In August I wrote a post about an AI tool that had cleared procurement while the use case hadn't. Four people used it four different ways, and one of them wasn't using the approved tool at all: they'd pasted three pages of a requirements doc into a personal free-tier account to reformat a table.
This issue is about that fourth person. The August post asked what an approved tool is allowed to do; this one is about the tools that never went through approval, what a ban does to them, and what a PM can do about them on one project without owning the policy.
The Shift
What is shadow AI at work?
Shadow AI is AI used for work outside whatever the organization has approved: a personal account, a browser extension, a free tool found on a Tuesday because the deadline was Wednesday.
It starts as a timing problem. On most teams a yes on a new tool takes weeks and the task takes hours, so people use what gets the work done and plan to ask later.
A ban leaves the deadline where it was, so the AI use moves to a personal browser tab, and what changes is who can see it. Nobody is the villain here: the person in the free-tier account is trying to hit the date you gave them, and the person who wrote the ban is trying to keep client data where it belongs.
Seen through this newsletter's usual lens, shadow AI is a late decision: the work needed an answer this week, and the answer is booked for next quarter.
The System
What can a project manager do about shadow AI without owning the policy?
Make the AI use on your project visible, and send each unapproved use to the person who decides, with a date.
1. Ask without an audit. At the next retro, ask what the team uses AI for, with any tool, and say you're asking to find gaps. Without that framing you'll mostly hear about the approved tool.
2. Keep one list. One line per use: the task, the tool, and the type of information that goes in, on a page the team already opens.
3. Send each gap to whoever decides, with a date. For every use without an approved fit, send the person who approves tools (IT, procurement, a risk function, an AI committee) one question: this tool, this task, this type of information, yes or no by a date. The decision stays with them; your contribution is the date.
4. Say what's allowed meanwhile. Until the answer arrives, tell the team what to use for that task, an approved tool if one fits or the manual way, and the day the answer is due. A stop with no date sends the work straight back to the personal tab.
The Receipt
The receipt is the message that opens move 1. Send it in the team channel a day before the meeting, so people have time to think:
Quick ask before Thursday's retro: which AI tools are you using for work right now, approved or not, and for what? I want to see where our approved options fall short of the work, and get faster answers on the ones we're missing. Nobody gets in trouble for an answer here.
Promise only what's yours to promise. The no-blame line covers how you respond, and any rule your organization already has still applies.
The AI Assist
Run this on the list from move 2, in a tool you're already approved to use. Describe information by type, such as client names or pricing, and leave the data itself out.
Below is a list of how my team uses AI at work: one line per use, with the task, the tool, and the type of information that goes into it. After it is the list of tools my organization has approved. First, group the uses by task. For each use, say whether a tool on the approved list could do that task, using only the list I gave you; if you can't tell, mark it UNKNOWN. Then, for every use with no approved fit, draft a one-sentence question to the person who approves tools, in this form: "Can we use [tool] for [task], with [type of information] going in? We need a yes or no by [date]." Leave the date blank. Don't judge whether any use is safe or allowed; that decision belongs to the approver.
What comes back is a set of questions ready for move 3, each with a blank for the date.
The Move
This week, send the receipt message before your next retro, then turn the first unapproved use you hear about into one dated question to whoever approves tools.
Reply prompt: which task sends your team to an unapproved AI tool most often? Reply with the task. I'm collecting them, without names, for a future issue.
P.S. New here? Subscribe free and the Execution Signal Starter Pack lands in your inbox: get the Starter Pack.