AI Transformation Office — Sponsor + Core Champions + Extended + external partner

An org-chart blueprint for the dedicated team that runs AI transformation across a small-to-enterprise company. Four roles: (1) Sponsor — an Executive VP / C-level who personally uses AI daily, owns the program, attends weekly coordination, sets the "AI-native company" intent; (2) AI Lead — the head of AI who builds infra and evangelises; (3) Operations Director — runs the weekly cadence, owns the bonus / incentive program, manages the academy curriculum; (4) External Partner — a small embedded firm or two-three contractors who own the methodology, run diagnostic interviews, build the initiative map, and provide muscle through pilot weeks 1–4. Around them: 7 Core Champions (3-5 leaders extracted from regular duties for ~50% of their time) + 12 Extended Champions (1 per major team, evangelising and building in their function) + a "patient zero" — someone who started the wave and may or may not still be there.

Plus an internal model-router engineer role (in-house ML engineer upskilled as the "which model when" expert — Opus 4.6 for hard tasks, Haiku for cheap ones), and three documented cadence variants with different time/intensity profiles: a 30-min × 3x/week voluntary standup; a 2-hour × 1x/week call with 60/30/10 cohort split; and a 2-hour × 2x/week mandatory share-out. Plus a peer-routing counter-pattern that explicitly replaces formal Champions for solo-mandate operators: instead of identifying and training Champions, 1:1 every employee then route them sideways to peers who are already nailing it. Both patterns coexist in the playbook — choose by operator profile.

Documented at one large fintech: 21 diagnostic interviews → 164 initiatives mapped (worth ~1–2 bn local currency/year aggregate) → 17 prioritised top initiatives → 4-week launch plan → AI Office formed with named owners. After 1 quarter: 87% of IT using AI daily, +50% YoY engineering productivity, 50+ Claude Code power users. The transformation does not happen at the company level — it happens at the team level, with a dedicated coordination layer. Without the Office, AI usage stays scattered and pilots stall in "PoC limbo".

The cadence variants matter because operator profiles differ. A mid-size IT firm leader running a champions-led adoption ($2K spent on Codex tokens in 3 weeks, model-router engineer in-house) prefers a low-overhead voluntary standup. A small franchise founder pushing a 2-month transformation prefers a high-touch 2-hour weekly mandatory call with explicit cohort tracking (after ~2 months: 60% AI-optimists, 30% observers, 10% holdouts — the "60/30/10 split" became a useful adoption-curve diagnostic). The peer-routing counter-pattern is what a solo-mandate operator at a mid-size company uses instead of a Champions Network — and it converted skeptics to adopters in ~10 days at the reference site without ever naming a Champion. Subscription-tier self-upgrade ($20 → $200/mo paid out of personal pocket) is a leading indicator nobody games — a small fraction of employees self-upgraded at one reference site after 8 weeks of the 2-hr/wk cadence.

1. Find your Sponsor first. A C-level / Executive VP who already uses Cursor or Claude Code personally every day. If you don't have one, do not start. The whole structure depends on the Sponsor's weekly attendance + personal vouching ("we are an AI-native company"). At one reference deployment, the Sponsor was a self-built personal-OS user with an "AI OS on a VPS" — credibility came from personal practice, not slide deck.
2. Name an AI Lead. Someone who can both ship and evangelise. The Lead should be 70% builder, 30% communicator. They own infra: the centralised LLM access, the policy framework, the corporate Claude/Cursor subscription, the data layer, the eval suite.
3. Name an Operations Director. Often the head of operations / chief of staff. Owns: weekly coordination cadence, the bonus / incentive program, the academy curriculum, the workshops, the executive scoreboard. Not technical — but understands compounding rituals.
4. Designate a model-router engineer. Upskill an in-house ML engineer (or hire one) to be the go-to person on "which model when" — Opus 4.6 for hard tasks; Haiku for high-volume cheap tasks; deep-research models for analyst work; reasoning-effort tier per task class. This role moves from "Sponsor's note in the Telegram channel" to a shared knowledge base over the first quarter. Without it, every team makes the model-choice decision from scratch and the org spends 2–3× more than needed.
5. Engage an external partner (optional but recommended). A small firm or 2–3 contractors who own the diagnostic methodology and shadow the program through the first 4–8 weeks. Reference example: a small embedded transformation firm — three named partners embedded in week-1 pilots, owning the dashboard and the diagnostic. Implication for a 30–100 person startup: a small external team can compress a 3–6 month transformation into 4 weeks of named milestones, then exit.
6. Identify Core Champions (7). People who already build with AI in their day job; willing to be extracted ~50% to a dedicated AI Office team. Each Core Champion brings their own validated case: a release-notes agent (2-3 hrs → 10 min), a vibe-coded merchant-cabinet agent, a speech-analytics MVP, a multi-agent dev pipeline, etc. Their "before/after" is the proof.
7. Identify Extended Champions (12, 1 per team). Not extracted from their team — they keep their role but are the explicit AI practitioner inside each function: Engineering, Frontend, Backend, Mobile, QA, Data, Marketing (per-region as needed), Sales, CS, HR, InfoSec. Their job: bring real tasks to weekly workshops, leave with working workflows, propagate to their team.
8. (Counter-pattern — for solo-mandate operators.) Skip Champion identification entirely. Instead: 1:1 every employee, watch for who is nailing it spontaneously, then route others sideways: "the eCom team — the folks there are doing it really well; go to them; they'll teach you." Peer-to-peer cascade. Founder doesn't pull people up; they send them sideways. Reference operator's outcome at a mid-size retailer: from "they spun a finger at my temple" to "people support me and run with me" measured in days, no Champion ever appointed. Choose this counter-pattern when (a) you have unilateral mandate, (b) you have enough founder time for the 1:1 rounds, and (c) under ~50 people total.
9. Acknowledge the patient zero. Often a former CTO / founding engineer who started using AI, handed out Claude Code licences to friends, and may have departed. Don't pretend they didn't exist — they shaped the early culture.
10. Run a 4-week three-stream launch plan. People / Pilots / Education in parallel.

