Skip to main content

APPROACH

Governed AI, described precisely.

"Human in the loop" is a reassurance. This is the actual boundary, where it sits, and what sits on each side of it.

The boundary is architectural, not editorial.

In most contract and assurance tools, a language model reads the material and expresses an assessment, and a person reviews what it produced. The model reaches the conclusion; the human checks it.

Westheath Commercial Intelligence inverts that. Governed, deterministic rules execute the test. Language models assist retrieval, classification and drafting. They are structurally prevented from asserting what a piece of evidence means — not by policy, and not by prompt, but because no path exists by which a model's output becomes a governed assertion.

Judgement begins at an act attributable to a named person, and every act after that carries their name.

The platform may

  • Extract and classify
  • Compare and cross-reference
  • Execute governed rules
  • Assemble evidence chains
  • Draft for review

Only a person may

  • Determine what evidence means
  • Approve a finding
  • Close a finding
  • Issue a report
  • Exercise commercial judgement
The boundary is maintained by architecture. No path exists by which a model output becomes a governed assertion.

Four principles, built in.

Evidence before AI

The platform is designed so that a conclusion cannot be stated without the evidence that supports it.

Principle
Evidence before AI
In practice
Every generated statement carries the artefacts it rests on
What it prevents
A confident conclusion with nothing behind it

Rules before AI

Governed, deterministic rules are the execution engine. The same inputs produce the same result every time, and the run is replayable. Language models are additive — they assist retrieval, classification and drafting.

Principle
Rules before AI
In practice
Governed rules run first; language models assist retrieval and drafting
What it prevents
An inconsistent result that depends on model mood

Judgement stays human

The platform may extract, classify, compare, cross-reference and draft. It may not approve a finding, close a finding, or issue a report. Those acts carry a name, and that name must be a person's.

Principle
Judgement stays human
In practice
No approval, closure or report issuance without a named person
What it prevents
A finding that nobody can be held to

One source of truth

Evidence exists once, content-addressed. Every output is generated from it, so no two documents can quietly disagree about what the evidence said. Changing evidence creates a new record, not a replacement.

Principle
One source of truth
In practice
Append-only, content-addressed evidence; no silent edits
What it prevents
Two documents that contradict each other about what the evidence said

Confidence is a state, not an adjective.

A gate that is never withheld is not a gate.

Five points at which only a named person may decide. No gate may be bypassed, delegated to the platform, or approved in bulk.

The five decision points and the platform boundaries around them are set out in the commercial assurance governance model.

Local by default. Cloud by choice.

External AI services are not a dependency of the design.

For programmes handling commercially sensitive cost data under public scrutiny, that is a question procurement asks early, and it is better answered by architecture than by a data-processing addendum.

Read the security position

Evidence before AI. Governance before automation. Judgement stays human.

Request a briefing

A working conversation about how any of these applies to your programme.