Governance is not the opposite of shadow AI. Delay is.

You already did the thing.

Now there’s a form about it.

The case was urgent. The customer was waiting.

Your manager needed an answer before lunch.

So you opened the private AI chat, pasted the least you thought it needed, and got the useful sentence.

Then the official route arrived. Sponsor, purpose, data assessment, architecture review, next forum date.

You’re not hostile to controls because you needed a faster answer.

You’re tired. You’ve got a queue, a person waiting at the far end of it, messages that don’t get less urgent because a governance calendar exists.

And the uneasy sense that answering quickly may have created a different problem.

Accepting the tool wasn’t a theory of corporate rebellion.

It was you trying to finish the work.

This isn’t a story about a dangerous machine. It’s a story about a slow one.

The apparatus built to control outsourced work has become slow enough, in places, that people treat its safe route as unavailable.

Then your employer calls the workaround shadow AI.

As though the shadow were the first fact.

2008 — A change request needs three signatures. The work waits for the signatures to find each other.

2016 — The governance forum meets next Thursday. Your customer case is open today.

2024 — You solve it with a private AI chat before lunch. It feels like competence, then like concealment.

2026 — The incident review asks why you didn’t use the official route.

The apparatus had reasons

Vendor governance starts after award, when the agreement your employer signed has to become actual work.

In your first thirty days a mobilisation workshop turns that contract into a service boundary.

Named executives. A contact directory.

A reporting pack.

Registers for risk, issues, actions, decisions, demand, change.

Its first useful product is your calendar. Operational review, tactical service review, strategic steering session.

That’s not theatre. It’s how contradictory direction through side channels becomes visible.

And those forums do different jobs, which is the part people forget.

Your weekly operational review clears actions, service data, incidents, work in flight. Your monthly tactical review reconciles capacity, forecast demand, spend, risk, customer feedback, corrective plans.

Your quarterly strategic forum looks at outcomes, market signals, material change, relationship health. It isn’t a louder operational meeting.

During a recovery, the operational rhythm can go daily for a defined stabilization window. Then it stops, explicitly.

The separation matters because not every request is a crisis, and not every crisis is a change.

Demand intake captures your sponsor, desired outcome, affected users, timing, dependencies, security concern, and whether the thing even sits inside baseline scope.

Then your requests get ranked. Mandatory obligation. Customer or safety impact. Value. Urgency. Dependency. Effort.

Only the next four to eight weeks of credible capacity gets committed. The rest stays a visible backlog with another review date.

That’s governance at its best. Not denial, not endless consultation, not a slide about partnership.

It gives your ordinary request a place to enter, a decision it can reach, an owner who can say no, and an escalation when the answer has gone slow.

The apparatus exists because suppliers, systems, customers, data and money don’t become harmless just because you’re in a hurry.

Where it gets in the way

The failure starts when a forum can show activity without directing the service.

Your dashboard says ninety-five of a hundred items were on time. It omits the acceptance date, the buyer-caused delay, and the rework pushed back onto your customer.

It says 99.9% of workflow runs succeeded while you’re still waiting on a human approval, and a rejected case quietly becomes a different route.

Government digital guidance names four KPI categories: cost per transaction, user satisfaction, completion rate, digital take-up. Not because four boxes solve governance. Because a scorecard needs cost, outcome, experience and adoption together.

But the measurement that changes the argument is decision latency.

Business days from your complete request entering its agreed route to an authorized decision. Report the median and the ninetieth percentile separately, or don’t bother.

A median of three days with a ninetieth percentile of thirty-five days is not a ten-day service.

It’s a service where a minority of your cases has vanished into ambiguity, politics, or authority nobody can exercise.

That tail is where your private tab opens.

Action closure hides the same trap. It isn’t the share of actions marked green. It’s actions closed by the due date, with evidence, against actions due.

Watch your aging chart. A 95% closure rate coexists happily with five structural actions rolled forward for a year.

