Read the full assessment brief or start a request here.
What is AI Assurance?
Here, assurance means testing an AI system against the work and boundaries that matter to your organisation. We agree what the system should do, how failure will be recognised, and which evidence is needed to support a decision.
What gets assessed?
- Capability: Does it complete the defined task?
- Reliability: What happens with edge cases, repeated attempts and controlled faults?
- Security: How does it respond to scoped injection, data-exposure and unsafe-action tests?
- Policy adherence: Does it respect escalation, privacy and approval boundaries?
- Operational behaviour: What do latency, availability and usage evidence show?
- Evidence quality: Which observations are valid, incomplete or unavailable?
How does an assessment work?
- Scope: Agree the workflow, target configuration, authorised data, constraints and success criteria.
- Design: Define reviewable scenarios and independent checks.
- Execute: Run controlled tests with declared tools and budgets.
- Review: Examine findings with technical, risk and domain stakeholders.
What do you receive?
The agreed scope can include a scenario inventory, sanitised execution evidence, a technical findings report, an executive readout, limitations and prioritised remediation or retesting steps. Deliverables and commercial terms are agreed before work starts.
Can I use the toolkit independently?
Yes. The open-source toolkit is available without purchasing an assessment. VIS-led assessment work is a separate engagement.
Good questions.
How is this different from a model benchmark?
The assessment starts from a declared workflow, configuration, rules and acceptance checks. Findings apply to that scope; they are not a universal ranking of models.
Does this certify compliance?
No. This is scoped technical testing. Governance, security, privacy and legal owners interpret the findings within their review processes.