How confident are you in the AI you already run?
When an AI application is in use, the important questions become practical. How is its quality tested? What can it access? How are mistakes detected? Can the team explain what changed and why?
Our Evidence Review examines the system as it exists and identifies gaps between the controls you need and those you can demonstrate.
[ + ]WHAT WE EXAMINE
- /01
The system's purpose, users and operational risks.
- /02
Current quality tests, benchmarks and test data.
- /03
Access controls, human approval and logging, as relevant.
- /04
Monitoring, incident handling and change management.
- /05
Applicable assurance requirements identified with the client's legal and compliance teams.
[ + ]WHAT YOU RECEIVE
A prioritised findings report with practical recommendations, dependencies, suggested remediation work and questions requiring business or legal decisions.
- FOR
- ORGANISATIONS ALREADY BUILDING OR RUNNING AI.
- WE EXAMINE
- PURPOSE AND RISKS, QUALITY TESTS, ACCESS AND LOGGING, MONITORING AND CHANGE MANAGEMENT.
- YOU RECEIVE
- A PRIORITISED FINDINGS REPORT WITH PRACTICAL RECOMMENDATIONS AND SUGGESTED REMEDIATION.
- SCOPE AND FEE
- AGREED WITH YOU BEFORE THE WORK STARTS.
| GAP | EXPOSURE | FINDING | CLOSING IT |
|---|---|---|---|
| 01 | HIGH | Accuracy is not measured in production. The only evidence is the launch test, run once. | Eval suite in CI plus live sampling. |
| 02 | HIGH | Logs record the answer but not the retrieved sources, so no decision can be reconstructed. | Add retrieval logging. |
| 03 | MED | Human oversight exists on paper. The override is used by nobody and tested by nobody. | Define and test an override path. |
| 04 | MED | No record of the data the model was evaluated on. | Build a versioned test set. |
| 05 | LOW | Technical documentation is current, but has no owner. | Name an owner. One meeting. |
An excerpt from a findings report. Invented system, invented findings. Not a legal assessment.
[ + ]WHO SHOULD USE IT
This is for organisations already building or running AI and seeking a more defensible way to evaluate it. If you haven't chosen a use case yet, our discovery service is a better starting point.
US
How confident are you in the AI you already run?
Tell us about the system, who uses it and what it can access. We will tell you whether a review is the right next step.