Govern AI in the systems where it acts.
Policies say what AI may do. Aribot checks what it does: which AI is in use, what each agent is allowed to touch, what it did, and what it cost, with a record you can hand to an auditor.
Find the AI nobody registered
Shadow AI discovery looks across connected repositories for model endpoints, vector databases, agent frameworks and leaked AI keys, so governance starts from what exists in your code.
- Unmanaged models and AI services surfaced from your own code
- Each finding owned and tracked like any other risk
- The free AI readiness self-assessment covers the organisational side
Give every agent only what it needs
The agent governance hub lists your fleet, its entitlements and what each agent has used. The scope designer shows what a chosen set of permissions would allow and block before you grant it, and each agent gets its own assessment.
- Fleet, entitlements, traceability and assurance in one hub
- Watch-only mode first: see what would have been blocked, change nothing
- Switch to enforce when the boundary is right, one capability at a time if you prefer
Stop an agent that steps out of line
The actions queue collects what needs a decision. Block an agent, reduce its access or reissue a narrower permission, and the decision is recorded with its reason.
- Block cuts the agent back to its minimum permissions and clears its session
- Reduce access lets the current safe action finish, then removes the risky permission
- Every agent call passes the same checks, whether it arrives over MCP, the API or the app
A record you can prove, and a bill you can explain
Agent and AI actions are written to a tamper-evident audit trail that can be re-checked independently. AI cost shows spend by model, feature and agent, so usage and budget are discussed with the same numbers.
- AI audit trail: any later change to a record shows up on verification
- AI cost: spend by model and by agent, from your own usage
EU AI Act evidence, assembled as you work
The EU AI Act readiness view compiles a conformity dossier for each AI system from its architecture, controls and findings. Framework coverage reports show where you stand against the NIST AI RMF, ISO/IEC 42001 and related frameworks.
- Start with the free AI readiness self-assessment
- Continue with a detailed, evidence-backed assessment on Ayurak Trust
AI policy is written once. AI acts all day.
Most organisations now have an AI policy. It names approved tools, sets out who may use customer data with a model and asks teams to register new AI systems. Then the real work happens somewhere else. A developer adds a model call to a service to summarise claims. A data team stands up a vector database for a pilot that quietly becomes production. A support tool ships with an agent that can open and close tickets. An assistant connected to a code repository is given a token that can read every project, because that was the quickest way to make the demo work.
None of this is reckless. It is how useful software gets built. But it leaves three gaps that a written policy cannot close on its own. Nobody can say with confidence which AI is running. Nobody can say what each agent is allowed to touch, as opposed to what it was meant to touch. And when something goes wrong, nobody can produce a record of what the AI did that would survive an auditor asking whether it was edited afterwards.
Regulation has caught up with this. The EU AI Act expects providers and deployers to know their AI systems, keep records and keep a human able to intervene. The NIST AI RMF and ISO/IEC 42001 ask for an inventory, assigned responsibility, measurement and retained evidence. These are technical questions dressed as governance questions, and they are answered with technical evidence.
Aribot answers them where the AI acts: in your repositories, at the point an agent calls a tool, in the record of what happened and in the bill. The policy stays yours. Aribot shows whether reality matches it.
Three people, one shared picture
AI governance fails when each owner keeps a separate spreadsheet. The AI lead tracks what was launched, the security team tracks access, and compliance tracks obligations. Aribot gives all three the same inventory, the same record and the same numbers.
AI leadYou are asked to ship assistants and agents quickly and to answer for them afterwards.
A fleet view of every agent, what it is allowed to do and what it has done, and a scope designer that lets you size authority before you grant it, so a launch is not held up by a policy debate nobody can settle with evidence.
CISOYou need to know which AI is running, what it can reach and how quickly you can stop it.
Shadow AI found in your own repositories, every agent call held to the same checks whether it comes over MCP, the API or the app, and a stop you can use in the middle of an incident without waiting for a deployment.
Compliance leadYou must show regulators and auditors what AI exists, how it is controlled and what it did.
A tamper-evident audit trail you can re-check, AI spend you can explain, a conformity dossier for each AI system under the EU AI Act, and coverage against the NIST AI RMF and ISO/IEC 42001.
Shadow AI, found in your own code
Governance has to start from the AI that exists in your code. Shadow AI discovery looks across the repositories you connect for the signs of AI in use: calls to hosted model endpoints, vector databases, agent frameworks and AI provider keys left in code.
Each finding lands in one list with the provider, the model or service, where it was found down to the file and line, whether a key is hard-coded, and a risk level. From there it is treated like any other risk: give it an owner, decide whether it is an approved system to register or something to retire, and track it to a close. A key found in code is a finding in its own right, because it gives anyone who reads the repository a way to spend your money and reach your data.
The list exports to CSV for the people who still live in spreadsheets, and it tells you which AI systems your EU AI Act work must cover. When the next audit asks which AI systems you operate, the answer comes from your code and from nobody's memory.
