AI ControlsGuardrails
AI governance that decides who can see what.
Guardrails are the rules, settings and records that decide what AI can reach in your business and what your people may put into it. We build the controls, the documentation and the evidence trail behind them to the standard of our ISO 27001 and ISO 9001 certification, and we leave legal advice to your advisers.
Certified in the Claude Partner Network Microsoft Solutions Partner ISO 27001 ISO 9001 Independent since 1999
Deliverables
What you end up holding
Documents and configurations you can put in front of a client, an assessor or your board.
Which AI tools are in use, by whom, on whose accounts and what each one can reach.
Track · GovernAgreed by your leadership in a half-day workshop, written so your people can follow it.
Track · GovernWhat is allowed, what is not and who decides when someone asks for something new.
Track · GovernA written list of what your people may put into an AI tool.
Track · GovernSo nobody can set up their own AI account on your domain.
Track · GovernYour vocabulary, your tone and the lines that should not be crossed, built into the tools your first team uses.
Track · AdoptFrom a test we run on your first team's own work. It closes the phase.
Track · AdoptWho can reach what, checked before anyone's access grows, with the waiver recording an informed decision.
Permissions reviewScored against how often the work happens and how much decision-making it needs, including the candidates where a plain workflow wins.
Build · DiscoverWhat it does and why is recorded next to the skill, connection or agent itself.
Build · MakeReviewed on a set cadence, with drift caught and corrected.
Trackside · ReviewingWhat was used, which tickets were raised and how each was resolved.
Trackside · RunningEvery item on this list is produced by work described on the Track and Build pages. Nothing here is a separate purchase.
What the written policy has to contain, and why a template will not hold →
The engineering fact
Where the boundary is
A connected AI tool can open anything its user's account can open.
Ask it for a file that account can reach and it will find it. No setting narrows the connection to just what the tool ought to see.
So the group of people who are connected is the only boundary that holds, and neither a policy statement nor an assurance from us can stand in for it. That is why we cap the group, name it and review it before it grows.
The useful question about any AI rollout is therefore "who is connected, and what can each of those people already open", rather than "what can the AI see". Most businesses have never been asked it.
The permissions review
A check of who can see what, before anyone beyond your first team connects.
Permissions review
Nothing widens until we check who can see what.
The review sits on the route between your first team and everyone else, and we do not skip it to meet a deadline. We can promise that because it follows from the engineering fact above.
There are three ways through it.
- Your first team connects under a signed waiver that names them and caps its size.
- Widening to a defined team takes a review of that team's permissions.
- Rolling out across the business takes a full review of who can reach what.
The waiver records an informed decision rather than an approval, and it does not cover anything we already know to be wrong. If a widening would break a guardrail your leadership set in Govern, the first phase of Track, it does not proceed until that is resolved.
What we do not claim
No legal advice, and no promise of compliance with a named regulation.
- We do not give legal advice, and we are not your compliance function.
- We do not certify your legal position, and we do not promise compliance with any named regulation. We build the operational controls that position depends on.
- We do not claim your data can be technically fenced off inside a connected account, because no one can do that technically.
- We do not describe governance as a phase you complete. It runs from the first day to the last and is reviewed on a set schedule.
We tell you only what we know for certain.
How this stays current
Inside Trackside, our monthly service once you are live, access and policy get a light check each month and a documented review each quarter.
- Stable disciplines Data classification, access, logging, review and training have been stable for decades, so there is no new discipline to invent when we apply them to a new category of tool.
- Reviewed on a cadence Inside Trackside, Reviewing checks access and policy on a set cadence, with a light check each month and a documented review each quarter.
- Drift gets caught If something moves away from a guardrail that Govern set, that is where it surfaces and gets corrected, before it becomes the thing you have to explain.
Who does the work
- Screened before they start Every person on the team is BS7858:2019 screened and DBS checked before they touch your systems or your data, on every engagement and as standard.
- Audited by a third party ISO 27001 and ISO 9001 are third-party assessed, and we also hold Cyber Essentials. More than 60 of the team are engineers.
Assurance
The questions a compliance lead asks
Each answer says what we do and what we leave to your advisers.
You are an IT company, not a legal or compliance function.
Correct, and we do not give legal advice. We build the controls, the documentation and the records your legal and compliance position rests on, and we hold ISO 27001 and ISO 9001 for how we do that work.
Your legal advisers tell you what position to take. We make that position real and evidenced inside your systems.
How do I know your governance is real and not a document?
Because a third party audits it and it changes production configuration. Govern brings sign-in under your company account and restricts who can create AI accounts on your domain. Both of those are settings in your systems.
Can you guarantee our client data will not leave our control?
No. Nobody can promise that. What we can do is keep the group of people connected to AI small, know exactly who is in it and check before it grows. We also write down what your people are allowed to put into a tool. Who is connected is the thing that protects you.
What happens if someone breaks the policy?
Whatever your policy says. You decide that in Govern, in advance, so nobody is working it out in the moment. In Trackside we watch for AI use that has not been approved and tell you when we find it. If you want it blocked, that is chargeable work and we will tell you before we do it.
Will this satisfy our insurer or our assessor?
Not on its own, because we do not set their questions. We answer the questions they have asked you, in writing, and we tell you which ones your current position does not cover. Send us the questionnaire.
Who is accountable when the AI gets something wrong?
You are, which is why Adopt defines where a person must review output and where it can stand alone. Governance here is mostly a question of who checks what and when. We write those points down and name who holds each one.