There is no shortage of “AI transformation frameworks” online. Most of them are consulting output: an impressive-looking diagram, five to eight phases, colour-coded, with deliverables that correspond suspiciously well to the consultancy’s service offering.
This one is not that. It’s the roadmap I actually used, in order, to build the AI Factory that shipped 55+ autonomous agents at 300% ROI. Every phase is described with what actually happened, what the gate criteria were, and where programmes commonly die.
If you run one programme against this roadmap, it will be more honest than most of what you’ll read elsewhere. It will also be harder, because honest is harder than impressive.
Seven phases. Some run in parallel; most don’t. The ordering matters.
Phase 1, Framing (weeks 0–2)
Outcome: written decision about whether AI is a project or a capability.
This is the first phase because the rest depends on it. An organisation that frames AI as a project will budget it as a project, staff it as a project, and kill it like a project. An organisation that frames it as a capability will build an operating line, a standing team, and a backlog.
Specifically:
- Budget classification. Operating line, not project line. Three-year horizon minimum.
- Ownership. Named executive sponsor, one. Not a committee.
- Charter, but the right kind. Not a project charter; a factory charter. Scope statement, principles, out-of-scope list, first-year ambition.
- Success metrics. Throughput-based (production agents in operation, hours saved) not milestone-based (phases completed, deliverables shipped).
Gate: the CFO and the CEO sign the framing document. If either hesitates on “operating line”, the programme is not ready to proceed. Either convince them or pause. Building on a project framing and hoping to convert later has a failure rate approaching 100%.
Phase 2, Process inventory (weeks 2–12, parallel with phase 3)
Outcome: a scored, ranked backlog of candidate processes across priority functions.
This is the phase most programmes try to skip, and the phase that most reliably separates the 5% of programmes that ship from the 95% that stall.
At the Oman conglomerate we mapped 250+ processes across 12 functions, documented 165 SOPs, and scored every process on automation potential, risk, and marginal ROI. The methodology is in Process Inventory: The Moat Nobody Maps.
Key decisions during this phase:
- Scope. Pick two functions first (usually Finance and HR or Shared Services). Expand later.
- Team. One AI lead plus two process analysts per function. Internal people, not consultants.
- Method. Three passes: breadth interviews, depth walkthroughs, scoring. Output is a working spreadsheet, not a slide deck.
Gate: the scored backlog is published, and the AI steering group has approved the first-cohort priorities (5–10 processes). No agent build starts before this gate.
Common failure: skipping or truncating this phase to “move fast”. This is the single most expensive decision a programme can make. The time saved is measured in weeks; the cost is measured in years of wasted rollout.
Phase 3, Platform and governance foundation (weeks 2–16, parallel with phase 2)
Outcome: platform stood up, governance controls operational, first agent-development environment ready.
This phase runs in parallel with phase 2 because the inventory work doesn’t need the platform yet, and the platform build is a long-lead item. By the time the first-cohort backlog is ready, the platform should be ready to receive it.
Platform components (for an enterprise-grade deployment):
- Orchestration: n8n or LangGraph, depending on agent complexity profile.
- LLM plumbing: LangChain or custom.
- Retrieval: permission-aware enterprise RAG (Pinecone, Weaviate, pgvector).
- Integration: MCP servers for each system agents need to reach.
- Observability: from day one, not added later.
- Cost controls: per-workflow budgets and token-spend attribution.
Governance components (see AI Governance on ISO 9001 for the full treatment):
- AI policy, approved at board level.
- AI procedure, integrated into the QMS.
- AI register.
- Risk-assessment template (EU AI Act-aligned).
- Incident-handling process.
- Management review cadence.
Gate: platform is running end-to-end with a test agent; governance controls are audit-ready; first real agent can start development.
Common failure: treating governance as phase-4 work. Bolted-on governance is the reason production rollouts stall at the legal review stage. Build it into the platform from the start.
Phase 4, First-cohort agents (weeks 12–28)
Outcome: 5–10 agents in production, measured, generating value.
Now the factory output starts. First-cohort agents are chosen from the top of the scored backlog: high automation potential, moderate risk, high marginal ROI. Almost always direct-labour-substitution agents in Finance or Shared Services.
Build pattern per agent:
- Week 1: detailed design, integration mapping, evaluation plan, governance sign-off.
- Week 2–3: build, internal testing, evaluation harness calibration.
- Week 4: UAT with real business users, governance review, production deployment with human-in-the-loop on day one.
- Week 5–6: oversight scaling down as evaluation metrics stabilise; value measurement begins.
First-cohort ambition: 5–10 agents live by end of phase. Each with named owner, criticality tier, SLOs, cost envelope, evaluation harness, audit trail.
Gate: at least five agents running in production for 30 consecutive days; measured value delivered; first ROI report to the steering group.
Common failure: over-engineering the first cohort. The goal isn’t elegance; it’s proof of throughput. Ship five simple, reliable agents faster than you ship two complicated ones.
