For risk and AI governance leads
AI your risk review can actually assess
AI governance and guardrails for enterprise AI means each use is classified by risk, tested against agreed thresholds before release, limited in what it can see and do, and logged so a reviewer can rebuild what happened. We build those controls into the system and leave the evidence with your team.
01 How we work
What a reviewer gets from the work
-
A risk class for every use
Each AI use is classified by its impact on people and the business. The class sets the controls, the depth of testing and who approves the release.
-
Thresholds written before testing
Correctness, groundedness, refusal behavior and resistance to prompt injection are tested against thresholds agreed with you before any results are seen.
-
Limits on data and actions
Assistants and agents get least-privilege access to data and tools. A person approves consequential or irreversible actions, and every action is logged.
-
The frameworks you already use
Controls can be aligned with the NIST AI Risk Management Framework or ISO/IEC 42001 where your organization uses them. Alignment is not certification, and your own review decides.
What we'll tell you
We cannot promise that a system will pass your review or an audit. We can make sure the evidence your reviewers ask for exists and is easy to read.
Review questions
Pick the questions your reviewers will ask
Tick the questions that come up in your AI risk review. For each one you see the evidence a reviewer expects, and what we would build or change first to produce it.
Review questionnaire Tick the questions you expect
What to prepare
Tick a question to see the evidence a reviewer expects and what to build or change first.
-
Start here
What is this AI allowed to decide on its own?
- Evidence
- A written scope for the use, its risk class and the actions that need a person's approval, signed off by the business owner.
- First
- Define the approval points in the workflow and block consequential actions until a named person approves them.
-
Start here
How do you know the answers are correct?
- Evidence
- Evaluation results against a representative test set, with thresholds agreed before testing and the known limitations written down.
- First
- Build an evaluation set from real cases and run it on every change to the model, the prompts or the retrieval index.
-
Start here
Can it reveal data a user should not see?
- Evidence
- An access model showing which data each role can reach, plus test results for unauthorized disclosure through the assistant.
- First
- Enforce access at retrieval time with each user's own permissions, not a shared service account, and test it.
-
Start here
What happens if someone tries to manipulate it?
- Evidence
- Prompt-injection and misuse test results, the defenses in place, and how blocked attempts are recorded and reviewed.
- First
- Add input and output validation, restrict which tools can be called, and include injection cases in the evaluation set.
-
Start here
Can we rebuild what happened in a past decision?
- Evidence
- Logs that record the request, the sources used, the model version and any action taken, kept for your retention period.
- First
- Turn on structured logging for each request and decision, with retention set by your policy rather than a default.
-
Start here
Who owns each model once it is live?
- Evidence
- A model inventory that names the owner, version, risk class and next review date of each AI use.
- First
- Create the inventory with an owner for each use, and monitoring that alerts that owner when quality drifts.
How it starts
-
List the AI uses
We inventory the AI uses in place or planned, with their data, users and owners, and agree a risk class for each one.
-
Agree what evidence looks like
With your risk team we write down the tests, thresholds and records each risk class needs before it is released.
-
Close the highest-risk gaps
We build the missing controls on one use first: evaluation, access limits, approval points and decision logging.
-
Hand over the evidence
Your reviewers assess the evidence and decide. The approach extends to other uses only once they are satisfied with the first.
The usual engagement model: Advisory and architecture
Ready to talk it through?
Start a conversation02 A typical situation
What this usually looks like
A head of risk might come to us with an internal assistant that is ready to launch, and a committee still asking who checked its answers, what data it can see and how a wrong answer would be traced.
We would classify the assistant, run an evaluation against thresholds the committee agrees, enforce access with each user's permissions and switch on decision logging. The committee would get a short evidence pack it can question, not a slide deck.
03 Questions
Questions people ask
No. The audit or review decides. We build the controls and the evidence it will look at, and point out where a gap remains.
No. We work to the policies you already have, and controls can be aligned with the NIST AI Risk Management Framework or ISO/IEC 42001 where they apply. Your risk team keeps the decisions.
No. Controls follow the risk class. A low-impact internal summarizer needs far less than an agent that changes customer records, and over-controlling low-risk uses slows teams down without reducing risk.
Yes. We can assess an existing system against your evidence requirements, run the evaluation and injection tests, and list what to build or change first, in priority order.
04 Start
Give your reviewers evidence they can assess
Book a call with an engineer. We will go through your AI uses and the evidence your review expects, and say where the gaps are.
What happens next
- You pick a time, or send us a message.
- A person at Nexstella reads what you shared before we talk.
- You leave with a clear next step, even when that step is to wait.
Or write to [email protected]. Send a brief instead
Meet
Booking as .
Loading the calendar. If it does not appear, it opens on Cal.com too: open it in a new tab
Send us two or three times that suit you. A person replies within one business day. The invite goes to the address above.
Prefer that we send you times?
What happens next
- You pick a time, or send us a message.
- A person at Nexstella reads what you shared before we talk.
- You leave with a clear next step, even when that step is to wait.
Contact
Send us a message
Pick the topic that fits and tell us what you are working on. A person replies within one business day.
Want the full picture? Read Data governance, cybersecurity and responsible AI.
Also written for you: Application security review , Enterprise AI to production .
Last updated