We test machine-learning systems for bias, security holes and regulatory gaps, then hand you a plain-English report with fix priorities. No jargon walls. No 200-page filler.
Request your AI safety audit
Every AI system has different failure modes. We match the audit scope to the risk profile of your specific deployment, not a one-size checklist.
We run your model against demographic sub-groups, measure disparate impact ratios, and flag where protected characteristics correlate with outcomes. The deliverable includes a ranked list of bias vectors with suggested mitigations, typically ready in two to three weeks.
Prompt injection, data poisoning, model inversion: we simulate the attacks that matter for your architecture. Our red-team exercises follow OWASP ML Top 10 guidelines and produce a severity-graded findings report within four weeks.
The EU AI Act, UK AI Safety Institute guidance, ISO 42001 and sector-specific rules (FCA, MHRA) each demand different documentation. We map your system against the frameworks that apply and tell you exactly what is missing, what is adequate, and what needs rewriting.
Models degrade after deployment. We set up statistical monitors that track prediction quality, data distribution shifts and feature importance changes on a weekly cadence. When something slips, you get an alert with context, not just a number.
From first call to final report, most engagements wrap up in six to eight weeks. Here is what happens at each stage.
We spend 45 minutes understanding your model, its training data lineage, deployment context and the regulatory obligations you face. No charge for this call.
You share model cards, data dictionaries, architecture diagrams and access credentials for a sandboxed environment. We sign an NDA before anything moves.
Our engineers run bias probes, adversarial attacks and compliance checks against the agreed scope. This typically takes three weeks, sometimes four for large-scale deployments.
You receive a structured findings document: executive summary, detailed technical findings, risk ratings and a prioritised remediation roadmap. We walk through it together on a video call.
For 30 days after delivery, our team answers questions, reviews your fixes and re-runs specific tests at no extra cost. After that, ongoing monitoring is available as a separate retainer.
If yours is not listed, the contact form below goes straight to an engineer.
No. We work inside a sandboxed replica. If creating one is impractical, we can run tests against exported model artefacts and a representative data sample. Production credentials are never required.
Financial services, healthcare, recruitment tech and public-sector decision systems make up most of our work. We have also audited recommendation engines for e-commerce platforms and content moderation classifiers for social media firms. If your sector is not on that list, get in touch and we will tell you honestly whether we are the right fit.
A single-model bias review starts around £8,500. A full-scope audit covering bias, security and regulatory compliance for one deployment typically lands between £18,000 and £32,000 depending on model complexity and data volume. We quote a fixed fee after the scoping call, so there are no surprises.
Yes, and this is actually one of the most common requests. When source code is unavailable, we treat the model as a black box and use input-output probing techniques. The depth of findings is somewhat narrower than a white-box audit, but it still surfaces the bias and security issues that matter for compliance.
Clients on our monitoring retainer receive regulatory update briefs within 10 business days of any published amendment. For one-off audit clients, we offer a discounted re-assessment if significant legislative changes land within six months of report delivery.
Describe the system you want audited, and an engineer will reply within one working day.
Phone: +44 1306 474465
Email: [email protected]