Phase 5, Adoption (weeks 18–32, overlapping phase 4)
Outcome: agents not just deployed, but actively used, at target utilisation.
This is the phase most easily measured in its absence. An agent “deployed” but not used is a cost centre. Adoption is a programme discipline, not a training exercise.
Adoption work per agent:
- Champions. One per function. Internal people, respected by their peers, bought into the specific agent they’re championing.
- Co-design sessions. SOPs updated with the practitioners who actually run the process. Not handed down.
- Weekly usage targets. Explicit. Measured. Reported.
- Feedback loop. Every week, champions surface friction; friction becomes agent backlog items.
- Storytelling. The first measurable wins get told across the organisation. Repeatedly. By name.
Gate: target utilisation reached (typically 70%+ of the eligible population using the agent weekly by end of phase 5).
Common failure: treating adoption as a launch event. Launch is a day; adoption is a quarter. Organisations that hit 70% utilisation compound. Organisations that hit 20% quietly de-fund.
Phase 6, Scale-out (weeks 28–52)
Outcome: second and third cohorts deployed; factory producing agents at steady-state cadence.
By this phase the platform is proven, governance is operational, the first-cohort agents are generating measurable value, and adoption is strong. The factory can now move from hand-crafted to production cadence.
What steady-state looks like:
- Cadence: 3–6 new production agents per month.
- Unit economics: build cost per agent dropping (platform amortises, patterns emerge, reusable components grow).
- Function coverage: expanding from initial 2 functions to 6–8 over this phase.
- Governance coverage: mature, reviewed at the quarterly management review.
- Cost base: trending down per agent, up in absolute terms (more agents, lower per-agent cost).
Gate: steady-state cadence maintained for at least two consecutive quarters; function coverage expanded; management review cycle operational.
Common failure: scaling platform and team ahead of demand. Over-built platforms collapse under the weight of their own complexity. Let the demand pull the capability; don’t push capability at the business.
Phase 7, Embed and compound (year 2+)
Outcome: AI is an organisational capability, not a programme. The factory runs itself; the backlog replenishes; the P&L impact compounds.
This is the phase most programmes never reach, because they were killed or de-funded somewhere in phases 2 through 4. But if you reach it, it’s the payoff.
What embedding looks like:
- AI roles are normal roles. AI product managers, agent engineers, process analysts. On the org chart. Progression path mapped.
- Agents are part of the work. Onboarding includes agent use. SOPs assume agent assistance. New processes are designed with agents in mind.
- The backlog is the roadmap. No separate “AI strategy” is needed; the backlog, scored and prioritised, is the plan.
- The P&L impact is in the operating model. Hours saved, errors avoided, capacity unlocked, measured in the management accounts alongside everything else.
- The governance is routine. Quarterly reviews are unremarkable. Audits pass. New regulations are absorbed without drama.
This is what “AI transformation” actually means, once stripped of the consulting vocabulary. It’s not a destination; it’s a reframing of how work gets done.
What the roadmap looks like on a calendar
For a typical mid-to-large enterprise starting from zero:
| Months | Phase | Key deliverables |
|---|---|---|
| 0–1 | 1. Framing | Factory charter, exec sponsor, budget classification |
| 1–4 | 2. Process inventory | Scored backlog for two priority functions |
| 1–4 | 3. Platform and governance | Platform stood up, governance controls operational |
| 3–7 | 4. First cohort | 5–10 agents in production |
| 5–8 | 5. Adoption | Target utilisation reached |
| 7–12 | 6. Scale-out | Second and third cohorts, steady-state cadence |
| Year 2+ | 7. Embed and compound | Capability, not programme |
Year 1 is where you find out whether the organisation has the discipline to run a factory. Year 2 is where the compounding begins. Year 3 is where the 300%-type ROI becomes defensible against even the toughest CFO.
The honest caveats
Three, because honest matters more than impressive.
This roadmap works if you already have basic IT maturity. If your organisation is still stabilising its core systems, add six months at the front for data readiness. Agents can’t operate on data you don’t have.
The timelines assume competent execution. With the wrong people, every phase doubles. This is why I spend so much of my time now in fractional leadership roles, because “wrong people in the wrong phase” is the single most expensive drag on AI programmes.
The 300% outcome is upper-quartile. A well-run programme on this roadmap should expect 150–300% ROI over three years, not 300% in year one. The factory compounds; the business case must be built on the compounding, not on a single-year rocket ship.
The one-sentence version
If you want the seven-phase roadmap compressed to one sentence: frame it as a capability, map before you build, stand up platform and governance together, ship the first small cohort with discipline, treat adoption as a separate programme, scale only once steady-state is proven, and embed the result into the operating model rather than the project portfolio.
That sentence is the whole plan. The seven phases are just how you execute it.
If you want to walk through this roadmap against your organisation’s starting state, book a discovery call. Or explore the AI Factory service and the case study behind the 300% ROI number.