// Playbook 04 — Transformation programme setup
Setting up an AI-Native programme.
A six-week playbook for turning an AI ambition into a working programme. Frame the outcomes. Decompose into workstreams. Scope the first three. Operate on a cadence that doesn't bury the team. Worked example throughout: Halcyon Financial.
the programme
· with stop conditions
· 14 steps
goals tolerated
The brief.
Worked exampleAn AI ambition without a programme.
Halcyon Financial is the same regulated digital financial services platform from Playbook 03. The CoE is up and running. Now the executive has committed to something bigger: be AI-Native within twelve months. The slides say so. The all-staff said so. Investors heard the same thing on the last earnings call.
On the ground, the operating reality is six teams making six different bets. A CX tool pilot in CX. A vendor evaluation in compliance. A bottom-up agent experiment from one engineering pod. No shared definition of success. Vendor pitches arriving weekly. An ambition, but not yet a programme.
The new Transformation Lead's first six weeks are about building the structure that makes the ambition tractable — without adding bureaucracy, freezing experiments in flight, or pretending the answer is a Gantt chart.
with a stop condition
decomposed
· with defined scope
survive Phase 1
// And the people you'll meet
The Transformation Lead
You, the readerAccountable for nothing. Responsible for the programme glue. The role this playbook is written for.
Maya Chen
CX Officer · 2.5 yrsHer KYC-notification idea lives in the customer self-serve workstream — queued for Q2 behind the rails being proven first.
Tom Nguyen
CX Team LeadOwner of the ticket-triage POC. Tests whether ticket data is clean enough to train against.
Priya Nair
Head of CXOwner of the knowledge-base AI-assist. Runs the workstream that proves the rails first.
Sam Patel
Engineering DirectorAccountable for the highest-risk pilot — the CX tool — and for whether twenty CX officers actually adopt it.
Anna Petrović
Head of ComplianceAccountable for both compliance workstreams. Both currently deferred until regulator-tolerance lands.
Four phases. Frame, decompose, scope, operate.
The framework// AI programmes don't usually fail at the model. They fail at the framing — and the decomposition that comes next.
Frame
Weeks 1–2Convert ambition into 2–3 concrete outcomes with measurable proxies and explicit stop conditions.
Decompose
Weeks 2–3Break each outcome into workstreams. RACI per workstream. Decision rights. The smallest team that can ship.
Scope
Weeks 3–4Pick the first three workstreams. Sequence by learning, not just impact. Name partners and what they don't own.
Operate
Weeks 5–6Three cadences. Adoption and value measured separately. Decision log public.
Phase one · Frame.
Weeks 1–2The goal of the first phase is tractability. "Be AI-Native" is a slide. A programme needs two or three outcomes a sponsor can defend, a Transformation Lead can decompose, and a sceptic can falsify. Without a tractable frame, every workstream becomes a debate about what it's for.
Surface the strategic ambition.
Sit with the exec sponsor and the business unit heads. What does "AI-Native" actually mean — for this business, this year, this customer? Translate the aspiration into specific business contexts: faster resolution, lower compliance cost, higher engineering throughput. Three or four candidate territories, no more.
Define 2–3 outcomes, no more.
Each outcome ties to a measurable business proxy and a time horizon. Not "use more AI" but "reduce CX ticket resolution time by 40% in nine months". Three is the limit. Five outcomes is no programme.
Write the stop conditions upfront.
The conditions under which the programme stops. Most AI programmes don't survive contact with reality because no one defined what failure looks like before they started — so the programme can't be stopped and just gets quieter. Name the trigger for stopping each outcome.
Get a 90-day picture in writing.
Two paragraphs, signed off by the sponsor, that describe what visible progress looks like at the end of week 12 — not the end-state, the early-confidence checkpoint. The picture is what you'll compare reality to in Phase 4.
Three outcomes, three stop conditions.
The Transformation Lead spends the first ten days in 1:1s — exec sponsor, Heads of CX, Compliance, Engineering, People & Culture. The shortlist of candidate outcomes comes back at eleven. The discipline of Phase 1 is the cut to three.
- Outcome 1Reduce CX median ticket resolution time by 40% in 9 months. Proxy: median TTR per ticket category. Stop if: no measurable improvement by month six.
- Outcome 2Automate 30% of compliance review work by Q3. Proxy: hours saved per week × analyst count. Stop if: regulatory tolerance not secured by week eight.
- Outcome 3Establish an AI-literacy baseline across 80% of engineering in 6 months. Proxy: certification completion + applied-use evidence. Stop if: voluntary uptake below 40% at month three.
The 90-day picture: workstream one live for 30% of the CX team, workstreams two and three with scope locked and budget committed, sponsor on the record at the Q1 board update.
The hardest conversation in Phase 1 is the cut. Sponsors will want all five outcomes. Hold the line at three. A programme with five outcomes is five programmes pretending to be one, and none of them will get the attention they need.
Phase two · Decompose.
Weeks 2–3Decomposition is the work the playbook does that PowerPoint can't. An outcome is a destination. A workstream is something a small team can be given on Monday and report back on by Friday. This phase converts strategy into the units of delivery — and is where most AI programmes either earn their structure or lose it.
