Programme charter
Eight AI ideas, one scorecard, two of them funded
A generative-AI product team had eight promising research directions and no way to choose. I built a gated funding model where each idea carries a hypothesis, a target metric and a benchmark, and only two reach a proof of concept.
- Candidate ideas reduced to funded proofs of concept
- 8 → 2Candidate ideas reduced to funded proofs of concept
- Each carries a hypothesis, a target and a published benchmark to beat
- Metric per ideaEach carries a hypothesis, a target and a published benchmark to beat
- Failing the gate ends the idea rather than escalating it
- Stop is an outcomeFailing the gate ends the idea rather than escalating it
The client
A consumer generative-AI product company working on image and video generation — model fine-tuning, face and identity work, lip sync, super-resolution — with a research team generating more good ideas than the budget could carry.
The engagement
A programme charter across three parallel value streams, with a staged gate model: a funded assessment, then a scorecard, then two proofs of concept at a fixed sum each, then a single MVP — or a stop.
The problem
A research-driven AI team can always name eight things worth trying, each defensible, each expensive. Without a gate, the portfolio is decided by whoever argues best in the room, effort spreads across everything, and nothing reaches production. Meanwhile two unglamorous streams — a cloud architecture review with open findings and a commercial discount the company had lost — sat outside anyone's plan because they were not exciting.
What I did
I made each idea state its own falsification condition. Every candidate in the backlog carries a hypothesis expressed as a number, the published benchmark it has to beat, an estimated duration and cost, and a named owner — so the scorecard, not the meeting, chooses the top two. The funding ladder then works in one direction only: a paid assessment produces the scorecard, the top two get a fixed proof-of-concept budget each, and only a proof that meets its declared metric earns the MVP budget. Anything else pivots or stops, which is written into the charter as a legitimate outcome rather than as a failure. I also insisted the two boring streams sit in the same charter as the exciting one, with the same milestone discipline, because the architecture remediation was a real risk and the lost discount was real money — and neither would have survived in a document about generative AI.
What was built
A single charter covering three value streams at once — generative feature development on a gated funding ladder, cloud architecture remediation against a formal well-architected review, and recovery of a reseller discount through the distribution chain — each with named milestone owners, and the charter written to feed a tracking dashboard directly rather than to sit in a folder.
On the table at the end
- Programme charter and tracking blueprint
- Scored backlog of eight candidate ideas with hypothesis, target metric, benchmark, duration, cost and owner
- Gate model from assessment through two proofs of concept to MVP or stop
- Tracking dashboard fed directly from the charter
- Remediation plan and milestone owners for the cloud review
What it changed
Replaced 'which model shall we try next' with a funding gate. Eight candidate ideas were reduced to two funded proofs of concept against declared metrics, with an explicit stop or pivot outcome designed in — so the programme could fail cheaply instead of expensively.
How it ran
- 01
Make every idea falsifiable
Hypothesis as a number, a public benchmark to beat, duration, cost and an owner — so an idea can lose on evidence rather than on advocacy.
- 02
Fund in one direction
Assessment, then scorecard, then two fixed-budget proofs of concept, then a single MVP — each gate paid for by the result of the previous one.
- 03
Write stop into the plan
Pivot or stop declared as an acceptable outcome up front, which is the only way it ever actually happens.
- 04
Keep the unglamorous streams visible
Cloud remediation and a lost commercial discount carried in the same charter with the same milestone owners as the AI work.
- 05
Charter feeds the dashboard
The document written as a data source for tracking rather than as prose, so status came from the plan instead of from a weekly reconstruction.
Other work
All case studies →- Insurance
Sequencing the data foundation before anyone bought an AI agent
A financial group wanted AI scoring across five affiliates. I proposed one affiliate, eighteen weeks, and a data readiness report before a single model — because the alternative is an agent trained on data nobody has reconciled.
- Private banking
A validated proof of concept in ninety hours
An Asian bank wanted to know whether an AI copilot would help its relationship managers. The first iteration took nineteen hours; the whole validated proof of concept took ninety and scored four and a half out of five on acceptance.
- Manufacturing
Three AI pilots aimed at one line of the balance sheet
A building-products manufacturer had capital trapped in inventory and a digital team already delivering wins. I proposed three pilots tied to working capital rather than to a technology roadmap.
Something similar on your plate?
Thirty minutes, no deck. I will tell you whether it is worth doing at all.