Skip to content

Governance & Compliance

AI governance, installed as working systems

Six layers, from AI inventory to audit trail, shipped one scoped build at a time. Each build has its own signed acceptance criteria, and we run the controls after launch, measured against those criteria.

Updated November 2026

The problem

The binder is current. Production is not.

Shadow AI is already in production

Employees adopted AI tools nobody approved or recorded. You cannot govern systems you have not found.

Policies exist. Nothing enforces them.

The acceptable-use PDF is signed and filed. No control in production checks a single clause of it.

EU AI Act obligations are landing

Obligations attach by risk class. Without an inventory and risk tiering, you cannot say which obligations are yours.

Audit evidence is rebuilt by hand

Every regulator or customer question triggers weeks of screenshots and email archaeology instead of a query against an audit trail.

The approach

How we set up AI governance for your company.

Most AI governance arrives as a binder of policy documents and committee charters, while the systems in production keep changing underneath it. We install governance as working systems instead. Each technical control in the stack below ships as software that runs in your environment, against acceptance criteria you signed before the build started. The human steps, such as a DPIA review, get a named owner and a recorded decision.

We work in one scoped build at a time, each with its own signed acceptance criteria. The first build is usually an inventory and risk tiering of everything already running, including the shadow AI your teams adopted without asking. The exposure map that comes out of it decides which layer we install next.

This is how we already work. A European pharmaceutical regulator runs an AI compliance scanner we built: 620+ marketing assets scanned, 11 rules per scan, around 2 minutes per asset, down from 2 to 3 hours of manual review. The scanner applies the rules on every run and keeps a record of each check, which is the evidence an auditor asks for.

First build: inventory and risk tiering · Then: layers in order of exposure

The framework

The QueryNow governance stack.

Six layers, each installed as working controls. For every layer you see what we install and the failure it prevents. The last line is an example of acceptance criteria we would sign before the build starts.

  1. AI Inventory

    Know what is running before you govern it.

    What we install

    • Shadow AI detection
    • AI tool register
    • Approval route for new tools
    • Risk tiering
    • Ownership assignment
    • Model registry

    The failure it prevents

    An unsanctioned chatbot has handled customer data for months. Nobody has recorded it or assessed its risk, and no one owns it.

    Acceptance criteria we would sign

    Every AI system in use, including tools surfaced from network and SaaS logs and AI features inside department software, appears in the AI tool register with an owner and a risk tier. A request for a new tool reaches a named approver through one intake form. A spot check by your team finds zero missing entries.

  2. Data Foundation

    Trace every model back to the data that feeds it.

    What we install

    • Source tracking
    • Lineage mapping
    • Quality validation
    • Freshness monitoring
    • Data bias screening

    The failure it prevents

    A model quietly retrained on stale, unvetted data, and nobody can say where that data came from.

    Acceptance criteria we would sign

    For any dataset feeding a tiered system, your team sees in one view where the data came from and when it last passed its quality and freshness checks, with bias screening results attached.

  3. Data Security & Access

    Control who and what can touch the data.

    What we install

    • Encryption
    • Anonymization
    • Role-based access
    • Least privilege
    • Key management

    The failure it prevents

    A retrieval path that hands restricted records to employees who were never cleared to see them.

    Acceptance criteria we would sign

    A test account outside the approved role cannot retrieve restricted records through the model or its logs. Encryption keys are replaced on the agreed schedule, and your security team can show the recorded changes.

  4. Model Assurance

    Prove each model does what its documentation says it does.

    What we install

    • Model cards (purpose and limits)
    • Performance benchmarks
    • Fairness testing
    • Red-teaming (tests that try to make it fail)
    • Drift detection

    The failure it prevents

    A model that degraded for months and was caught by a customer complaint instead of a dashboard.

    Acceptance criteria we would sign

    Each tiered model ships with a model card and a benchmark your team can rerun. A test that feeds the model deliberately shifted input data triggers a drift alert within 24 hours.

  5. Human Oversight

    Keep a named human between the model and the consequence.

    What we install

    • Decision review
    • Escalation paths
    • Override authority
    • Output validation
    • Accountability mapping

    The failure it prevents

    An automated decision that hurt someone, with no reviewer on record and no one accountable.

    Acceptance criteria we would sign

    In a sampled week of decisions, every output above the risk threshold shows a named reviewer and a recorded outcome. A reviewer override halts the downstream workflow.

  6. Compliance & Audit

    Turn the first five layers into regulator-ready evidence.

    What we install

    • EU AI Act mapping
    • NIST AI RMF and ISO/IEC 42001 mapping
    • Data protection impact assessment (DPIA, GDPR Art. 35)
    • Works council steps in Germany
    • Incident reporting
    • Audit trails

    The failure it prevents

    A regulator question that takes your team three weeks of manual reconstruction to answer.

    Acceptance criteria we would sign

    Every register entry carries its EU AI Act risk classification and the NIST AI RMF and ISO/IEC 42001 controls it maps to. A simulated incident produces the system record for the report from audit trail data, with no manual reconstruction, and the responsible team adds its assessment.

Frameworks and Germany

The frameworks your auditors name, and the steps Germany adds.

Layer six maps every control to the frameworks a customer audit or a regulator will ask about: the EU AI Act, GDPR, the NIST AI Risk Management Framework (NIST AI RMF) and ISO/IEC 42001, the international standard for an AI management system. One control then answers a US questionnaire and a European one with the same evidence. Our guide to NIST AI RMF and ISO/IEC 42001 sets out how the two fit together.

