AI governance
Know what your AI is doing, before someone asks you.
AI spreads through a company one team at a time, and the policy written about it rarely reaches the place where it is actually used. Nexus keeps the list of what is in use, the limits on who may use it, the checks on every answer and the record of what happened, in the same place the work runs.
Every use passes the same controls
Permission decides which AI may be reached and on what budget. The checks decide what may be sent, and what may come back.
- of every AI in use
- One listof every AI in use
- limits and spending caps
- Per teamlimits and spending caps
- checked before it leaves
- Every answerchecked before it leaves
- of what the AI did
- One recordof what the AI did
The three questions
AI governance is three questions a company has to be able to answer.
Not a framework and not a policy document. In practice, every conversation about AI risk comes down to these, and most organisations cannot answer any of them from a system.
Which AI is the company actually using?
In most organisations nobody can answer this, because teams adopt tools faster than anyone records them. Nexus keeps a single list of every AI model in use, who owns it and what it costs, so the answer comes from a system rather than from asking around.
Who is allowed to use it, and for how much?
Access is usually a password shared by whoever set the account up, with no limit attached. Nexus issues access per team, restricted to the AI they are meant to use, with a spending cap and an end date on it.
What did it do, and can we show someone?
When a customer complains, an auditor asks or a regulator writes, the question is what the AI was given, what it answered and who acted on it. Nexus keeps that as a record of the work itself, so it exists before anyone needs it.
What is in use
One list of every AI the company runs.
You cannot govern what nobody has written down. The list is kept by the platform the AI runs through, so it stays current without anyone maintaining a spreadsheet.
What exists, and whose it is
Every AI model available to the organisation is registered in one place with an owner against it, whoever provides it. Adding, replacing or withdrawing one is a decision taken once, not a change chased through every team that happened to use it.
What it costs, and to whom
Spend is tracked per model as it happens, so the question of what AI costs the business, and which department is spending it, has an answer someone can own rather than a number that arrives after the fact.
Which version answered
Models are replaced and updated. The platform keeps track of which one was live at the time, so an answer given six months ago can still be explained against the AI that actually produced it.
Who can use it
Give a department AI without giving it a blank cheque.
Access is granted per team, with the limits attached to it rather than written in a policy beside it.
Only the AI they need
A team gets access to the models chosen for their work and nothing else, even though everything runs on the same platform.
A spending cap that holds
Each access comes with a budget. Consumption stops at the limit instead of appearing on an invoice at the end of the month, and you can be alerted before a team reaches it.
An end date
Access granted for a pilot, a supplier or a fixed project stops working on the date you set, without anyone having to remember to withdraw it.
Without sharing the company account
Teams never hold the credentials for the underlying accounts. Access can be given, reduced or revoked centrally, which also means a departure does not leave an open door.
Separation between departments
Finance, HR and the service desk work in separate spaces, each with its own AI, its own documents and its own people. One department cannot reach another's material, which is usually the first thing asked about shared AI. Who signs in, with what role and with what authority to approve comes from the identity system you already run, and is covered on governance and audit.
What it may say
The things you cannot afford to happen are stopped, not reported.
Checks run on what goes into the AI and on what comes back, every time, as part of the work. Nobody has to notice.
Customer and employee data
Personal data is detected and stopped before it reaches a model, and before an answer carrying it reaches a person or another system.
Answers you cannot send
Responses that breach your content standards are blocked or rewritten rather than reaching a customer, an employee or a document.
Attempts to manipulate the AI
Deliberate attempts to make the AI ignore its instructions are caught at the boundary, which matters most when the text came from outside the organisation.
Answers in the wrong shape
If an answer does not match what the next step expects, it is rejected there instead of quietly producing a wrong entry in a business system.
Your own rules
The standard checks are combined with rules written for your organisation, because most of what matters is specific to your institution rather than generic.
Proof it was tested
Controls nobody has tested are an assumption.
Saying the AI is safe is not the same as having had someone try to break it.
Nexus Cerberus
Cerberus attacks your own AI deliberately, the way an adversary would, and reports what it found. It keeps a record of where each AI model came from, and issues signed reports that an auditor, a client or a regulator can treat as evidence rather than as a statement from the supplier. It tests the AI in the environment where it actually runs, not a demonstration version of it.
What it did
When someone asks why, the answer already exists.
Each use of AI is recorded as part of the work it belonged to, which is what turns an explanation into something you can show.
What it was asked
The information the AI was given, including the documents it was allowed to read, is kept with the case it belonged to.
What it answered
The response is retained as given, alongside which AI produced it, so nothing has to be reconstructed from memory later.
Who decided
Where a person reviewed the answer, their name, role, decision and reason are recorded against the same case, not left in an email thread.
What changed as a result
If the work updated a business system, what was sent and what came back sits in the same record, which connects the decision to its consequence.
What this does not do
The limits, stated plainly.
Controls are worth more when you know where they stop. These are the points that matter in an evaluation, and we would rather you heard them from us.
This does not make an organisation compliant
It gives you the inventory, the limits, the enforcement and the evidence. Compliance still depends on the decisions you make with them: which AI is permitted, which uses are approved, who reviews what, and how long records are kept.
Checks reduce risk, they do not remove it
Some categories of problem are caught reliably and others only partly. That is why anything consequential still passes through fixed rules and a named person before it reaches a customer or a system.
A spending cap says nothing about quality
A budget controls what a team can consume. Whether the answer was right is a separate question, handled by the checks, the testing and the review step.
Someone still has to own the list
The platform records what exists and what it costs. Deciding which AI the organisation is willing to run, and removing what it is not, stays a human decision with a name against it.
Architecture session
Bring the AI you already run under one set of controls.
If AI is spreading across teams faster than anyone can keep track of it, we can map what is in use today, what each team should be allowed to spend, and what record every use ought to leave behind.