11 min read · AI governance · AI strategy
How to build an AI governance framework for a mid-sized company
Capio team

A practitioner’s framework for mid-sized firms: a first-quarter build plan, clear risk tiers, approval gates and board metrics aligned to ISO/IEC 42001 and NIST AI RMF.
You can stand up effective AI governance in one quarter without adding a new bureaucracy. Start by agreeing principles and risk appetite, inventory every AI use and classify risk, set decision rights, then embed a few proportionate controls into delivery and operations.
Tie this to how you run pilots and move to production so good ideas pass through clear gates. Align with NIST AI RMF and move toward ISO/IEC 42001 over time, not all at once. Report progress and issues to the board monthly with a short pack.
What is an AI governance framework and why does a mid-sized company need one?
An AI governance framework is the policies, roles, processes and technical controls that guide how AI is designed, built, deployed, monitored and retired. It keeps AI aligned with your goals and risk appetite and makes performance, safety and accountability visible.
AI behaves probabilistically, learns from data and can drift over time. That is different from traditional IT and needs guardrails such as human in the loop, action limits, audit logs and a kill switch for unsafe behaviour. Getting this right helps a mid-sized firm scale AI safely and faster, instead of stalling in pilots.
How is AI governance different from traditional IT governance?
Traditional software is deterministic and easier to test exhaustively. AI models can be opaque, sensitive to inputs and can degrade with new data. Governance therefore focuses on:
- Risk tiering of use cases and proportionate controls
- Data quality, access and lineage as first-class concerns
- Human oversight for material decisions
- Ongoing monitoring for drift and incidents, not one-off acceptance
What outcomes should a mid-sized company expect if it gets this right?
- Faster time from idea to production because approval gates are clear
- Fewer reworks and incidents because risks are caught early
- Better regulatory readiness through an inventory and risk register
- Higher trust from customers and your board
What are the minimum building blocks of AI governance for a company of mid size?
You do not need a large function to start. A lean set of artefacts and roles will carry most firms through the first year.
- Principles and risk appetite agreed by leadership
- An AI inventory with owners, purposes and data sources
- Risk classification and a live risk register
- Roles and decision rights with a small committee that meets monthly
- Board reporting with a short monthly pack
- Policies and guardrails for data, model use, human oversight and escalation
- Delivery controls embedded in your pilot and release process
- Monitoring, incident response and audit logs
- Board reporting with a short monthly pack

What is minimum viable governance in year one?
Aim for simple, maintained artefacts:
- One page of principles and risk appetite
- A spreadsheet inventory of all AI uses, including shadow use
- A four-tier rubric for risk with default controls per tier
- Three lightweight policies: data use, model use, human oversight
- Approval gates wired into your pilot process
- A monthly half-page board update and a quarterly deep dive
Which roles and decision rights do we need from day one?
- Executive sponsor who signs off principles and appetite
- AI governance committee chair and coordinator
- Product owners for each AI use
- Legal, risk or compliance representative
- Data and security leads
- Engineering or platform lead
Give the committee real decision rights on approvals, exceptions and incident response, and time-box meetings to one hour monthly.
How do we classify AI use cases by risk and set proportionate controls?
Use case risk should drive controls. Keep the rubric simple so teams can apply it without a workshop every time.
What is a simple tiered rubric for risk and controls?
- Low: non-actionable insights, internal productivity, no customer impact
- Controls: owner named, basic logging, approved data sources
- Medium: content generation to customers with review, internal decisions with human validation
- Controls: prompt and output logging, human in the loop, action limits
- High: decisions affecting customers or staff, financial exposure, legal or safety implications
- Controls: documented justification, bias and performance testing, explainability where feasible, dual sign-off, rollback plan, kill switch
- Critical: safety, legal exposure, or regulated high-risk contexts
- Controls: all high-tier controls plus privacy and security reviews, stress testing, formal approval by the committee and executive sponsor, enhanced monitoring and audit
Which controls are mandatory for higher risk tiers?
At high and critical tiers require human in the loop for edge cases, audit logs for inputs and outputs, a defined action limit, rapid disablement, and a rollback plan. Where explanations are feasible, keep them with the decision record.

How do we build AI governance in the first quarter without adding a new bureaucracy?
Work in three short phases with visible artefacts each fortnight.
What happens in the opening weeks with the inventory, risk appetite and a charter?
Weeks one to three:
- Agree principles and risk appetite with the executive sponsor
- Stand up the AI governance committee and a one-page charter
- Build the first AI inventory from interviews and self-reporting, including shadow use
- Draft the risk tiering rubric and default controls
Deliverables: charter, principles and appetite, inventory v1, draft rubric.
What happens in the middle phase with policies, approval gates and impact assessment triggers?
Weeks four to six:
- Finalise three policies: data use, model use, human oversight
- Define approval gates for pilots and releases by tier, and wire them into your delivery workflow
- Define when an impact assessment is required and what it must include
- Stand up logging and monitoring basics for pilots
Deliverables: three policies, approval gate checklist, impact assessment template, logging plan.
What happens in the final phase with monitoring, audit logs and a board pack?
Weeks seven to nine:
- Pilot the gates on two or three real use cases and adjust the rubric
- Finalise audit log fields and retention
- Define incident response for AI issues
- Produce the first monthly board pack with leading and lagging indicators
Deliverables: tested gates, audit and monitoring checklist, incident playbook, board pack v1.