• Week 1: People — AI hiring plan, AI Office formed, bonus programme draft. Pilots — 2-3 live pilots kicked off (typically: a QA automation pilot, a sales copilot MVP, a churn prediction MVP). Education — Telegram channel launch with Sponsor's first vision post; AI Management OS install; Workshop #1 (Engineering — Claude Code + CLAUDE.md hands-on).
• Week 2: People — Champions network rollout (1 per team, bi-weekly challenges). Pilots — corporate VPN deployed (the headline quick win that unblocks dozens of engineers). Education — AI governance draft (privacy, regulated-data rules).
• Week 3: People — AI metrics enter Performance Review. Pilots — 100% AI adoption push in engineering (Claude Code Pro for all devs, CLAUDE.md in every repo, target 80%+ adoption). Education — Workshops #2 Sales, #3 Support, #4 Product/HR.
• Week 4: Milestones — AI Office staffed, motivation programme live, vacancies in progress, decision on scaling pilots to Phase 2, AI governance approved, adoption baseline measured.

11. Run the bonus / incentive program from Week 1. Three tiers (Adoption / Champion / Transformation Protection). For solo-mandate operators using the peer-routing counter-pattern, see the no-metrics counter-stance — bonus programs are not the only way.
12. Pick a cadence variant. Three documented options. Don't run all three; pick the one that fits the org and the operator. Switch later if needed.

• (a) 30-min × 3x/week voluntary standup — a bare-minimum cadence. Used by an operator at a mid-size retailer who runs everything else in 1:1s. Cost: 1.5 hrs/wk per attendee. Drives most adoption via the 1:1s; the standup is the show-and-tell venue. Best when peer-routing counter-pattern is the operating model.
• (b) 2-hour × 1x/week with 60/30/10 split — a small franchise founder's pattern: 30 mins theory from the founder on something they're personally learning; ~45 mins practical exercises; ~45 mins case-sharing from team members. After ~8 weeks of this, a 60/30/10 split crystallises (60% AI-optimists, 30% observers, 10% holdouts) — accept the split; don't try to force the 10%. Cost: 2 hrs/wk per attendee + founder driving content. Best for small orgs with one charismatic founder.
• (c) 2-hour × 2x/week mandatory share-out — the same franchise founder's earlier variant before tightening to 1x/week. Faster initial flywheel; higher attrition risk. Used during the initial 6-week "decide who can run with us" phase. Cost: 4 hrs/wk per attendee.

14. Track subscription-tier self-upgrades as adoption signal. Provide everyone a company-paid $20/mo seat. Watch who upgrades to $200/mo out of their own pocket — at a small reference site, a small fraction of employees self-upgraded after ~8 weeks. These are the deepest converts; protect them. The signal is hard to game because it costs the employee real money. Operator policy: "if anyone wants to grow above $20, come to me; we discuss how to optimise."

Org chart template (paste into Docs/Strategy/AI-Office.md and adapt):

• Don't start without a personally-practising Sponsor. This is the #1 failure mode. A Sponsor who can't use Claude Code lacks credibility; the Office becomes a comms function and stalls.
• Don't centralise the work; centralise the cadence. Champions stay in their teams. The Office coordinates; it doesn't build for the teams. If the Office becomes the bottleneck, you've inverted the model.
• Don't skip the bonus programme (in standard variant). Adoption stays at 30–40% without explicit incentive alignment. The Transformation Protection clause is the single most-mentioned culture signal in the reference enterprise's adoption story. (Counter-stance for solo-mandate operators exists — see the no-metrics case — but applies only when mandate replaces stakeholder justification.)
• Don't make the AI Office an empire-building target. Cap headcount of extracted Core Champions at ~7. If it grows, you're solving the wrong problem.
• Don't omit "patient zero". The legacy story matters — people remember who handed them the first Claude Code licence. Acknowledge it in the org map.
• Don't use pejorative names for the AI. Internal language matters. One operator's framing: "don't call the agent a pejorative — respect it, it's not stupid; learn to use it." Pejorative team vocabulary tracks adoption rate inversely.
• Two-cohort blind spot at enterprise scale. At a large org, only a fraction will be real power-users (Claude Code etc.); the rest will get AI baked into their existing tools (CRM copilots, etc.). Don't try to make the rest into power-users; design for two cohorts. The model-router engineer's KB serves the power-user cohort; the existing-tool integrations serve the rest.
• 2026-Q1 lesson: the first wave of generic AI training failed — "40% tried it, 15% used it". Cure: weekly per-function workshops where teams bring real tasks and leave with working workflows. Don't run generic AI literacy training and expect adoption.
• Cadence pitfalls. Variant (a) 30-min × 3x/wk fails if it's the only adoption surface; needs peer-routing 1:1s in parallel. Variant (b) 2hr × 1x/wk fails if the founder stops driving content before week 5; volunteers don't emerge until week ~6. Variant (c) 2hr × 2x/wk burns out within 12 weeks; treat as initial-phase only.