Why AI programmes die

They do not fail because of the technology. They die in the org chart.

Four failure modes, and the fixes everyone already agrees with. The interesting column is the second one: where the fix stops being advice and becomes something the software will not let you skip.

The four-mode diagnosis is a widely circulated industry observation, not this practice’s own. What is ours is the enforcement: where each fix is held by a gate, a refusal or a record.

Failure mode 1

Fragmented ownership

A CAIO, a CTO, a CIO and a COO all have a stake, and nobody has accountability. AI becomes a political football rather than a business capability.

Held by a gate

Every stage has named decision seats, and the release gate stays closed until the accountable seats have signed. Not a committee — a seat, with a signature that resolves to a person. A signature never papers over an open condition.

See the decision seats

Failure mode 2

Strategy follows spend

Licences get bought, pilots get launched, and months later someone finally asks what measurable problem this was supposed to solve.

Held by the record

Use cases are ranked by formula before anything is built, and the value stage separates recorded from measured. Four value lines and none measured reads as four forecasts, not four results — because if you cannot name the metric, you do not have a use case yet.

Watch the stages derive

Failure mode 3

Data blindness

Every use case hits the same quality, access and governance wall, and the people who know how to fix the data are the last ones invited to the strategy room.

Held by a gate

Data governance is an obligation with an owner, an SOP, named evidence and a test that every dataset traces to a source and a lawful basis. It is a delivery gate, not a workstream that starts after the pilot disappoints.

See the obligation and its test

Failure mode 4

Shadow execution

The most valuable AI work in the building is often a solo side project living in a spreadsheet. No sponsor, no budget, no path to scale.

Held by a refusal

The register holds every AI system including the ones nobody declared, and flags the ones with no owner or no human reviewing the output. An unregistered system is a finding, not an absence — and a production system with blocking work open is blocked.

See the register

Each fix is enforced by something that will not proceed without it — a closed gate, a refused record, or a field that prints [ TO BE COMPLETED ] rather than filling itself in.

Software cannot give a programme an accountable owner. It can refuse to move until one is named, which is the only part of this a system can honestly do.

AI maturity has very little to do with the technology. If the org chart looks familiar, the problem is the operating model underneath it — which is the thing this system is, rather than a tool that sits beside one.

→ Stage 1 · Discover

First, what you actually run — the processes, the systems and the AI already in use.

Discover