Map each outcome to 3–5 workstreams.
A workstream is a concrete deliverable a small team can ship — not a theme. "Knowledge-base AI-assist" is a workstream. "Customer experience uplift" is a theme. The test: can it have a date attached, an owner named, and a stop trigger?
Write a RACI per workstream.
Responsible, Accountable, Consulted, Informed. Names, not titles. The RACI exists to remove ambiguity at the moments where decisions slow down — not to look professional in a slide. If a workstream has more than one Accountable, the RACI is wrong.
Right-size the team — smallest that ships.
Two builders plus one decision-maker is often enough. Add a fourth only when a specific skill gap demands it. If a workstream needs eight people, the workstream is the wrong shape — go back to the decomposition.
Document decision rights and escalation.
Who can say yes. Who can say no. Who can say "not yet" — and what triggers an escalation. Written down before the first real decision, not negotiated under the pressure of one. Three lines per workstream is enough.
★ = selected for Phase 3 delivery
Seven workstreams, each with a name attached.
The CX outcome decomposes into four workstreams — three focused on officer productivity, one on customer self-serve. The compliance outcome breaks into two workstreams that both depend on a regulator-tolerance conversation that hasn't happened yet. The engineering AI-literacy outcome is one workstream, owned by People & Culture rather than the AI programme directly.
Maya Chen's KYC-notification idea from Playbook 03 lands inside the customer self-serve workstream — sequenced for Q2, not Q1. The improvement-idea pipeline from the CoE and the AI programme's roadmap reconcile cleanly because they were never trying to do the same job.
Each workstream gets a one-page RACI with the actual person named in every cell. The CX tool pilot: Sam Patel (Engineering Director) Accountable, with two ML engineers and a CX team lead Responsible. Both compliance workstreams: Anna Petrović (Head of Compliance) Accountable. The Transformation Lead — the only un-titled role on the chart — is Accountable for nothing, and Responsible for the programme glue.
RACI bloat. A RACI with twelve names on it is a list of who showed up to the meeting, not a model of how the decision happens. Aim for four names per workstream. If the consulted list is longer than the responsible list, the workstream is over-governed before it's started.
Phase three · Scope.
Weeks 3–4By Phase 3 you have a workstream map. The next discipline is to pick — three to start, no more. Sequence by what each workstream teaches you, not just by what it returns. Phase 3's purpose isn't maximum value in week one. It's maximum learning before the programme commits to its big bets.
Sequence by learning value, not just impact.
The first workstreams should answer questions the org doesn't yet have answers to: does the data hold up? Does the change land? Does the vendor deliver? A low-risk workstream proves the rails. A medium-risk one tests data quality. A high-risk one tests cultural appetite. Three workstreams, three different learnings.
Map dependencies and blockers explicitly.
Data access. Security review. Vendor onboarding. Change-management readiness. Anything that takes more than two weeks to unblock should be started in week 4, not week 8. The dependency map is the early-warning system for everything that can derail Phase 4.
Name partners — and what they don't own.
Strategy partner. Delivery partner. Training partner. Each named, each with explicit scope. The most common partner failure is implicit scope creep — the partner becomes load-bearing because no one wrote down what they didn't own. Document the boundaries before the first invoice.
Knowledge-base AI-assist.
Low risk- Duration8 weeks
- OwnerPriya Nair (Head of CX)
- PartnerInternal only
- DataExisting KB · no PII
Ticket-triage POC.
Medium risk- Duration12 weeks
- OwnerTom Nguyen (CX Team Lead)
- PartnerInternal · ML team
- DataAnonymised tickets · privacy review needed
Agent CX tool pilot.
High risk- Duration16 weeks
- OwnerSam Patel (Engineering Director)
- PartnerExternal delivery partner
- DataLive customer interactions
Three workstreams, three different questions.
The knowledge-base AI-assist starts in week 5 — low risk, internal-only, proves the technical rails. The ticket-triage POC starts in week 7 once the data privacy review (kicked off in week 4) returns. The CX tool pilot starts in week 9 — the external delivery partner is onboarded over weeks 4–8.
The delivery partner for the CX tool's scope is documented in a one-page charter: build the pilot, train the first 20 CX officers, hand off to internal team at week 14. What they don't own: ongoing operation, scaling beyond the pilot, change management across the broader CX org. The boundary is written down before the partner's first day on site.
The two compliance workstreams stay deferred until the regulator-tolerance conversation lands. Better to wait than to start without sign-off and have to roll back. The customer self-serve workstream (housing Maya's KYC-notification idea) is queued for Q2 — once the knowledge-base AI-assist has proven the rails.
Resist the urge to pick three workstreams that all teach the same thing. Three vendor-led pilots teach you about one vendor. Three internal builds teach you about engineering capacity. Spread the bets across what's unknown — the rails, the data, the culture — and you've bought yourself a real Q2 conversation.
Phase four · Operate.
