Security & governance
Autonomy without control is a liability.
Every AI system we deliver includes a baseline of security and governance controls. They are designed in from the first sprint, not added at the end.
Baseline controls
Ten controls in every deployment
Least-privilege access
Each agent receives only the permissions its task requires, under its own identity.
Scoped credentials
Short-lived, revocable credentials stored in a managed vault — never in prompts or code.
Approval gates
Thresholds that send financial, legal, or customer-facing actions to a person first.
Audit logging
Every observation, decision, tool call, and outcome is recorded and reviewable.
Environment separation
Development, staging, and production are isolated, with test data outside production.
Data minimisation
Only the fields a task needs are sent to a model; sensitive data is redacted where possible.
Evaluation
Test suites built from real cases run before every release to catch regressions.
Monitoring
Quality, cost, latency, and error dashboards with alerts to an owner.
Rollback & kill switch
Any agent can be paused instantly and changes reverted to a known-good version.
Human escalation
Clear routes for uncertain or out-of-policy cases to reach the right person.
In detail
How control is layered
Identity & access
- Separate service identity per agent
- Least-privilege, operation-level permissions
- Short-lived credentials in a managed vault
- Access reviews and revocation
Action control
- Allow-listed tools and operations
- Approval thresholds for consequential actions
- Rate limits and spending caps
- Kill switch per agent
Data handling
- Only required fields sent to models
- Redaction of sensitive data where possible
- Provider settings that exclude training on your data, where offered
- Retention aligned with your policy
Assurance
- Evaluation suites before every release
- Full audit logs of decisions and actions
- Monitoring for quality, cost, and errors
- Rollback to known-good versions
Being clear about limits
No AI system is error-free.
Language models can misread inputs, produce plausible but wrong output, or behave differently after a provider update. We don't claim otherwise.
What we can do is design systems where errors are likely to be caught, their impact is limited, and every action can be traced and reversed. That is the purpose of the controls above.
FAQ
Frequently asked questions
01Are you certified under SOC 2 or ISO 27001?
We do not claim certifications we do not hold. We apply the controls described on this page and support your own security review, questionnaires, and contractual requirements.
02Can AI systems run in our own cloud?
Yes, where needed. We can deploy into your cloud environment and use model endpoints that meet your data-residency or data-handling requirements.
03Do you guarantee compliance with regulations like HIPAA or GDPR?
Compliance depends on your organisation, contracts, configuration, and processes. We design systems to support the safeguards these regulations require and work with your compliance team, but we do not make blanket compliance guarantees.
04What happens if an agent does something wrong?
It can be paused immediately, its credentials revoked, and recent changes reviewed through the audit log and reverted where possible. We then add the case to the evaluation suite so the failure is caught in future.
Next step
Want to walk through our controls with your security team?
A strategy call is a working conversation about one workflow: what it involves today, what could be automated, and what a first build would look like.
What happens next
- 1Tell us about one workflow
- 2We review it before replying
- 3A working session, then a first-build proposal