What we believe about AI in real operations.

Seven principles that shape every system

1. 01

   AI should not replace the person doing the work

   The team running an operation carries context that a model does not. Our systems are built to remove repetitive effort and surface information faster, so the operator has more room to make the judgement call — not less.

   We design around the operator's day, not around a demo. When a system reduces the number of tabs a person keeps open, that is the outcome we are optimising for.

2. 02

   Every substantive answer should point back to a source

   A response without a source is difficult to act on and impossible to review. Where the workflow allows, we tie outputs to the underlying document, record, or dataset so a reviewer can verify the reasoning quickly.

   Source grounding is not a UI feature. It shapes how retrieval, prompting, and review are designed from the first sketch.

3. 03

   Hallucinations are a design problem, not a model problem

   Language models will produce confident text in the absence of information. Our job is to constrain the surface area where that happens: narrow the scope, require sources, surface uncertainty, and let the operator see when the system is guessing.

   A system that admits it does not know is more useful than one that fabricates. We design the seams to make that admission visible.

4. 04

   Automation should stop before the consequential decision

   Drafting a message, classifying a case, preparing a submission, clustering feedback — these compress well. Approving a refund, releasing a batch, filing a submission — these should stay with a person who is accountable.

   We identify the boundary during discovery and design the review step around it. Automation on the wrong side of that line becomes a governance problem later.

5. 05

   Humans should intervene when confidence, sensitivity, or novelty is high

   Confidence should route work, not hide it. Low confidence should escalate. High sensitivity should escalate. Anything the system has not seen before should escalate. The default is a review, not an action.

   We define the escalation policy with the customer during design, and revisit it during pilot as real cases surface.

6. 06

   Confidence scores are a routing signal, not a verdict

   A confidence score reflects the model's certainty inside its training and inputs. It does not measure whether the answer is correct in the world. We use scores to sort and route, and we test the calibration against real outcomes.

   Where the workflow permits, we prefer visible confidence and sources over a single opaque number.

7. 07

   There are decisions AI should never make

   Anything with regulatory, safety, financial, or human consequences beyond a defined threshold. Anything involving a person's employment, credit, health, or freedom. Anything where a mistake cannot be reversed on the same working day.

   We name these decisions during scoping. The system may prepare, cite, and recommend around them. The decision stays with a human.

How these principles show up in the work