AI Governance for Operations

Governance is an operations problem

Most AI governance work gets handed to legal, security, or IT, and most of it comes back as a document nobody reads. That's not a failure of effort. It's a failure of placement.

Early in my career, every purchase over $5,000 needed a sourcing justification form. Filled out by hand, routed, approved, before the money moved. It slowed people down. It's also the reason we could tell a board, or an auditor, exactly how a decision got made and who signed off on it. That's what governance actually is. Not a document. A control built into the moment the decision gets made. AI hasn't changed that. It's just changed who, or what, is making the call.

I spent 25+ years running operations, including $400M+ in indirect spend and a 10,000-contractor contingent workforce program at NXP. Every governance model that held up in that world had the same three parts: clear expectations, measurable outcomes, continuous feedback. AI doesn't need a new philosophy. It needs the one you already use for people, pointed at a new kind of worker.

Start with the Shadow AI inventory

Shadow AI is the unsanctioned use of AI tools by people trying to get their work done faster. In most mid-market operations it's already widespread and already touching customer data, and whoever's accountable for the risk usually can't see any of it. Leaders routinely underestimate it by an order of magnitude, because the honest answer is unsafe to give.

So make it safe. Run an amnesty inventory: a two-week window where anyone can report what they use and why, no discipline, no exceptions to that promise. Ask three questions and stop there. What tool do you use? What task does it do for you? What information do you paste into it?

The third answer is the one that matters. Tool names turn over every quarter. Data exposure is the risk that doesn't go away, and it's the only part of this a board is actually going to ask you about.

Sort the work into three lanes

Once you can see the usage, tier it by consequence, not by technology. A blanket policy fails for the obvious reason: it treats drafting an internal summary the same as scoring a supplier or screening a candidate, and those aren't the same risk.

Green is internal, low stakes, easy to reverse. Drafting, summarizing, reformatting, brainstorming. Sanction a tool, publish a short data rule, get out of the way. Speed is the whole point here.

Amber is customer facing or externally visible, but still reversible. Proposal language, supplier correspondence, first-pass analysis. AI drafts it, a named human approves before it leaves the building, and that approval gets logged.

Red is where I get careful, because red is regulated, contractual, or consequential to someone's livelihood. Hiring decisions, pricing commitments, contract terms, performance ratings. AI can prepare the inputs and structure them. A person makes the actual decision, every single time, and that line has to be designed in from day one, not discovered the hard way after something goes wrong.

Most organizations find that 70 to 80 percent of their real AI usage sits in green. That's the number that lets you move fast in public, because you can point to exactly where you didn't.

The Governor: manage Digital Associates like people

A Digital Associate is an AI worker with a defined job, a defined boundary, and a named human manager. A copilot speeds an expert up. A loose agent runs on its own with nobody watching. A Digital Associate has a manager. That's the entire difference.

The Governor is the layer that does that managing. In practice it's four short artifacts per associate, each one short enough to fit on a page.

A job description first, and write the never list before the yes list. What this associate handles, and what it explicitly never touches. That never list is your red line.

Then an escalation rule. The conditions where it stops and asks a person instead of guessing. Low confidence, an unusual amount, a counterparty it hasn't seen before, anything outside the pattern it was built for.

A performance standard comes next, the accuracy, cycle time, and exception rate you actually expect, agreed before launch, not argued about after the first bad week.

And one named manager. A single human who owns its output the way they'd own a direct report's work. Not a committee. Committees don't coach.

Then you run the exact cycle you already run with people. Review performance on a schedule. If an associate's underperforming, coach it, retrain it, or retire it. Your managers already know how to do this part. That's the whole reason adoption barely costs you anything.

The four controls that carry the weight

Governance frameworks tend to fail from bloat, so keep this to what actually matters.

Start with a data boundary: one rule, in plain language, about what never gets pasted into a general-purpose tool. Customer personal data, unreleased financials, contract terms, credentials. If your people can't recite it back to you, it's too long.

You also need a sanctioned path, meaning at least one approved tool that's actually better than the shortcut it's replacing. Take away an option without giving people something faster and you've just pushed Shadow AI further underground, not out of the building.

Anything sitting in amber or red needs an accountability record, the same way that old sourcing form did: a log of who approved what, and when. It's the difference between an incident becoming a conversation and an incident becoming an investigation.

Last is a review rhythm, a standing quarterly look at associate performance, exceptions, new tool requests. Governance without a calendar entry attached to it just quietly decays.

A 90-day sequence

The first thirty days are just looking. Run the amnesty inventory. Map where AI is already touching the work, and separately, map the low-value work your highest-paid people are still doing by hand. Those two maps overlap more than most leaders expect, and the overlap is your fastest payback.

From day 31 to 60, you sort and sanction. Assign every use you found to a lane. Publish the data boundary. Stand up one sanctioned tool for the green lane, and actually communicate it as the upgrade it is.

The last thirty days, pick one workflow and govern it properly, all the way through: job description, escalation rule, performance standard, named manager. One workflow done completely teaches an organization more than a policy that covers all of them on paper and none of them in practice.

After that, it's repetition. Add associates as the roster grows, and let the review rhythm carry the rest.

What to measure

Recovered time matters most: hours actually returned to your highest-compensated people, and where those hours went next. Watch the exception rate per associate too, and whether it's trending down. And keep an eye on sanctioned-tool adoption, because a low number there usually means Shadow AI never left. It just stopped answering the survey.

Stop measuring tool count, licenses issued, training completions. None of that tells you whether the work got better, or safer.

When outside help is worth it

Plenty of teams can run this whole thing internally. Bringing in outside help is worth paying for when what you actually need is a workflow-level inventory instead of a tool survey, a risk tiering you could defend to your own board, and a governance model built in weeks instead of quarters. The test is simple. If nobody in the building has both the operating authority over the workflow and enough AI fluency to argue with a vendor, you're going to end up buying the wrong thing.

That sourcing justification form is still the whole idea, just pointed at a different kind of worker. What are the things that only YOU can do? Everything else on that list is a candidate for a lane, a Governor, and a named manager.

Paula Taylor is CEO & Founder of AI Edge Strategy. She spent 25+ years in enterprise operations, including as Global Director of Indirect Procurement at NXP Semiconductors, where she managed $400M+ in indirect spend and a 10,000-contractor contingent workforce program.