← Back to blog
Artificial intelligenceGovernanceSMEIT project management

AI governance for SMEs: build a committee that actually decides

Published on November 19, 2026•by Pierre Coulanges•8 min read
AI governance for SMEs: build a committee that actually decides
Photo: Brett Jordan / Unsplash

On November 18, 2026, France Num devoted a conference to a practical question: how can companies turn AI governance principles into operational decisions? Data protection, accountability and cost control must now lead to approvals, restrictions and stop decisions—not only another policy document. This playbook describes the lightweight committee needed to do that. France Num’s conference on moving from AI governance principles to implementation provides the timely angle; the underlying issue is day-to-day management of AI use cases.

For the broader view of use cases, costs and public support, read our article on what works, fails and receives public funding in SME AI projects. This article addresses the next missing component: the mechanism that lets management approve, test, scale or stop each use case.

The starting point

Requests usually arrive through different channels. Sales wants meeting summaries, HR wants help reviewing applications, customer service is testing an assistant, and an employee may already be uploading internal documents to a public AI service. IT, the DPO and information security are then involved after the tool has been selected or used.

The target is to establish, within thirty calendar days, a decision channel capable of answering a complete request within ten working days. These are D1 project-management targets, not regulatory deadlines. They translate the “Govern, Map, Measure, Manage” functions of the NIST AI Risk Management Framework Playbook into a workable SME cadence: assign accountability, understand context, measure risks and make a decision.

The committee must not become another discussion forum. Each case should result in a recorded decision, a named owner, operating conditions and a review date. The AI policy defines common rules; the AI use register lists systems; the committee decides what happens to them.

Prerequisites

Prepare the following before the first meeting:

  • An executive sponsor who can accept risk, reject a project, approve spending or order an AI use to stop. IT cannot carry this accountability alone.
  • An AI portfolio coordinator, such as an IT project manager, transformation manager or shared CIO. This person collects requests, prepares decisions and monitors actions.
  • Experts involved according to the case: the process owner, DPO when personal data is processed, security lead for sensitive data and system connections, and procurement or legal teams for contracts.
  • A list of existing and proposed uses, even if incomplete. If none exists, follow our step-by-step AI use register playbook.
  • Contracts and settings for tools already in use, including data location, conversation retention, provider reuse of inputs, subprocessors and deletion options.
  • The internal cost of the affected process, including human time, loaded hourly cost, rework and software expenditure. The CNIL guide to selecting a generative AI solution recommends including internal, external and recurring costs rather than looking only at subscription fees.

The step-by-step process

1. Give the committee decision-making authority

Who does what: the executive appoints the sponsor, coordinator and consulted functions. The mandate states who may approve a test, commit funds, accept conditions or stop a use case.

Realistic duration: two hours. This is a D1 scoping estimate for translating the “Govern” function of the NIST AI RMF Playbook into a lean organisation.

Deliverable: a concise mandate defining scope, decision rights, escalation rules and the participants required for different data categories.

Add a role-based AI literacy matrix. Article 4 of the AI Act has required providers and deployers to support AI literacy since February 2, 2025, with supervision and enforcement rules applying since August 2, 2026, according to the European Commission’s AI literacy questions and answers. The objective is not identical training for everyone: document what the sponsor, administrator, user and human reviewer must know.

2. Create a single intake form

Who does what: the coordinator designs the form with a process owner, IT and the DPO. Employees should not request “a chatbot licence”; they should describe the task, expected decision and required data.

Realistic duration: three to five working days. This D1 estimate covers design, approval and launch, building on the inventory method in our AI use register playbook.

Deliverable: an intake record covering:

  • the business process and accountable owner;
  • users and people affected by the output;
  • data entered, generated and retained;
  • required human validation;
  • proposed tool, model and integrations;
  • fixed, variable and internal costs;
  • the measurable expected outcome;
  • the fallback method if the AI or provider is unavailable.

Power Automate, n8n or Make can create the case, request missing information and notify the appropriate reviewers.

3. Triage requests before the committee meeting

Who does what: the coordinator checks completeness. The DPO joins when personal data is involved; the security lead joins when the request includes confidential information, system access or automated actions.

Realistic duration: half a day to define the grid, followed by thirty minutes per request. These are D1 operating estimates. The criteria reflect the checks recommended in the CNIL questions and answers on using generative AI systems: contractual roles, security, international transfers, retention and possible reuse of usage data.