Sample a tenth of your closed actions each month and test the artifact, not the sentence somebody wrote beside it.

You can feel this failure before you can name it.

The form asks what you need. The service asks who is sponsoring it. The forum asks whether it belongs in the next planning cycle.

None of those questions is absurd on its own.

Together, arriving after the work has become urgent, they turn governance into proof that governance happened.

The committee sees its calendar.

You see a route that doesn’t arrive in time.

A service-level signal

None of which makes your workaround virtuous.

The private chat may receive data it shouldn’t. A departmental flow may run on an unreviewed credential. A process owner may not know it decides something material.

Security is right to see the risk. Your team is right to say it cleared a backlog central IT ignored. Both.

So deleting it on sight isn’t a serious diagnosis. Delete it today and you may interrupt payroll or customer service. Leave it and you keep an unknown risk alive.

The missing measure isn’t whether the committee wrote an AI policy.

It’s the time-to-deliver of the sanctioned alternative. How long until your legitimate need gets a safe, usable route, with templates, connector review, environments and support.

Microsoft says this out loud, which is more than most vendors manage. Overly cumbersome security prompts risky shadow-IT workarounds.

The point isn’t that security should surrender. It’s that latency is an operating cause of recurrence.

If the approved path can’t meet the moment where your work actually happens, prohibition measures a preference. Not a control.

That’s the correction. You’re not choosing speed over governance. You’re asking governance to become a usable operating system.

Speed has a boundary

Some controls can’t be turned into a friendly intake form and wished away.

Your workflow still needs a known data class, a controlled credential, declared decision authority, an accountable owner.

Good rules separate eligibility — whether your request may proceed — from routing — who decides the next question.

And when an upstream record is missing or contradictory, “otherwise approve” isn’t agility. For a material action the safer default is a visible exception queue with a service-level target on it.

Authority has to survive every handoff, too. A workflow can carry your identity and purpose, but authorization gets checked by the application protecting the resource. Not inferred from a token somebody obtained an hour ago.

An agent’s confident response doesn’t prove the outcome was authorized.

Which is why change control deserves better than contempt.

A real change describes your current and proposed state, then assesses service effect, price, schedule, security, data, readiness, dependencies, and what happens if you reject it.

Your material changes get a test environment, peer review, owner approval, a release record, monitoring, and a rollback plan.

That isn’t a demand that every urgent adjustment wait until next month.

It’s a demand that the route match the consequence.

And the procedure already has an emergency path. An authorized incident lead can direct containment for up to forty-eight hours under emergency authority, with the instruction, the cost assumption and the expiry logged.

Afterwards your tactical forum turns it into a corrective action or a formal change.

“Emergency” has to describe the event, not your impatience. But the existence of abuse doesn’t justify making every legitimate urgent case prove itself through a slow general route.

When authority is disputed, your escalation needs three states. Due. Overdue. Escalated.

Your service-level manager owns elapsed time. The policy owner owns a conflict about authority.

Keep those separate and operations staff can’t override a control just to make the queue look healthy.

What the review should ask

The incident review will still want to know what data moved, which credential was used, who owned the decision, and whether the result touched a customer, a payment, an access right, a regulated record.

Those aren’t gotchas. That’s how an organisation avoids turning a helpful answer into an invisible commitment.

An auditable workflow records your request, the source versions, the rule version, approvals and delegations, the command ID, the result, the retries, the exception handling and your override.

It shows when the event arrived and when the decision was made. Not just the final status.

But the review owes you one more question before it congratulates itself for finding your private tab.

What service level did the official route offer this need? And did it meet it?

Your employer built vendor governance to direct demand, change and escalation after award. It can use the same apparatus to learn from the route people took around it.

The private AI chat isn’t evidence that controls are obsolete.

It’s evidence, sometimes, that the controls became unreachable.

The correction is smaller and more demanding than the melodramatic version. Keep the boundary. Keep the evidence. Keep the person who can stop the harmful action.

Then measure how long it takes the safe route to arrive.