How an AI-adoption engagement actually runs — start to finish, on real data. Four moves: understand the organisation, map every task to where AI belongs, build a personal learning plan for each person, then drive adoption with prompts and activities. Grounded in a synthetic ~9,000-person build of Canary Co and O*NET-coded occupations.
The shape of the work
Most "AI adoption" stalls because it starts with the tool. This starts with the organisation — who does what, which tasks should stay human, and what each person needs to learn next. The tool comes last, once you know where it belongs.
Model the org from the ground up — divisions, units, roles, people, capability.
Decompose every role into tasks and route each to a three-tier AI model.
Target minus actual equals gap. Gaps become personal learning plans.
Prompts, activities and a 90-day roadmap that meet people in their real work.
Discovery · Workforce intelligence
You cannot adopt AI into an org you can't see. The first move builds a complete, internally consistent picture: 6 operating divisions, 17 business units, 149 distinct job titles, 43 O*NET-coded occupations, and 8,980 synthetic people — each with a seniority band, tenure, employment type, region, and a role grounded in a real occupation code.
Capability is scaled by seniority — a Senior generally exceeds a Junior against the same role competency target. That single rule is what makes the gap analysis in Phase 03 meaningful rather than random.
How the picture is built
The synthetic half. A generator creates people, IDs, the reporting hierarchy and personas — seniority, tenure, employment type, region. Same random seed in, same org out, so a demo is fully reproducible.
The O*NET half. Each role carries a real SOC code (e.g. Pharmacist 29-1051.00, Customs Broker 13-1041.06). O*NET supplies the tasks, skills and competency profile that drive everything downstream.
Task analysis · The three-tier model
A job is not one thing — it's a bundle of tasks, and they don't all belong in the same place. We decompose each occupation into its O*NET tasks and route every one to a three-tier model: stays human, human + AI, or AI-led with oversight. Across 43 roles that produced 176 mapped tasks.
None / human
Empathy, accountability, judgement, physical presence. The work that defines the role and carries the risk.
Test: would you want a person accountable if this went wrong?
Microsoft 365 Copilot
Knowledge work helped by drafting, retrieval and analysis. The human stays in the loop and owns the output.
Test: does a draft or a summary save real time here?
Copilot agent / autonomous agent
Structured, repeatable, low-judgement work — run by an agent with a human checkpoint, not a human keystroke.
Test: is this rules-based and high-volume enough to delegate?
Just over a third of tasks stay fully human. That is the headline that defuses the fear in the room: this is augmentation, and the most important work is precisely the part that doesn't move.
Logistics and telecoms carry the most rules-based volume; corporate functions lean toward Tier 2 — assist, don't replace. The mix tells you where to pilot agents and where to pilot Copilot.
Pick a role to see its O*NET tasks tiered and routed. Notice the pattern: the highest-importance tasks tend to stay human, while the highest-frequency, repeatable ones move to an agent.
The reassuring pattern
Across every role, the tasks that stay human are usually the most important ones — dispensing (4.9), counselling patients (4.7), advising clients (4.3), architecture calls (4.4). AI takes the repetitive volume. People keep the judgement. That is the sentence that gets a workforce on side.
Capability · Personal learning plans
Each role has a competency target from O*NET. Each person has a modelled actual. Target minus actual is the gap — and the largest gaps become a prioritised, individual learning plan with a concrete development action. Across the org that generated 14,031 plan lines for 5,908 people.
Junior Pharmacist · Health & Beauty · UK & Ireland Retail · Part-time, 3 yrs · SOC 29-1051.00
The radar tells the story at a glance: Frank sits below the role target on every competency, with the widest gaps on Science (4.2 → 3.0) and Critical Thinking (4.5 → 3.4). As a junior on a clinical-facing role, those gaps are high-priority — they touch the Tier 1 tasks AI will never take from him.
His plan, generated from the gaps:
The join: Frank's AI moment is the Tier 2 drug-interaction check and the Tier 3 dispensing-record agent — which free the time he reinvests in the Tier 1 clinical judgement his learning plan is built to strengthen. Phases 02 and 03 are the same coin.
Ann's repetitive ticketing moves to a Tier 3 agent — so her one plan line sharpens the exact human skill (persuasion, retention) her role now hinges on.
The same gaps, aggregated across all 14,031 plan lines, tell L&D where to invest first. The top of the list is unmistakable — and it is human skill, not tool training.
Adoption · Prompts · Activities · Roadmap
Now the tool earns its place. Adoption lands when it meets people inside the work they already do — so every prompt below is tied to a real O*NET task and its tier, and every activity is drawn from the actual development actions in Phase 03. Prompts for the Tier 2 work, agents for the Tier 3 work, learning for the Tier 1 work that stays theirs.
Tier 2Customs Broker · Classification
Human confirms the code and owns the declaration. AI does the lookup and the first-pass reasoning.
Tier 2Accountant · Reporting
Month-end management accounts. Saves 1–2 hrs/cycle; the accountant still signs the commentary.
Tier 2Software Developer · Review
Keeps the engineer making the architecture call (Tier 1) while AI accelerates the routine review.
Tier 2Pharmacist · Safety check
Surfaces risks fast; the pharmacist's clinical judgement (Tier 1) makes the call and counsels the patient.
Tier 3Customer Service · Triage
Runs unattended with a human checkpoint on escalations — frees reps for de-escalation and retention.
Tier 1Ops Manager · Stays human
Even a Tier 1 task can use AI to prepare. The accountable judgement stays with the manager.
Run real decisions in a safe room, then debrief structured against what good looks like. The single most-prescribed action across the org.
Work real incident reviews as cases. Builds the exact skill AI can't supply and the org needs most (1,528 gaps).
Hands-on, repeated practice with a more senior pair. How Dawn closes a 1.4 programming gap.
Persuasion, service orientation and active listening — the human edge once agents take the ticketing.
Personnel management and coordination for new and stretched managers like Angela.
30-minute, role-specific sessions using the library above on the person's own live work — not generic training.
Days 0–30
Days 30–60
Days 60–90
Governance is not a brake
Every Tier 3 agent ships with a human checkpoint. Every Tier 2 output has a named owner. The "what stays human" list is published, not implied. Adoption that skips this is the adoption that gets pulled after the first incident.
End to end. One coherent method from raw org structure to a 90-day adoption plan — not a deck of disconnected tactics.
Grounded in data. Every number on this page is computed from a real generated dataset of 8,980 people and O*NET occupation codes — reproducible from a seed.
Honest by design. Synthetic people, flagged placeholder content, explicit "what stays human." The method survives contact with a sceptical workforce.