How do we connect governance to pilots and production so good ideas scale?
Put the gates where teams already work. Use your pilot pipeline to enforce proportionate reviews, then carry controls into production.
What approval gates should a pilot pass before MVP and scale?
Use three gates that map to common delivery stages:
- Proof of concept: confirm owner, purpose, data sources, risk tier, basic logging and a defined success measure
- Minimum viable product: complete policy checks, human oversight plan, action limits, rollback plan, documented tests, approval by the committee for high tiers
- Enterprise solution: performance and bias re-test on production data, monitoring configured, on-call or incident process, board visibility for high and critical
Typical timings in mid-sized firms: proof of concept in two to six weeks, minimum viable product in one to three months, then harden into an enterprise solution owned by the line team.
How do we embed human in the loop and kill switches in workflows?
Decide what triggers human review, who reviews and how to record overrides. For disablement, define what events require a stop, who can trigger it and how to revert safely. Keep both in the runbook and test them.

How should a mid-sized company align with ISO/IEC 42001 and NIST AI RMF pragmatically?
Use the frameworks as direction, not as day one checklists. Map your minimum viable governance to them, then deepen over time.
- NIST AI RMF is a voluntary risk framework published in 2023, useful for structure and shared language across teams. See the official publication for detail: NIST AI RMF 1.0.
- ISO/IEC 42001 is the international management system standard for AI, published in 2023. It helps you formalise governance as a system. See the standard page: ISO/IEC 42001.
What can we adopt now, and what can wait?
Adopt now:
- Inventory, risk register and tiering rubric
- Roles, charter and decision rights
- Policies and delivery gates
- Monitoring basics and incident handling
Plan next:
- Formal management system artefacts and internal audit cycles aligned to ISO/IEC 42001
- Broader stakeholder engagement, training and periodic control testing
What evidence would an auditor or regulator expect to see?
- A current AI inventory and risk register with owners
- Documented principles, policies and approval records by tier
- Logs that show inputs, outputs, prompts and decisions for material use
- Monitoring, incident reports and corrective actions
- Board reporting and meeting notes for the governance committee

What does the EU AI Act timeline change for non EU mid sized firms that use or sell into the EU?
The AI Act entered into force in 2024, with phased application. The AI Omnibus entered into force in July 2026 and adjusted some dates. If you place systems on the EU market or use them in the EU, plan now against the official timetable.
- Background on entry into force and the regulatory framework: European Commission summary
- AI Omnibus entry into force: EU announcement
- Official timeline including updated high-risk application dates: Council of the EU timeline
Which transparency rules already apply and when do high risk rules bite?
Some provisions apply from August 2026, including enforcement structures. Updated high-risk dates are later. Stand-alone high-risk obligations in Annex III apply from December 2027 and embedded high-risk obligations apply from August 2028, per the Council timeline. Use the official pages above to confirm the latest details.
How do we avoid rework as the deadlines approach?
- Keep your inventory and risk register current and flag any EU exposure
- Classify use cases against Annex III categories early
- For likely high-risk uses, keep evidence now that will satisfy later obligations
- Review third-party vendor roadmaps against the official timeline and ask for written commitments

What good looks like metrics, board reporting and a one page charter
You can fit governance on one page for the board while keeping the working artefacts underneath.
What goes into a monthly board AI governance report?
- Top risks by tier and any incidents or near misses
- Pipeline summary by gate with blockers
- Control test results for high and critical uses
- Key decisions made and any approved exceptions
- Next month’s planned reviews and audits
What are the leading and lagging indicators to track?
Leading:
- Percentage of AI use cases in the inventory with an owner and tier assigned
- Percentage of pilots that pass the first gate on the first attempt
- Time from idea to proof of concept approval
Lagging:
- Number of material incidents and time to contain
- Percentage of high-tier uses with overdue re-tests
- Exceptions granted and time to close them
A one-page charter should name the sponsor and chair, state the purpose and scope, set decision rights and quorum, define meeting cadence and list the core artefacts it owns.

Frequently asked questions
Is governance going to slow us down?
Good governance speeds you up. It removes ambiguity with clear gates and default controls by risk tier. Teams know what to prepare and who approves what, so ideas move faster.
Do we need a named Chief AI Officer?
You need named accountability and decision rights. Some firms appoint a Chief AI Officer. Others ask an existing executive to sponsor AI and chair the governance committee. What matters is a visible owner and empowered product and platform leads.
Can we automate governance?
Parts of it, yes. You can automate inventory collection, access checks, logging, policy enforcement and some testing. Risk classification, exception handling and material approvals still need judgement. Treat automation as a control, not as the decision-maker.
How does CAPIO PRO approach AI governance?
We build governance that is proportionate, embedded and designed for handover.
- Readiness baseline, three to four weeks: A short assessment across seven pillars, leading to an AI Health Score and a clear starting point. Why: decisions move faster when leaders share the same baseline.
- Governance setup: Committee charter, responsible AI guardrails, human-in-the-loop rules, impact assessments where required, an AI inventory and risk register, and board reporting. Why: core artefacts with owners keep governance lightweight and real.
- Pilot gates: Proof of concept, then minimum viable product, then an enterprise solution owned by your team. Typical timings are two to six weeks for a proof, one to three months for an MVP. Why: gates tie governance to delivery and scale what works.
- AI operating system: An embedded programme over six to twelve months with a monthly steering committee that has decision rights. Why: cadence and ownership make governance stick.
- Fractional leadership: A fractional Chief AI Officer role that establishes the baseline, governance and first decisions in the first ninety days and plans its own exit. Why: leadership capacity without long-term dependency.
- Handover by design: We design the engagement for handover, not reliance. Why: your teams should own AI as a capability.
Where to start
Pick two live use cases and run them through the gates while you set the inventory, rubric and charter. Keep the artefacts short, make the approvals real and improve each month.


