Method
Four things, built in that order.
The practice came first, then the model it produced, then the platform that runs it, then the agents inside that. Each one exists because the one before it was not enough — and each still stands on its own.
01 · ADVISORY
Reading a portfolio as it actually is.
The register, the capacity position, the last few quarters of decisions. What is in the plan that would not survive a gate, and what the gates that would have caught it look like. The work starts with what exists, not with what should have existed.
- Reads
- The register as maintained, not as presented
- Reads
- The capacity position against real availability
- Reads
- The last few quarters of decisions, and their basis
- Returns
- A written account of what would not survive a gate
- Returns
- A design for the gates that would have caught it
02 · OPERATING MODEL
Two altitudes, one governed system.
One front door for demand, staged gates through delivery, and a written decision trail behind both. It runs on whatever tooling is already in place where that works — the model is the thing that matters, not what it is built on. An ask that cannot survive the front door was never a plan item; it was a preference with a sponsor.
Demand
Portfolio · once a quarterDelivery
Feature · every featureThe expensive decisions are taken while they are still cheap to change.
03 · PLATFORM
The model, built as software.
Horizon PMO is a portfolio intelligence platform for enterprise change — one live view of every project, programme and investment. It exists because the method needed somewhere to live: a register that can be compared, gates that can be failed, and a decision trail that survives the people who wrote it.
- Views
- 17 dashboards over one governed data model
- Tenancy
- Multi-tenant, with per-tenant isolation
- Identity
- Enterprise SSO and SCIM provisioning
- Access
- Row-level control enforced in the database, not the application
- Independence
- The advisory work does not depend on it
04 · AGENTS
Drafting and deciding are different jobs.
An agent runtime inside the platform. Agents read portfolio data, every step they take is recorded, and any write stops at a person. The approval is a database gate, not a convention — a run cannot complete a write without a recorded human approval, and cannot be approved past a recorded error.
- Agents do
- Read, draft, score, challenge
- Agents cannot
- Complete a write without a recorded approval
- Agents cannot
- Be approved past a recorded error
- Recorded
- Every step, every tool call, every input and output
- Reserved
- Direction, judgement, accountability
Where the gates sit
Gates fire before a line of code is written, not after.
Roughly half of them sit ahead of the build. The expensive mistakes are caught at design, where they are cheap to fix — not in production, where they are not.
Delivery
The same rule, all the way down.
Everything above decides what belongs in the plan. This is what happens to it afterwards — the same discipline, applied to building rather than choosing.
AI works inside the loop. It does not sign. The reviews at those gates are run as AI panels — several models, independently, against the same design, because one opinion repeated is not corroboration. Direction, judgement and accountability stay with a named person, dated and on the record.
THE DECISION TRAIL
What was proposed, who challenged it, what changed, and who signed. Critical designs are never signed off by their own author.
Start here
Start with the portfolio you already have.
Bring the plan as it stands. The first conversation is a read of it, not a pitch for any of the four.