CapabilityAI and governance

    AI agents your compliance team can sign off

    Short answer

    We put governance into the systems that run AI: an inventory of every AI use, a written mandate and limits for each agent, human sign-off where it matters, evaluation before release and an audit trail you can hand to a reviewer.

    ControlsPer agent
    Inventoryevery AI use, owner, data
    Mandatewritten per agent
    Sign-offnamed person, enforced
    Evaluationbefore release and on change
    Audit trailexportable
    How an agent gets signed offPer agent
    1. 01Listedowner · data and risk classaccepted whenevery AI use has an owner and a risk class in the register
    2. 02Mandate writtenmay read, may do · limitsaccepted whenthe agent stops at its action limit
    3. 03Testedagreed cases · before releaseaccepted whenthe agent passes the agreed cases, and again after any model or prompt change
    4. 04Person signsnamed · enforcedaccepted whena risky action waits until the named person approves
    5. 05Loggedaudit trail · exportableaccepted whena reviewer can export every action, version and approval

    Problems

    Why governance has to live in the system

    P01

    Nobody knows where AI is used

    Teams adopt assistants and agents one by one. There is no list of what runs, on which data, and who owns it.

    P02

    Agents act without written limits

    An agent that can read, write or pay needs a mandate. Without one, nobody can say what it was allowed to do.

    P03

    No evidence for the regulator

    When an auditor asks how a decision was made, the prompts, model versions and approvals are scattered or lost.

    P04

    Policy on paper, not in the system

    The AI policy exists as a document, but nothing in the tooling enforces it.

    What is included

    AI inventoryA register of every AI use: purpose, owner, data, model and risk class, kept current as part of delivery.
    Mandates and limitsEach agent gets a written mandate: what it may read, what it may do, spending or action limits and when it must stop.
    Human sign-offRisky actions and releases wait for a named person. The system blocks them until that person approves.
    Evaluation before releaseAgents are tested against agreed cases before they go live, and again after every model or prompt change.
    Audit trailEvery agent action, model and prompt version, check result and approval is logged and exportable for review.
    Mapped to your frameworkControls mapped to the rules your compliance team works with, such as the EU AI Act, ISO/IEC 42001 or internal policy.

    How an engagement runs

    1. Inventory

      We list the AI already in use and the uses you plan, with owners and data.

    2. Controls design

      Mandates, limits, sign-off points and evaluation criteria, agreed with your compliance team in writing.

    3. Build into the tooling

      The controls go into the systems that run the agents, so the policy is enforced, not only written.

    4. Evidence and handover

      A working audit trail, a runbook and your team trained to add the next agent under the same controls.

    Common questions

    Q01Is this legal advice?

    No. We build the controls and the evidence. Your compliance and legal teams decide what the rules require; we map the controls to their reading.

    Q02Do we need this if we only use AI assistants?

    Usually a light version: an inventory, data rules and a log. Agents that act on systems or money need mandates, limits and sign-off.

    Q03Can it cover agents we did not build?

    Yes, as long as they run through systems we can connect to. Third-party tools are listed in the inventory and wrapped with the same logging where possible.

    Q04How is it priced?

    Per milestone against acceptance tests agreed in a written scope, after a scoping call. No price before scope.

    Next step

    Start with an inventory

    A scoping call with Oleg. You get a written scope for the inventory and the first controls before any build starts.

    Oleg Pun

    CEO, Vareger

    Oleg Pun, CEO, Vareger