Weeks 5–6Phase 4 is where the programme becomes a habit. Three cadences carry the weight: the weekly that surfaces blockers, the monthly that re-prioritises, the quarterly that re-grounds against strategy. The cadence has to be predictable enough that the team prepares for it — and tight enough that it doesn't bury them.
Stand up the three cadences.
Weekly programme stand-up — 30 min, six people, blockers and decisions only. Monthly steering — 60 min, sponsors plus workstream leads, re-prioritisation and value check. Quarterly value review — 2 hrs, broader leadership, outcome metrics against the 90-day picture. Each cadence has a different job. Mixing them is the most common failure.
Track adoption and value separately.
Adoption is whether people are using it. Value is whether outcomes are moving. Adoption is leading; value is lagging. Two metrics, two dashboards, two cadences. The most common reporting failure in AI programmes is celebrating adoption as if it were value — "200 active users on the CX tool" with no measurable change in resolution time.
Publish the decision log within 48 hours.
Same discipline as the CoE. Three lines per decision: what was decided, why, who owns the next step. Decisions, not minutes. If a forum can't be summarised this way, it didn't make decisions.
Three cadences, one dashboard each.
Weekly stand-up: Mondays, 30 min, programme lead plus three workstream leads plus the sponsor's chief of staff. Standing agenda — blockers, decisions needed, what's shipping this week. Forty minutes is too long. Forty-five is a status meeting in disguise.
Monthly steering: first Tuesday, 60 min, sponsor plus heads of CX, Compliance, Engineering. Reviews the value dashboard, re-prioritises against the three outcomes, signs off on partner changes. Quarterly value review: January, April, July, October. Two hours, the full leadership cohort, outcomes vs. the 90-day picture, stop-conditions check.
- AdoptionWeekly active users per workstream · target 60% of seat licences by week 12
- ValueMedian ticket resolution time, tracked weekly · baseline captured week 1
- Decision logPublic to all of Engineering and CX · 48h SLA
- Stop-conditions checkQuarterly · explicit go / hold / stop call per outcome
The cadence that gets dropped first is always the quarterly value review — it feels less urgent than the weekly. Defend it harder than the others. The weekly catches problems; the quarterly catches whether you're still solving the right problem.
Ambition without decomposition is a wish list.
Three cadences. Weekly, monthly, quarterly.
Operating rhythm// Once the four phases are operating, the programme runs on three rhythms. Each has a different job. Mixing them is the most common failure.
The stand-up.
- Surface blockers
- Confirm what's shipping
- Decide what needs a decision
- Escalate what needs the monthly
The steering.
- Review value dashboard
- Re-prioritise across workstreams
- Sign off partner / scope changes
- Publish decisions within 48h
The value review.
- Outcomes vs. 90-day picture
- Stop-conditions check per outcome
- Refresh the workstream map
- Re-secure sponsor endorsement
Four ways this fails.
Common pitfalls// Every AI programme that didn't survive failed in one of these four ways. Watch for them in the first 90 days — they're easier to fix early.
The pilot trap.
Endless proofs-of-concept that never scale because no one defined what "scale" looks like. Six months in, three pilots are technically successful and zero are in production. Engineering capacity quietly burns.
Every pilot has a written scale gate before kickoff. Three criteria, one of them measurable. If the gate isn't met by week 12, the pilot is pulled — not extended.
Vendor drift.
The external partner quietly expands scope, becomes load-bearing, embeds in workflows that aren't documented. Exit becomes expensive. The programme loses optionality without anyone making a decision to give it up.
Define what the partner doesn't own in writing. Quarterly partner review against the original charter. Anything the partner is doing that wasn't in scope gets a yes/no/renegotiate call, not a quiet continuation.
The metric mismatch.
Adoption metrics get celebrated as if they were value. "Two hundred users on the CX tool!" — but no measurable change in resolution time. The programme looks like it's working until the quarterly review asks the lagging question.
Adoption and value live on separate dashboards. Never combine them. The weekly tracks adoption; the quarterly tracks value. Different cadence, different question.
The ambient anxiety.
The programme is positioned as "transformation" and triggers quiet fear across the org. Workstreams stall in ways no one can name. Engagement metrics drop. The conversation everyone is having is the one the programme charter doesn't reference.
Name what the programme is — and isn't — honestly, early, often. Don't make promises the programme can't enforce. Frontline staff can spot an unenforceable commitment in a single all-staff. Better silence than a promise that won't hold.
By week six, you should have.
Starter checklist// Twelve items. If you can tick all twelve at week six, the programme is operational. Anything missing is debt — it'll surface in month three.
Using this in practice.
CloserThe discipline is the thing.
This playbook is a starting point, not a prescription. Every AI programme runs into its own version of vendor pressure, internal politics, and the seductive comfort of building another pilot instead of shipping the first one. Halcyon Financial is one shape. Yours will be different.
What travels is the discipline, not the artefacts. The workstream map will look different in your tool stack. The RACI will have your faces. The partner charter will have your vendor's logo on it. The shape of the work — frame, decompose, scope, operate — is what doesn't move.
If you're standing one of these up and want to talk through where it's getting stuck, I'm happy to.