Do Not Make Them Use Your AI

A chief operating officer put a slide up with two numbers on it. Seat licenses issued: four thousand. Weekly active users: three hundred and ninety. She wanted to know how to close the gap, and the room treated it as an adoption problem.

It is not an adoption problem. Her people use AI all day. They are just not using hers.

Contents

1. The mandate is the only part that is not working
2. Forcing the company AI is deskilling by policy
3. What you actually own is the system of action
4. Six objections, taken seriously
5. This is a recruiting instrument, not a perk
6. Where I am probably wrong

01The mandate is the only part that is not working

Start with the thing that is already true and has been for years. In Microsoft and LinkedIn’s 2024 Work Trend Index, 78 percent of AI users said they were bringing their own AI tools to work. Eighty percent at small and medium-sized companies. That was a global survey of 31,000 knowledge workers, and it was two years ago.

Nothing since has reversed it. A Resume Now survey of 1,020 employed American adults, fielded in May 2026, found 76 percent had used an AI tool they found and signed up for themselves to do work tasks. In the same survey, 41 percent said their employer had given them no tools, no training and no guidance, and 52 percent said their employer provides either nothing or only free public tools.

Put those two findings next to each other and the shape of the problem inverts. The employee did not wait. The employer did.

Figure 01

Bringing your own AI was never a generational quirk

• Gen Z85%
• Millennials78%
• Gen X76%
• Boomers and older73%
• All AI users78%

• By generation
• Overall

So the mandate is not creating behavior. It is relabeling behavior that already exists, and pushing the part it cannot see further out of view.

That is the first thing I tell a client, and it is usually the first argument. Because the mandate feels like a control. It is not one.

Littler’s 2026 employer survey, covering more than 300 United States C-suite executives, in-house counsel and HR leaders, found 68 percent now have a formal policy governing workplace AI use. In the same survey, only 55 percent have a formal review or approval process for AI tools, and only 54 percent restrict what information can be entered into them.

A policy is a document. A gate is a control. Most companies bought the document and called it the control.

If you have a policy and no gate, the mandate is doing one job: moving liability onto the employee. Every organization I have said that to has recognized it, and none of them enjoyed it.

02Forcing the company AI is deskilling by policy

I have written before that a job is a bundle of tasks and AI eats the top slice first: the hardest, most judgment-heavy, most interesting part goes to the machine, and what is left is thinner than what you hired for. That piece treated it as a market force arriving from outside.

A mandate is the same mechanism, arriving from inside, by decision.

When you require a strong practitioner to work in a weaker tool, you do not protect the top slice. You just change who removes it. Her judgment is still doing the hard part; it is now doing the hard part with worse instruments, slower, and without the accumulated context that made her fast. You have not reduced the risk. You have reduced her.

This is the difference I keep coming back to. Protecting the business and confiscating the superpower are two different acts that use the same vocabulary.

1. Protecting the business

   What data are you putting in. Where does the work product live. What gets reviewed before it goes to a customer. Who is accountable for the output.

2. Confiscating the superpower

   Give us your prompts. Put your method in the handbook. Use only the approved model. Your workflow is now company property.

The first set builds trust. The second destroys it, and it destroys it specifically among the people you least want to lose, because they are the only ones who had anything to confiscate.

Standards are sacred. Procedures are negotiable. A company has every right to say what “finished” means, what “reviewed” means, and what “confidential” means. It has no business specifying the instrument.

03What you actually own is the system of action

Here is where the advice turns constructive, because “do not mandate” is only half of it, and the half most consultants stop at.

Every enterprise has spent thirty years building systems of record. The ERP, the CRM, the ledger, the HRIS, the ticketing system. These are stores of facts. They are not going anywhere and they were never the contested ground.

What is being built now, unevenly and mostly by accident, is the layer above them: the place where an intent becomes a permitted transaction. Call it the system of action. It is the surface where something outside the company asks to change something inside it, and gets an answer that is either a receipt or a refusal.

That layer is the only part of this you genuinely own. Not the model. Not the interface. Not the employee’s method.

Figure 02

Three layers, and only the middle one is yours

Layer 3 · Belongs to the person

The personal intelligence layer

The employee’s own model, memory, prompts, agents and accumulated method. Portable by nature. It arrives on the first day and leaves on the last.

↓ asks to act

Layer 2 · Belongs to the company

The system of action

Where intent becomes a permitted transaction. Identity brokered between two AIs, a published catalog of what can be done, and a receipt for every action taken.

• Identity
• Capability catalog
• Scopes
• Receipts
• Idempotency
• Revocation

↓ writes, reads, reverses

Layer 1 · Already built

Systems of record

ERP, CRM, ledger, HRIS, ticketing, document store. The facts. Unchanged by any of this except in who is allowed to touch them and how that touch is logged.

