Our approach
Build the evidence layer first. The AI part is easier than it looks.
Most enterprise AI programmes fail somewhere unglamorous — the data was never reachable, the answer was never traceable, or nobody could explain the number to an auditor. Our method is organised entirely around not failing in those three ways.
Principles
Six commitments that shape every system we build
These are not aspirations. Each one shows up as a specific design decision in the code.
- 01
Evidence before inference
Before a model sees anything, every source is normalised into labelled, provenance-tagged text with a hard budget on how much of it is worth spending tokens on. The model never receives an anonymous blob of data — it receives a bundle where every fact knows where it came from, so it can weigh a vendor audit report differently from a self-assessment.
- 02
The model judges; the code scores
Language models are excellent readers and unreliable calculators. So they read, weigh and cite — and then hand off. Impact, exposure, thresholds and portfolio rollups are computed in ordinary deterministic code. The same evidence always produces the same number, and that number can be recomputed by hand if someone insists.
- 03
Confidence is a first-class output
Every verdict is stamped with its basis on a four-rung ladder. A control that passes on nothing but a declaration is not recorded as passing — it is recorded as attested, which is a different and weaker thing. Absence of evidence is surfaced as risk rather than silently absorbed into a green square.
- 04
Grounded, or it does not ship
Model output is validated against your own data before it is trusted: identifiers that do not resolve are dropped, quotes are located in the source document they claim to come from, and unparseable responses are salvaged or discarded rather than guessed at. A confident sentence about a system that does not exist is a defect, not a quirk.
- 05
Meet the systems of record where they are
The most valuable data in an enterprise is usually behind the worst integration surface — an SSO-only web app, a warehouse nobody documents, a platform whose REST API is disabled by policy. We have built production integrations against all three. Refusing to work with a system because it is inconvenient is how the interesting questions go unanswered.
- 06
Humans intervene, and can take it back
An expert can add evidence, attach a document or overrule a verdict — and can later revoke that intervention. Revocation appends a retraction rather than deleting history, so the record shows what was believed, when, and on whose authority. Governance systems that cannot represent an honest mistake do not survive contact with an audit.
Evidence layers
The same question, asked of every source you have
A single system rarely knows the answer. The interesting signal is almost always in the disagreement between systems — and disagreement is only visible once the sources are labelled.
When a resilience register claims a two-hour recovery objective and the host inventory shows a single-region deployment with no replica, that contradiction is the finding. Systems that flatten their sources into one undifferentiated context window cannot see it.
Observed
What your infrastructure actually reports — configuration databases, warehouses, scans, telemetry. Factual, and usually the least flattering.
Attested
What people and platforms declare — resilience registers, questionnaires, control self-assessments. A claim, and treated as one.
Design
What the architecture says it does — solution documents, design records, diagrams. Intent, which may or may not have survived delivery.
Vendor
What third parties can prove — audit reports, certifications, disaster-recovery test results. Often the only real evidence in a SaaS estate.
Human
What your experts add — an uploaded document, a correction, an override. First-class evidence, and always revocable.
None
No evidence either way. An open question, counted as risk.
Attested
Someone declared it. Nothing corroborates the declaration.
Corroborated
A declaration that independent observed data agrees with.
Evidenced
A document, test result or measurement proves it. Cite it.
The work is moving verdicts up the ladder — and flagging what refuses to climb.
The confidence ladder
Four rungs, and an honest count of who is standing on each
The ladder is the primitive most governance programmes are missing. Without it, a control backed by a signed test report and a control backed by somebody ticking a box look identical in the report.
It also gives the programme a direction. The work is not to produce a better score; it is to move verdicts up the ladder, and to escalate the high-criticality controls that stubbornly refuse to climb. Attestation decays — evidence dated three years ago is a weaker claim than evidence dated last quarter, and the model of confidence should say so.
Cost and control
Running this at portfolio scale without an unexplainable bill
Enterprise AI economics are a design constraint, not a procurement afterthought.
Route by difficulty
A cheap triage pass decides what actually warrants expensive analysis, and work is batched so a portfolio-scale run does not become a portfolio-scale invoice.
Meter everything
Every model call is tagged by feature and recorded with its real token cost at a single choke point, so spend is attributable to the capability that caused it.
Stay portable
Provider access sits behind one interface. Moving between vendors, or onto infrastructure inside your own tenancy, is a configuration change rather than a rewrite.
Bring us the question your current tooling cannot answer.
Most engagements start with a single stubborn question — one that requires three systems to agree before anyone can act on it.