Responsible AI

Define what your AI is allowed to do.

Test AI on your cases, control its access, and decide which actions need approval. We build and document these controls with your team so you can review results and respond to failures.

What we deliver

  • Evaluation cases and results

    Tests drawn from your use case, agreed acceptance criteria, and a report of results and failures to address before release.

  • Access and action controls

    Defined data sources, permitted actions, and permission checks in the application, with input and output validation suited to the task.

  • Approval and escalation rules

    Decisions about which actions need review, who reviews them, and when missing information or a failed check must stop an action.

  • Records for investigation

    Logs and operating procedures with agreed coverage, access, and retention, so your team has records to investigate outputs and actions.

When teams come to us

  • A pilot is heading for review

    Your security team needs to see what was tested, what failed, and which controls are in place.

  • An agent is ready to act

    It can update records or call tools, but the permissions and approval rules are not yet clear.

  • An unexpected result needs an explanation

    Your team needs the records to trace what happened and decide what to change.

How we decide which actions need review

Agree what needs approval

Before launch, we agree with your team which actions require approval and who reviews them. Corrections from that review help us improve instructions and tests, including cases the first version missed.

Decide where autonomy is appropriate

We assess autonomy for each type of action using evaluation results, the consequences of an error, and your review requirements. Fewer corrections alone do not justify removing oversight. We define monitoring, escalation, and the conditions for returning an action to human review.

Illustrative example Permissions for an order draft
  1. Source

    Read the approved catalogue

    Use the part descriptions and availability shared for this task.

  2. Permitted action

    Prepare a draft

    Use that information to prepare an order for the team to review.

  3. Approval required

    Confirm a shipment

    Require an authorised person’s approval before this action can run.

If permission is missing

Stop the action and refer it to the agreed reviewer.

This is a policy example for one task. Tests also need to cover denied requests and attempts to skip approval.

Our own workflow

Our own hiring

We use an agent to turn interview transcripts into candidate profiles. Corrections during development helped us refine its instructions and handle cases the first version missed.

How we build it

  1. Risk map

    Identify possible failures, who they affect, and the conditions your team needs to address before release.

  2. Control design

    Agree permissions, approval requirements, escalation paths, and the records to retain.

  3. Evaluation before launch

    Test representative cases and failures against the agreed criteria. Document unresolved issues.

  4. Oversight in production

    Define who monitors results, responds to incidents, and can pause an action or return it to review.

Failures to plan for

We assess these cases and define how the system and your team should respond:

  • Incorrect answer

    An answer contains false or unsupported information. Test with representative questions, check the cited sources, and define when to withhold an answer or ask a person.

  • Unauthorised access or action

    The system attempts to read restricted data or perform an unapproved action. Enforce permissions when accessing data and executing tools, and test denied requests.

  • Missing records

    The team cannot reconstruct an unexpected result. Define the inputs, sources, model versions, and approvals needed for investigation, along with limits on what is stored.

  • Required approval skipped

    An action runs without the review it requires. Enforce approval before execution and define how to pause the action, notify its owner, and review the incident.

These controls help reduce risk and detect failures. Evaluation and monitoring show where further work is needed.

Works well with

We scope these controls with the product, its data, and the people who will operate it.

Frequently Asked Questions

Does this certify compliance?
No. We deliver the technical controls, evaluation results, and documentation agreed for the project. Your legal, security, or compliance reviewers assess those materials against their requirements. We document the scope and any unresolved gaps.
How do you balance review with useful automation?
We test the workflow with the people who use it, including the time and effort required for review. Your team decides which actions can run within defined permissions and which need approval, based on evaluation results and the consequences of an error.
Who handles an AI error?
Before launch, we agree operational owners for monitoring, incident review, and pausing actions. The operating procedures identify who receives an alert, who investigates it, and who can approve a change or a return to service.
Does this slow the project down?
Evaluation and review take work. We scope them with the rest of the build, based on the use case, the available data, and the consequences of an error. Testing before launch helps identify problems while there is still time to change the design.
How do you handle data that can't leave our environment?
We design the deployment around that restriction and verify where data is processed, stored, and logged. If a hosted model would require a transfer you do not allow, we assess models that can run inside your environment. Any permitted external service must also meet your requirements for retention and training use.
What can we share with security reviewers or a board?
We agree the package with your team: evaluation cases and results, permitted actions and approval rules, the scope of the logs, and procedures for handling failures. It records the controls implemented, their limitations, and issues that remain open.

What does your team need to check before launch?

Talk to an engineer