Your AI Policy Is Losing to a Browser Tab
Photo by Photo by Vitaly Gariev on Unsplash on Unsplash
Every company that has tried to stamp out shadow AI has run the same experiment. Legal drafts a policy. IT blocks a couple of domains. There is a Slack announcement, a lunch-and-learn, and a sincere promise that an approved tool is coming. Six weeks later a product manager is pasting a redlined contract into a free chatbot on her phone, because the sanctioned option takes nine clicks and a manager's sign-off.
That is not a discipline problem. It is a product problem, and it gets fixed the way product problems get fixed: by making the safe path also be the fast path.
The Approved Tool Is Almost Always the Slower One
Run a stopwatch test in your own company. Have someone time the sanctioned workflow from intent to answer. Then time what people actually do. The gap is usually measured in minutes, and sometimes in the number of people who have to be involved before anything happens.
Nobody chooses the slower tool when the deadline is Thursday. If your approved assistant requires a ticket, a data classification step, and a training module that has not been updated since the last model release, the outcome is already decided. People will route around it. They will do that quietly, because the alternative is a conversation about why they are being slow.
The uncomfortable implication: every shadow tool in your company is a small, unflattering product review of the thing you shipped.
Figure Out What People Are Actually Trying to Do
Before you write another policy, go look at the workflows. Not the org chart of tools — the workflows. In most US companies right now it is a short list:
- Summarizing a long document before a customer call
- Drafting a first pass at an email that has been written four hundred times before
- Turning messy meeting notes into tickets
- Explaining an error message in a legacy system nobody wants to touch
- Sanitizing a spreadsheet export into something a human can read
None of those tasks require a company-wide platform decision. They require a good answer in under a minute. If your governance model treats each of them as a procurement event, employees will treat your governance model as an obstacle to be routed around.
Three Questions That Beat a Ban List
When an unsanctioned tool shows up, resist the urge to send a reminder email. Ask three things instead:
What was the task? Get the actual sentence someone typed, or as close as they can reconstruct it. This tells you what to build.
Where did the data come from? If the answer is a customer contract or a support ticket, you have found a boundary problem worth fixing this quarter, not a compliance problem worth announcing.
What was wrong with our tool? Ask it plainly, and do not flinch. The answers are the cheapest product research you will get all year.
Do this for a month and a pattern emerges. It is almost never one giant gap. It is a handful of small frictions that add up to a workaround.
Ship the Fast Lane, Then Fence It
The practical move is to build a sanctioned lane that is measurably faster than the shadow one, then put the guardrails inside the lane rather than at the door. In practice that means:
- One click from the tools people already live in — the browser, the doc, the ticket queue
- No approval step for the common case; approval only for the genuinely sensitive one
- Clear, visible labeling of what data class the session is handling
- Automatic redaction of account numbers, SSNs, and contract terms when the document is classified
- A visible history so employees can see what happened to their prompt
That last item matters more than it sounds. People trust tools they can audit. A black box with a compliance badge gets bypassed the first time it gives an answer that feels wrong.
Instrument the Boundaries, Not the Doors
The instinct after an incident is to lock everything down. It works for about a quarter and then the workarounds come back, only now they are invisible. A better approach is to instrument the boundary: log what data classes are leaving, alert on spikes, and treat every alert as a product signal rather than a disciplinary file.
If your data-loss tooling fires ten thousand times a week and nobody reads it, you do not have controls. You have a very expensive log file. Tune it until the signal means something, even if that means starting with three rules that actually get investigated.
The Metric That Tells You It's Working
Track two numbers side by side: sanctioned AI usage inside your work tools, and the volume of traffic hitting blocked or unknown AI domains. If the first goes up and the second goes down, the product is winning. If the first goes up and the second stays flat, you have built a nice tool that nobody uses and the workaround is still alive.
Governance by memo produces compliance theater. Governance by product produces adoption and, incidentally, a much smaller attack surface. Employees are not looking for a fight over which chatbot they get to use. They are looking for the fastest way to finish the task in front of them. Give them that, with the guardrails already baked in, and the policy document becomes almost redundant.