The Problem
The leadership team had approved 15 different AI pilots over 18 months. None had reached production. Each pilot had its own tech stack, its own vendor relationship, and its own definition of success. The CTO described it as "innovation theatre — lots of activity, zero business impact."
The core issue wasn't technical capability. It was the absence of a decision framework for what to build, how to govern it, and when to cut a failing experiment.
The Approach
We ran a structured 8-week engagement split into three phases:
Phase 1: Landscape Audit (Week 1-2)
- Mapped all 15 pilots against business capability, cost, technical risk, and strategic alignment
- Interviewed 12 stakeholders across product, engineering, compliance, and executive team
- Identified 3 root causes: no AI operating model, no vendor evaluation criteria, no production readiness standard
Phase 2: Framework Design (Week 3-5)
- Designed an AI operating model tailored to FinTech regulatory context, working through the four things any AI operating model has to settle: who owns each initiative, where decision rights sit between engineering, product, and compliance, what risk gates apply before something reaches production, and an honest inventory of what the team could actually build against what the roadmap assumed
- Created a pilot evaluation scorecard (strategic fit + technical feasibility + compliance burden + ROI horizon)
- Defined production readiness gates with clear ownership per stage
- Built an AI decision rights matrix clarifying who decides what (CTO vs Head of Product vs Compliance)
Phase 3: Prioritisation & Roadmap (Week 6-8)
- Scored all 15 pilots against the new framework
- 6 initiatives passed the threshold; 4 were killed; 5 were deferred with explicit re-evaluation triggers
- Built a phased delivery roadmap with owners, milestones, and go/no-go decision points
- Established a monthly AI governance review cadence
The Outcome
| Metric | Before | After |
|---|---|---|
| Active AI initiatives | 15 (none in production) | 6 (2 in production within 12 weeks) |
| Decision velocity | Months per pilot decision | Days per initiative review |
| Compliance confidence | Low (no framework) | High (auditable decision log per initiative) |
| Team clarity | Diffuse ownership | Clear owners per initiative |
Within 12 weeks of engagement end, the first 2 initiatives were in production. The client reported that the decision rights matrix alone removed 80% of the cross-team friction that had blocked earlier progress.
What Changed
Before the engagement, AI decisions were made in hallway conversations. After, they were made in a structured governance forum with clear criteria, documented rationale, and explicit owners. The CEO described it as "the difference between hoping AI works and knowing why it will."
Where This Pattern Shows Up Elsewhere
Nothing about this engagement was FinTech-specific once you strip out the regulatory detail. The AI Operating Model built here is the same four axes — ownership, decision rights, risk gates, capability inventory — applied against a different regulatory context than a healthtech or manufacturing team would use. What repeats is the underlying failure: a pile of pilots that each work in isolation and never compound into capability, because nobody decided who is allowed to say a pilot is done.
The decision rights matrix is the more general artefact of the two. AI Decision Rights: Who Decides What When No Human Is Watching covers the same question this FinTech had to answer — CTO, Head of Product, or Compliance — in a form that applies wherever a model is now making a call that used to require a person to sign it: a claims decision in healthcare, a quality flag on a manufacturing line, a price change in retail. The domain changes. The question of who owns the call, and what happens when nobody assigned it, does not.
The harder version of this problem is still ahead of most of the initiatives that made it through this engagement's gates. A pilot with a person reviewing every output can survive a missing decision rights matrix longer than it should. A pilot where the AI is now taking actions on its own cannot. Your AI Agents Are Making Architecture Decisions Nobody Assigned Them That Role is the next version of the same governance gap this engagement closed — worth reading before the next 6 initiatives start shipping agentic features rather than model outputs a human still checks.