AI governance & secure delivery

AI Governance & Secure Delivery

AI needs to work in the real world — securely, measurably and with appropriate human oversight.

We design AI delivery around the organisation's existing security, governance, data and operational requirements rather than treating them as an afterthought.

The considerations below describe how we approach delivery. Controls are agreed and implemented with each organisation rather than assumed.

Data & privacy

Before an AI system is designed, it should be clear what data it may use, where that data lives and who may see it.

• Data minimisation

  Use the smallest amount of data that allows the use case to work, rather than connecting everything available.

• Appropriate data access

  Define which systems and records the solution may read, and under whose permissions it operates.

• Retention

  Agree how long inputs, outputs and logs are kept, and what is deleted at the end of that period.

• Sensitive information

  Identify personal, clinical, legal or commercially sensitive material early and decide how it is handled or excluded.

• Approved AI and model providers

  Work within the model and vendor list the organisation has approved, rather than introducing new processors by default.

• Data residency

  Establish where processing and storage must take place, and select services that can meet that requirement.

• Privacy requirements

  Align the design with the organisation's own privacy assessments, records of processing and data protection advice.

Security & access

AI systems inherit the security requirements of conventional software while introducing additional considerations around models, prompts, retrieved data, tool access and third-party AI services.

• Identity and access

  Integrate with the organisation's existing identity provider rather than creating separate user stores.

• Least privilege

  Grant each component only the permissions it needs to perform its function.

• Role-based access

  Where roles are meaningful to the use case, restrict data and actions by role.

• Authentication

  Require authenticated access to interfaces, services and administrative functions.

• Logging

  Record activity so that use, changes and failures can be reviewed after the fact.

• Secrets management

  Hold credentials and keys in the organisation's approved secret store, not in code or configuration files.

• Untrusted input reaching the model

  Treat documents, emails, web content and user text as untrusted. Instructions hidden inside that content — prompt injection — should not be able to change what the system does.

• Integrity of retrieved information

  Control what can enter a knowledge base and who can change it, so answers cannot be steered by material that should not be there.

• Tool and action permissions

  Where a system can act — send, write, update, spend — scope those actions tightly and require confirmation for the consequential ones.

• Preventing data leakage

  Limit what the system can retrieve for a given user, and control what leaves the environment, including what is sent to a model provider.

• Handling model output

  Treat output as untrusted content rather than as instructions or safe code, especially where it feeds another system.

• Client infrastructure requirements

  Build within the organisation's tenancy, network boundaries and platform standards where those apply.

AI evaluation

An AI system is only trustworthy if its behaviour has been tested against criteria agreed in advance.

• Defined success criteria

  Agree what a good answer or action looks like before building, and what would count as unacceptable.

• Accuracy and quality testing

  Test against a representative evaluation set drawn from real cases rather than demonstration examples.

• Hallucination evaluation

  Check whether outputs are grounded in source material, and report results against the evaluation set used.

• Failure scenarios

  Examine what happens with missing data, ambiguous requests, edge cases and service outages.

• Human review

  Decide which outputs need a person to confirm them before they take effect.

• Testing before production

  Evaluate in a controlled environment before the solution touches live work.

• Ongoing monitoring

  Continue measuring quality after launch, since data, models and usage all change.

Governance

Governance answers a simple question: if this system gets something wrong, who knows, who decides and who is accountable.

• Use-case risk assessment

  Assess each use case on consequence, reversibility and the sensitivity of the data involved.

• Responsibility and ownership

  Name the business owner of the system, not only the technical team supporting it.

• Vendor and model risk

  Choose models and suppliers against the requirement, record why, and consider what depends on that provider — data handling, availability, model changes and the route to an alternative.

• Auditability

  Keep enough of a record that a decision or output can be traced back to its inputs and sources.

• Human oversight

  Keep appropriate human review around consequential steps, and make it practical rather than nominal.

• Change management

  Treat prompt, model and data changes as changes, with testing and approval.

• Documentation

  Document how the system works, its limitations and what to do when it fails.

Responsible adoption

AI risk can arise from the model, data, architecture, integrations, vendors and the way people use the system. Effective governance needs to address all of these layers, including the people doing the work.

• Acceptable-use guidance

  Set out plainly what staff may and may not use AI tools for in their role.

• Staff training

  Teach people how the system works, where it is reliable and where their judgement is still required.

• Role-specific controls

  Match permissions and guidance to what each role actually needs to do.

• Escalation routes

  Give people a clear way to raise a wrong or unexpected output, and a person who responds to it.

• Adoption monitoring

  Track whether the system is being used as intended, and where it is being worked around.

Enterprise procurement

Raga AI can work through an organisation’s normal supplier onboarding, information-security review and contracting process as part of preparing an engagement. Tell us what your process requires — onboarding forms, security review, due-diligence questionnaires or the contract itself — and we will work through it with you.

Documentation and due-diligence requirements are confirmed individually for each engagement, so that what we tell you reflects the position at the time rather than a standing claim.

Enterprise contact

For procurement, security review, due-diligence questionnaires or privacy questions, contact info@ragaai.co.uk and mark your message for the attention of procurement or security. This reaches us directly — we would rather use one working address than publish a mailbox that is not monitored.