relevant knowledge written down somewhere an agent can read it, which is part of Phase 2.
What happens if the first department does not prove out?
The Prove gate exists for that case. Fix the scope, the knowledge or the rules and re-test, or stop before spending on scale. A roadmap that cannot say no at a gate is not a roadmap.
Related reading
What is AI transformation?
The definition, and what changes first.
How we work
The five-phase method we use inside each build.
How to build an AI agent
The six-phase playbook for one production agent.
AI governance framework
Owners, autonomy levels and approval rules for every agent.
