Blueprint and business case
An agentic core with a human gate on every state change
A thirty-year financial house wanted autonomous AI in a regulated business. I designed the agents to propose and people to approve, then priced it against a payback date instead of a benefit narrative.
- Modelled payback on the first-year business case
- 8 monthsModelled payback on the first-year business case
- Each mapped to a specific technical control, not a policy line
- 6 frameworksEach mapped to a specific technical control, not a policy line
- Agents propose; nothing changes state without approval
- Human gateAgents propose; nothing changes state without approval
The client
A thirty-year-old financial house in a smaller emerging market, expanding into US and EU jurisdictions, with most of its clients non-resident — and therefore facing three regulatory regimes at once rather than one.
The engagement
A transformation blueprint with a costed proposal: six-week discovery, twelve-week MVP, then scale — sold on a capital cost, an annual running cost and a payback date.
The problem
The client wanted agentic AI inside a regulated financial business operating across three jurisdictions simultaneously. Two failure modes were already visible. The first is autonomy: an agent that can act rather than propose is a compliance incident waiting for a date. The second is the business case — AI programmes in financial services are routinely sold on a benefit narrative, which survives exactly until the first budget review.
What I did
I made the human gate an architectural property rather than a policy: agents propose, a person approves, and nothing mutates state without that approval, with the workflow state durably checkpointed so an approval can be paused and resumed rather than lost. Each of the six applicable regulatory frameworks was translated into a concrete control — immutable write-once storage for record retention, private network paths with no public egress, per-jurisdiction key management — so compliance shows up as architecture decisions a reviewer can inspect, not as a chapter nobody reads. Then I costed it properly: capital cost, annual running cost including the licence lines everyone forgets, and a benefit model built on analyst productivity and avoided regulatory exposure, producing a payback date rather than a percentage. Phasing followed the same discipline — six weeks of discovery, a twelve-week MVP proving the supervisor pattern with three sub-agents under human gates, then scale.
What was built
A supervisor agent orchestrating specialist sub-agents for onboarding checks, investment scoring and compliance, with durable checkpointed state and a human approval gate before any state change; a headless content platform as the retrieval backbone so the knowledge the agents reason over is current within seconds; and compliance-led engineering that maps each regulatory framework to a specific technical control rather than to a policy paragraph.
On the table at the end
- Enterprise transformation blueprint, around forty pages
- Costed proposal with capital and operating cost and a payback model
- Digital core and discovery decks
- Control mapping from six regulatory frameworks to concrete architecture decisions
What it changed
Turned an open-ended 'we want AI' ambition into a phased programme with a stated capital cost, a stated annual run cost and a first-year payback inside eight months — and set the autonomy boundary before the first line of code rather than after the first incident.
How it ran
- 01
Set the autonomy boundary first
Agents propose, humans approve, state changes only after approval — decided before architecture, because retrofitting a gate is a rewrite.
- 02
Make the knowledge layer current
A headless content platform as the retrieval backbone with native vector support, so the agents reason over content that is seconds old rather than a stale index.
- 03
Translate regulation into controls
Six frameworks, each mapped to specific storage, network and key-management decisions a reviewer can point at.
- 04
Cost it like an investment
Capital cost, annual run cost including licences, benefit model from analyst hours and avoided exposure — expressed as a payback date.
- 05
Phase against proof
Discovery, then an MVP that proves the supervisor-and-gate pattern on three sub-agents, then scale — each phase buying the right to fund the next.
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.
- Investment research
Seventeen architecture decisions before the first sprint
An AI research platform for analysts, where the hard parts were not the models. I wrote the decisions down — including prompt-injection defence and per-subscription isolation of the vector store — before anyone started building.
- Corporate banking
Winning back the last call after two meetings had failed
A long-standing account was one bad meeting from closing. I rebuilt the pitch around the client's own operating numbers and a working prototype, and cut every claim we could not source.
Something similar on your plate?
Thirty minutes, no deck. I will tell you whether it is worth doing at all.