Every agent call passes the same checks
An agent is only as safe as the authority it holds. In Aribot every agent call, whether it arrives over MCP, the API or the app, passes the same sequence of checks before anything happens, and the result is recorded whatever the outcome. The checks are listed below in plain terms. The figure follows one call that is stopped part-way.
- 1Authenticate. The agent proves who it is. An unknown caller goes no further.
- 2Check licence. The agent is one your organisation has brought under governance.
- 3Consent and scope. The person or system the agent acts for agreed to this kind of action.
- 4Role and company. The action stays inside the agent's role and your company's data.
- 5Allowed?. The specific action is within the authority this agent was granted.
- 6Real-time stop check. Nobody has blocked or reduced this agent since its last call.
- 7Threat check. The request does not match a known threat for this system.
- 8Run action. Only now does the action happen.
- 9Meter. Use and cost are counted against the agent and its budget.
- 10Record. The outcome, allowed or blocked, is written to the audit trail.
Watch-only first, then enforce
In watch-only mode, every check runs and every call is recorded, but nothing is stopped. You get a list of what would have been blocked and why. That list is the conversation you need to have with the team that owns the agent: either the agent needs the permission and the scope should grow, or it does not and the call should be refused. Switch to enforce when the list holds only calls you want stopped, for the whole fleet or one capability at a time.
Size the scope before you grant it
The scope designer shows what a chosen set of permissions would allow and block, using what the agent has used so far, so you grant the smallest scope that lets it do its job. The per-agent assessment then puts the authority an agent holds beside the authority it has used, which is where standing risk tends to hide.
Stop an agent in the middle of an incident
When an agent misbehaves, the question is how fast you can stop it without breaking everything around it. The actions queue collects what needs a decision, from agents whose use has drifted beyond their scope to blocks raised in enforce mode, and offers three responses.
- Block
For when the risk is now. The agent keeps only the bare minimum, its live session ends, and whatever it tries next is refused without anyone redeploying anything.
- Reduce access
The safe action already under way finishes, then the risky permission is removed, so a half-done task does not leave data in a broken state.
- Reissue narrower
The agent gets a new, smaller permission that covers what it legitimately needs, and keeps working inside it.
Every decision is written to the audit trail with who made it and why, so the post-incident review starts from facts.
A record you can re-check
Logs are easy to keep and easy to doubt. When a regulator or an auditor asks what an AI system did, the follow-up question is always whether the record could have been changed since. Aribot writes agent and AI actions to a tamper-evident audit trail. Anyone with access can re-check it at any time, and if a record was altered after it was written, the check shows where the record stops being trustworthy.
The trail holds allowed calls and blocked ones, access changes with their reasons, and the spend attached to each. It exports as an evidence pack for an audit, so the person asking does not need an Aribot account to review it.
And a bill you can explain
AI cost shows spend by model, by feature and by agent, from your own usage, with budgets, alerts and a forecast. Finance, engineering and the AI lead talk about the same numbers, and an agent that suddenly costs five times more than last week is a question someone asks the same day. Nobody waits for the end of the quarter.
Six weeks at Halvermere Insurance
Halvermere Insurance is an invented mid-sized insurer, and every number here is illustrative. Its board has asked the CISO, the head of AI and the compliance lead for one thing before renewal season: a clear answer on what AI the company runs and how it is controlled, ready for the EU AI Act.
- Week 1Where we stand on paper
The compliance lead takes the free AI readiness self-assessment. The result shows strong policy but weak evidence: there is an approved-tools list and an AI register, but no way to confirm the register is complete, and no record of what agents do.
- Week 2What is running
The platform team connects Halvermere's repositories to Aribot. Shadow AI discovery lists eleven findings. Seven are known systems. Four are not: a model endpoint in the claims service, a vector database in a pricing pilot, an agent framework inside the support tool, and an AI provider key committed to that same tool's repository. Each gets an owner by the end of the week, and the key is rotated.
- Week 3Watch, do not block
The claims-triage agent and the support agent are brought under governance in watch-only mode. Over ten days the would-have-blocked list fills up. Most entries are harmless reads the scope should allow. One is not: the claims agent occasionally asks to delete claim records, a permission it was given during testing and never needed.
- Week 4Size the scope, then enforce
The head of AI uses the scope designer to grant the reads the agents use and to leave deletion out. Enforce mode goes on for the claims agent. Its next delete request is stopped at the "allowed?" check and recorded with its reason, exactly as in the figure above. Claims handling carries on as before.
- Week 5An incident, handled in minutes
A change to the support tool makes its agent loop on a ticket and run up model spend. AI cost raises the budget alert the same afternoon. The CISO uses Reduce access from the actions queue: the current reply finishes, the agent can no longer reopen tickets, and the fix ships the next day without the agent being switched off.