I have argued elsewhere that there is no universal rule about bringing your own AI, and I still think that. This is not a rule. It is a gate, and a gate is specific: it knows who is asking, what they may do, in what scope, and what happens if they are wrong.

The engineering work is unglamorous and finite.

• Publish a capability catalog
• Broker identity between two AIs
• Scope every grant
• Log a receipt per action
• Make writes idempotent
• Make revocation instant
• Name the sensitivity of the data
• Set a return date on borrowed context

Do that and the employee experiences intelligence while the enterprise experiences governed transactions. Those are the same event described from two sides, and the second description is the one that survives an audit.

It also gives you something the mandate never could: a log. Right now, the work your best people do with their own AI is invisible to you. Not because they are hiding it, but because there is no surface on which it could have been recorded. Build the surface and the shadow stops being a shadow.

04Six objections, taken seriously

I do not think the security and legal objections are stupid. I think they are real risks attached to the wrong remedy. Each one is asking for something a mandate cannot give it and a gate can.

Figure 03

What each objection is actually asking for

The objection	What it is actually about	What a mandate does	What a gate does
Data leaves the building	Which classes of data may cross which boundary, and for how long	Bans a tool. The data still leaves, through a personal account you cannot see	Classifies the data, scopes the grant, sets a return date, and records the crossing
Audit and e-discovery	Being able to reconstruct who did what, when, on what basis	Produces chat logs from the tool nobody uses for real work	Produces a receipt for every action against a record, regardless of which model asked
Regulated data	A statutory duty attached to specific fields and specific people	Applies one rule to all data, which means the rule is calibrated for the least sensitive case	Refuses at the field level. The regulated capability is simply not in the catalog for that identity
Model provider terms	Whether your content trains someone else’s model	Solves it for the sanctioned tool only, which is the tool carrying the least real work	Makes terms a condition of the grant. No acceptable terms, no capability, whoever the vendor is
Output quality	What “reviewed” means before something reaches a customer	Confuses the instrument with the standard, and enforces the instrument	Enforces the standard at the point of action, and is indifferent to how the work was produced
The leaver	What the departing employee keeps and what they must return	Assumes the method was company property, which is unenforceable and was never true	Revokes access in one action. Method walks out; your records and your capabilities do not

There is one honest version of the mandate, and it is worth saying. If your organization handles data where a single unlogged crossing is a reportable event, and you have not built the gate yet, then restricting tools is a legitimate stopgap. It is a stopgap. Treat it as a dated decision with an owner and a replacement plan, not as policy.

05This is a recruiting instrument, not a perk

The retention argument for permissive BYOAI is well worn by now. The recruiting argument is not, and it is the larger one.

A capable person evaluating your company in 2026 is running a calculation nobody puts in the job description. They are asking whether the way they work will survive contact with your organization. They have built something. They want to know if you will let them use it.

You can answer that question in the first interview, and almost nobody does. “Bring your own stack; here is what it can reach on day one” is a more persuasive sentence than any compensation band, because it tells the candidate the thing compensation cannot: that their capability will not be flattened to the median on arrival.

Autonomy has become part of compensation. It is not a soft benefit. It is a component of the offer with a price attached, and the companies that have not noticed are paying the difference in cash without getting the credit.

The reverse is also priced, and it is priced badly. There is a threshold where the gap between what a person can do alone and what the organization permits them to do gets wide enough that staying becomes irrational. That is not a resignation event. It is a slow, quiet, entirely rational drift, and the mandate is one of the reliable ways to cause it.

06Where I am probably wrong

Already true

Employees brought their own AI years ago and have not stopped. Most organizations that wrote a policy did not build a control. Three quarters of executives describe their own AI strategy as performative. None of this is a forecast; it is measured, and the measurements are cited below.

What has to happen for the advice to work

Someone has to own the gate as a product, not a project: a published capability catalog, identity brokered between a personal AI and an enterprise, a receipt per action, and revocation that takes seconds. If that work does not get funded, “do not mandate” collapses into “do nothing,” which is worse than a mandate.

Where I am probably wrong

The gate may cost more than the mandate returns for the next two or three years, particularly in heavily regulated industries where the catalog has to be justified field by field. If the build is slower and dearer than I think, the mandate wins on economics for a while and I lose the argument on cost rather than on principle. I also assume most capable employees want to bring their own instrument. A real fraction want a good tool handed to them and are right to want it. That is a supply obligation, not a contradiction, but if it turns out to be most of the workforce rather than some of it, this advice applies to a smaller group than I am claiming.

The instruction stands.

You do not want a company where everyone uses your AI. You want a company where everyone’s AI can reach your work, and leaves a receipt when it does.