Deliverable: a triage record assigning the request to a clear route: approval with conditions, controlled experiment, enhanced assessment or immediate rejection.

Immediate rejection should remain possible when the purpose is unclear, necessary data cannot lawfully be used, no one can review the output or a non-AI method already meets the need with less risk.

4. Run a bounded business test

Who does what: the process owner provides test cases and the current baseline; the project manager defines the protocol; the integrator configures the solution; users record errors, corrections and human interventions.

Realistic duration: four to six weeks. This D1 estimate allows several real operating cycles without turning the experiment into a permanent pilot. It follows the method described in Profitable AI project: validating a first workflow in six steps.

Deliverable: a test protocol containing the baseline, acceptance criteria, permitted data, required logs, spending ceiling, validation owner and shutdown procedure.

Do not measure only whether the output looks convincing. Measure human review time, rework, traceability, variable costs and the ability to return to the previous process.

5. Make and record the decision

Who does what: the coordinator presents the evidence, the process owner confirms the expected benefit, specialists state their conditions and the sponsor decides. A vendor demonstration is not evidence that the business protocol succeeded.

Realistic duration: ninety minutes per month, with simple cases handled asynchronously. This is a recommended D1 cadence based on the Manage function of the NIST AI RMF Playbook, which calls for an explicit determination of whether an AI system achieves its intended purpose and whether deployment should proceed.

Deliverable: a dated decision stating authorised uses, restrictions, accountable owner, metrics, approved expenditure, next review and the reasons for rejection or shutdown.

6. Turn approval into an operable system

Who does what: the project manager coordinates the process team, IT, DPO, security and integrator. The process owner accepts the final workflow and owns its operational results.

Realistic duration: two to four weeks for a contained workflow, followed by quarterly reviews. This D1 estimate assumes the test produced usable acceptance criteria and follows the approach described in moving from isolated AI tests to a business workflow.

Deliverable: an operating file containing the procedure, access rights, monitoring, human validation, incident handling, cost tracking, AI literacy evidence and exit plan.

How to measure whether it works

Use indicators that can be calculated directly from the intake and decision records:

  • Decision lead time = decision date minus the date on which the complete request was received. This measures the committee, not the requester’s drafting speed.
  • First-pass decision rate = cases decided during their first review divided by cases reviewed. A decline usually indicates missing information or unclear criteria.
  • Test-to-decision rate = completed tests resulting in scaled deployment, restricted use or a documented stop decision divided by completed tests. A justified shutdown is a useful outcome.
  • Cost per retained use case = internal committee time, test expenditure and operating costs divided by the uses still authorised at the next review.

Display these indicators in a short dashboard containing the request, owner, status, committed cost, primary risk and next decision. The dashboard is not intended to rank teams; it should expose stalled cases and uses that continue without justification.

Mistakes that cause the project to fail

  • Turning the committee into a demonstration club. Every session must end with a decision, a precisely assigned information request or a shutdown.
  • Applying a full assessment to every request. Without proportionate triage, the DPO, security team and executives become bottlenecks.
  • Confusing policy with approval. An AI policy defines shared rules but does not approve the data, provider or workflow for a specific use.
  • Starting with the tool. A request phrased as “deploy this assistant” prevents comparison with conventional automation, an existing software feature or removal of the task.
  • Allowing experiments to continue indefinitely. Without acceptance criteria, a spending ceiling and an owner, a pilot may consume resources without reaching operations.
  • Scaling without reversibility. Document the manual fallback, data export, access removal and supplier replacement process before making operations dependent on AI.
  • Ignoring individual AI use. A ban without an accessible approval channel drives usage outside the register. Submitting a request must be easier than bypassing the process.

Getting support

D1 Consulting can define the committee mandate, decision matrix, use-case portfolio, test protocols and delivery governance through our IT Project Management and Shared CIO service.

We can also build the intake channel, approval workflows, notifications and controlled agents with Power Automate, n8n or Make through our Process Automation & Optimisation service. Each decision then creates an operational audit trail without adding disproportionate manual administration.

👉 Book your free 30-minute diagnostic: we will identify the first AI decision workflow to structure and the deliverables required to put it into operation.

An automation or digital transformation project?

Let's discuss your challenges and see how we can support you.

Contact us