- Week 6The board pack
The compliance lead opens the EU AI Act readiness view for the claims system and finds the dossier mostly assembled from work already done. The audit trail is re-checked and exported as an evidence pack. Halvermere books a detailed assessment on Ayurak Trust to put it all in a form the board and the auditors accept.
Nothing in those six weeks needed a new policy. It needed the policy Halvermere already had to be checked against what its AI was doing.
EU AI Act evidence, assembled as you work
Regulatory evidence is cheapest when it is a by-product of running the system well. The dossier for each AI system fills in as you model it, fix its findings and govern its agents, and the gaps are listed so nobody discovers them the week before an audit. The same evidence feeds every framework report, so one piece of work counts towards several obligations and is done once. The table below shows where each public standard draws its evidence from.
Mapping to a standard shows where evidence comes from. It is not a certification, and conformity decisions remain with your organisation and its assessors.
From a free self-assessment to evidence
The free AI readiness self-assessment is the quickest way in. It takes minutes, needs no sign-up to start and shows where your organisation stands against the EU AI Act, the NIST AI RMF and related frameworks on policy and process.
Aribot then shows whether your AI matches those answers: what is running, what each agent may do, what it did and what it cost. When you need a formal result, a detailed assessment on Ayurak Trust, carried out with Ayurak's team on assessment templates by Aristiun, backs each answer with evidence and reports it in a form you can hand to a board or an auditor. Existing Trust customers can go straight there.
Frequently asked
- Do we have to switch on enforcement straight away?
- No. Most teams start in watch-only mode. Every agent call is checked and recorded, and you see which ones would have been blocked, but nothing is stopped. When the would-have-blocked list contains only calls you want stopped, switch to enforce, for the whole fleet or one capability at a time.
- What counts as shadow AI?
- Any AI in use that nobody has registered: a call to a hosted model from a service, a vector database added to a data pipeline, an agent framework in a support tool, or an AI provider key committed to code. Discovery looks across the repositories you connect and lists each finding with where it was found, so it can be given an owner and a status.
- What happens when we block an agent?
- Block cuts the agent back to its minimum permissions and clears its session, so its next call is refused. Reduce access is the gentler option: the current safe action finishes, then the risky permission is removed. Either way the decision, who made it and the reason are written to the audit trail.
- Why is the audit trail called tamper-evident and not tamper-proof?
- Because the honest promise is detection. Anyone with the right access can re-check the trail at any time, and if a record was changed after it was written, the check shows where. That is what an auditor needs: not a claim that nothing could ever change, but a way to prove nothing did.
- Does this only cover agents we built ourselves?
- It covers any agent or assistant that reaches your systems through the governed gateway, including assistants such as Claude, ChatGPT or Gemini connected over MCP. Shadow AI discovery covers what lives in your own repositories. AI you use only as a website, outside your code and your gateway, is the part the free readiness self-assessment helps you capture as policy.
- How does the EU AI Act dossier get built?
- For each AI system you model in Aribot, the EU AI Act readiness view gathers its architecture, the controls in place and the open findings into one dossier, and shows what is still missing. It is kept current as the system changes, so nobody writes it in a rush before an audit. Classification decisions and sign-off stay with your people.
- What does AI cost show?
- Spend by model, by feature and by agent, from your own usage, with budgets, alerts and a forecast. When a team asks for a bigger model or another agent, the conversation starts from what the current ones cost.
- How do the free self-assessment, Aribot and Trust fit together?
- The free AI readiness self-assessment shows where your organisation stands on policy and process, in minutes. Aribot governs the AI that is running. A detailed assessment on Ayurak Trust, done with Ayurak's team, backs each answer with evidence, much of which Aribot already holds, and reports it in a form a board or auditor accepts.
- Can we try it on our own agents before buying?
- Yes. You can start free with Aribot and connect an assistant in a few minutes. For a fleet of agents, book a call and Ayurak's team will walk through onboarding in watch-only mode with you, so you see your own would-have-blocked list before anything changes.
Where you will find it once signed in
Demos and related pages
The standalone product for governing AI agents in real time.
Bring your agents under governance, step by step.
Use Aribot from Claude, ChatGPT, Gemini or your IDE.
Every agent call through one governed gateway.
Granted authority beside the authority used.
What an identity holds, next to what it has ever used.
Set-up steps for each AI assistant and tool.
Detailed, evidence-backed framework assessments.
A self-assessment against the EU AI Act, the NIST AI RMF and more.
Start with the free self-assessment. Go deeper with evidence.
- Step 1, free
Answer the AI readiness self-assessment online. You get a view of where you stand against the EU AI Act, the NIST AI RMF and related frameworks, in minutes.
- Step 2, with Ayurak's team
A detailed, evidence-backed assessment on Ayurak Trust: each answer supported by evidence, reviewed with our team, and reported in a form you can hand to a board or an auditor.
Pick your starting point
Most teams begin with the free AI readiness self-assessment, then govern their agents in Aribot. Existing Ayurak Trust customers can continue straight to a detailed assessment.
