Skip to content
BennettTechnologies

AI Governance & Security

Governance is part of the architecture. Not an afterthought.

Enterprise AI has to balance capability with predictable control. We engineer the boundaries — who can access what, which actions need approval, how output is checked and how behaviour is recorded — into the systems themselves.

Governance domains

Eight areas we design for when agents meet production.

Governance documents matter, but controls that only exist on paper do not stop an agent at runtime. Each of these domains maps to concrete engineering.
  1. 01

    Identity and access

    Determine which user or agent can access which operation or dataset.

  2. 02

    Data boundaries

    Enforce tenant, location, record and organizational segmentation.

  3. 03

    Human approval

    Require explicit authorization for selected sensitive or irreversible actions.

  4. 04

    Output validation

    Evaluate model responses before relying on them for consequential decisions.

  5. 05

    Maker-checker architecture

    Separate generation from validation or approval where appropriate.

  6. 06

    Auditability

    Record meaningful actions, decisions, relevant context and policy outcomes.

  7. 07

    Model governance

    Assess providers, models, data handling and change management.

  8. 08

    Operational resilience

    Handle timeouts, malformed output, unavailable providers and rollback.

Risks and controls

Concrete risks, specific controls, honest residual risk.

No control removes risk completely. We document what each control addresses and what remains, so decisions about autonomy are made deliberately.

Risk 01

An agent retrieves data from another customer account.

Engineering controls

Tenant scope derived from the authenticated session, never from model-supplied parameters; isolation enforced in the API and data layer; automated cross-tenant tests.

Residual risk

Defects in application authorization code still need review and testing; isolation is only as strong as the layer that enforces it.

Risk 02

A model recommends an unsupported operational action.

Engineering controls

Tools expose only defined operations; recommendations are validated against business rules; uncertain or out-of-policy suggestions go to review.

Residual risk

A plausible but wrong recommendation can still reach a reviewer, so reviewers need the evidence behind it.

Risk 03

A write operation bypasses a required approval.

Engineering controls

Approval enforced server-side on the operation itself, not in the prompt; approval state stored and checked at execution; audit of every attempt.

Residual risk

Approval policies must be kept current as workflows change.

Risk 04

A model or tool update changes system behaviour.

Engineering controls

Pinned model versions where available; regression evaluation before changes; staged rollout and rollback; monitoring of quality signals.

Residual risk

Evaluation sets cover known cases; novel inputs can still behave differently.

Risk 05

Sensitive information enters the wrong model context.

Engineering controls

Data classification; redaction before model calls; provider and region selection based on data handling terms; minimal context by design.

Residual risk

Classification and redaction rules can miss unusual data formats and need periodic review.

Risk 06

An external integration becomes unavailable.

Engineering controls

Timeouts, retries with limits, circuit breakers, graceful degradation and clear user messaging; fallback models where justified.

Residual risk

Some features will be unavailable during an outage; the goal is predictable failure, not zero failure.

Maker-checker architecture

Separate the component that proposes from the one that approves.

In a maker-checker design, one component — often an AI agent — produces output or proposes an action. A separate process checks it before execution: deterministic validation, a policy engine, a second model with a narrow evaluation task, or a person.

We apply it where the risk warrants it. Maker-checker reduces the chance that a single faulty step causes harm; it does not, on its own, guarantee correctness or security. Permissions, validation rules and monitoring still do much of the work.

Maker-checker workflowA maker component proposes an action. A separate checker — automated validation, policy or a person — reviews it. Approved actions execute and are logged; rejected actions return with reasons.MAKERAgent proposesdraft · change · replyCHECKERValidate / approverules · policy · personEXECUTEApply & recordscoped write · auditrejected · with reasons

Governance design framework

A simple, repeatable lifecycle.

We use the same sequence whether governance is part of a new build or applied to AI features already in production.
  1. 01Discover
  2. 02Classify
  3. 03Design Controls
  4. 04Implement
  5. 05Test
  6. 06Monitor
  7. 07Improve

Discover & classify

Inventory AI use cases, data flows, providers and actions; classify them by data sensitivity and impact.

Design, implement, test

Choose controls proportionate to the risk, build them into the system and test them like any other requirement.

Monitor & improve

Watch real behaviour, review incidents and near-misses, and adjust controls and autonomy accordingly.

Frameworks such as the NIST AI Risk Management Framework and ISO/IEC 42001 can be useful references when structuring an AI governance programme. Our engineering work is not a compliance certification or legal assessment; where you need one, we work alongside your compliance and legal advisers.

Introducing AI into sensitive workflows?

Tell us which systems and decisions are involved. We'll help you decide where AI can act, where it should recommend, and where it should stay out.