European companies have two more steps. When an AI tool processes personal data in a way that is likely to put people at high risk, GDPR Art. 35 asks for a data protection impact assessment (DPIA) before it goes live. We assess with your data protection officer whether one is needed and prepare it where it is. In a German company with a works council, a system that can monitor staff needs the council's agreement before it goes live, and Copilot counts as such a system. We prepare the works council information pack and the technical annexes for a works agreement with your counsel. Regulated companies can see the full set on our AI governance page for life sciences.

Most companies we meet start this work because of Microsoft Copilot (formerly Microsoft 365 Copilot). Turning it on for everyone raises the governance questions at once: which SharePoint sites it can search, which labeled files a data loss prevention policy stops it from using, who may build agents and what staff are told about using it. Our Copilot AI Foundation sets up the data controls, the AI policy, the AI tool register, the agent rules and the training before company-wide rollout. For agents built on any platform, our guide to governing AI agents covers ownership, risk tiers and retirement.

How it works

How a governance build runs.

Inventory before policy

The first build finds everything already running, including shadow AI, and tiers it by risk. The order of the next layers follows from what that inventory shows.

Criteria signed on day one

Each layer ships against written acceptance criteria you sign before the build starts. Your team can watch each criterion pass or fail.

Built in your environment

Controls run inside your tenant against your identity model. Nothing is hosted on our side and handed over later.

Evidence as a byproduct

Every automated control writes its own audit trail as it runs, and owners record their reviews and approvals next to it. When a regulator asks, your team starts from that record.

Your team operates it

Each build ends with handover. Your team gets the runbook and the access to own the control without us.

EU AI Act-aligned by construction

Layer six maps every system to its risk class, and every implementation we deliver is aligned with the EU AI Act.

Our method

From one scoped workflow to an agent in daily use.

  1. Day one

    Scope and sign

    We define the workflow and the acceptance criteria it has to meet, and agree them in writing before the build starts.

  2. The build

    Build

    We build the agent in your environment, with Claude Code and automated evaluation against your own data, and test it against the criteria you signed.

  3. After launch

    Run

    We run, monitor and improve the agent after launch, measured against those criteria, with a named person on your side approving what it produces.

Compliance & security

  • Built to SOC 2, HIPAA and GDPR standards
  • EU AI Act-aligned delivery
  • Microsoft Solutions Partner

How we engage

Four ways to work with us.

We scope one governance layer with you and sign written pass-or-fail acceptance criteria on day one. Scoping also names who on your side owns the layer, the access we need and the reviewers who sign off. A full six-layer install runs as repeated builds on the same terms, and you can stop after any build with working controls in hand.

Advisory

Strategy, roadmap and readiness for an AI program.

Build

One scoped workflow at a time, with acceptance criteria agreed in writing before the build starts.

Run

Managed operations after launch, measured against those criteria.

Outcome-linked terms

Commercial terms can be tied to the acceptance criteria you sign.

Every engagement starts with one scoped workflow and a written definition of done. Commercial terms are set in the statement of work.

Frequently asked

Questions buyers ask.

What is the QueryNow governance stack?

Six layers that cover the lifecycle of AI in your company: AI inventory, data foundation, data security and access, model assurance, human oversight, and compliance and audit. Each layer is a set of working controls we install and test in your environment, with the policy text it enforces attached.

How is this different from a governance framework engagement?

A framework engagement produces documents that describe what should happen. We install systems that make it happen. Each layer combines technical controls, which we install and test in your environment against signed acceptance criteria, with named owners who record the reviews and decisions a control cannot make, so each rule in the policy has an owner and a check.

We already have AI in production. Where do we start?

The first build is an inventory and risk tiering of everything already running, including shadow AI surfaced from network and SaaS usage. The result is an AI tool register with an owner and a risk tier for every system, and that register decides which layer we install next.

How do the builds work for governance?

Each layer is one scoped build with its own signed acceptance criteria. We scope the layer, sign the criteria on day one, build in your environment, and hand over working controls once every criterion passes. A full six-layer install runs as repeated builds on the same terms, and you can stop after any build with working controls in hand.

Have you built governance systems before, or is this new?

We have delivered 200+ production deployments on AI and the Microsoft platform since 2014, including an AI-driven digital workplace for Rockwell Automation serving 28,000+ employees in 80+ countries. The controls in this stack are the same kind of build: scoped and signed up front, then shipped and run.

Does this cover the EU AI Act?

Layer six maps every inventoried system to its EU AI Act risk classification, with audit trails and incident reporting attached to it. We build systems to SOC 2, HIPAA, and GDPR standards, and every implementation is aligned with the EU AI Act.

Do you map to NIST AI RMF and ISO/IEC 42001?

Yes. Layer six links each control to the NIST AI Risk Management Framework and to ISO/IEC 42001, as well as to the EU AI Act and GDPR. If you plan to certify an AI management system under ISO/IEC 42001, the register, the risk tiers and the audit trails are the evidence an auditor reviews. The certification itself comes from an accredited body.

What about the DPIA and the works council in Germany?

We assess with your data protection officer whether a data protection impact assessment (DPIA) is needed under GDPR Art. 35, and prepare it where it is. Where a works council has a say over the tool, we prepare the information pack and the technical annexes for a works agreement with your counsel, before the tool goes live for staff in Germany.

Research paper

The six-layer AI governance stack: Governance that ships, not governance that stalls

AI governance in six layers, installed as running systems instead of policy binders.

Read the paper
Get the designed PDF

Ready to ship AI that actually works?

Every engagement starts with one scoped workflow and a written definition of done. Commercial terms are set in the statement of work.

  • Built to SOC 2, HIPAA and GDPR standards
  • EU AI Act-aligned delivery
  • Microsoft Solutions Partner