Before buying AI tools, write the operating model

Start with a plain inventory row:

Workflow	Owner	Input data	Output user	Review gate	Next review
Support reply drafting	Head of Support	Ticket text, account notes	Support agents	Human approval before sending	30 days

If the team cannot fill in that row, it is not ready to buy another AI tool.

AI adoption usually becomes messy before it becomes strategic. One team buys a writing tool, another tests an internal assistant, a third uploads sensitive spreadsheets into a model, and no one can answer which use cases are approved, who can revoke access, or what happens when the output is wrong.

An operating model prevents that drift. It does not need to be heavy. It needs to define ownership, workflow inventory, risk classification, review gates, and the minimum evidence required before an AI workflow becomes normal business process.

The NIST AI Risk Management Framework is a useful reference because it treats AI risk as something organizations map, measure, manage, and govern over time NIST AI RMF.

Use this when

Use this when more than one team is experimenting with AI and leadership needs a shared way to approve, support, limit, and retire use cases.

It is especially useful for operations, enablement, technology, and security leaders who need practical coordination before a full governance program exists. The goal is not to centralize every prompt. The goal is to make sure repeated work has a named owner, a known data boundary, and a clear review path.

Skip it when

Do not build a company-wide operating model for one narrow experiment. If a small team is testing a low-risk workflow with public data, a project checklist may be enough.

Do not use this as a substitute for legal, security, procurement, or compliance review when the workflow affects regulated data, employees, customers, medical decisions, legal advice, financial eligibility, or safety-critical work. The operating model should route those cases to existing review functions, not replace them.

What to do

1. Name a single accountable owner for AI adoption.
2. Maintain a visible inventory of active and proposed AI workflows.
3. Classify each workflow by data sensitivity, business impact, reversibility, external exposure, and human review.
4. Define which workflows can launch locally and which need security, legal, procurement, compliance, or executive review.
5. Require evidence before scaling: example inputs, sample outputs, failure cases, reviewer notes, and operating limits.
6. Decide who can approve new tools, who can revoke access, and where prompt/output logs are stored.
7. Revisit approved workflows on a schedule because tools, policies, vendors, and risks change.

A useful first version can be a spreadsheet with a short policy beside it. Do not wait for a full governance platform if the current problem is that no one knows which workflows exist.

Watch the boring risks

The biggest early risk is not model quality. It is unclear responsibility. If no one owns the workflow, no one owns the data exposure, incorrect outputs, access control, vendor review, or rollback plan.

Before approving a workflow, check whether the tool stores prompts, uses customer data for training, exposes logs to administrators, supports access controls, and allows data deletion or export. Also check whether the workflow changes customer-facing text, employee records, financial analysis, security decisions, or source code.

The operating model should make review visible before the work feels routine. Once a workflow becomes a habit, teams stop noticing the assumptions inside it.

Other ways to handle it

For a very small company, use a lightweight approval checklist and a shared spreadsheet.

For a regulated organization, treat this operating model as intake only. The actual approval path should connect to established risk, privacy, legal, procurement, and security processes.

For a team that only needs experimentation rules, publish a short acceptable-use policy instead. That can be enough until workflows repeat across teams or touch sensitive data.

Try this next

Create a one-page AI workflow inventory with these fields: workflow name, owner, input data, output audience, tool, risk level, review gate, approval status, rollback owner, and next review date.