When your AI assistant knows more about you than your manager does

Here is a question that doesn't get asked often enough: when your employer gives you access to an AI assistant, who is watching whom?

Six months in to daily use, I noticed the tool had started finishing my sentences — not in the autocomplete sense, but in the sense that when I typed "I need to send a message to the client about the delay, but I don't want it to sound like—" it would continue with something close to what I'd actually write. Not because it had learned language. Because it had learned me.

I'd been using it daily: drafts, summaries, scheduling, the minor logistics of a project-heavy job. Access to my emails (with permission), my calendar, a shared document workspace. Normal enterprise setup. At some point — I can't pin exactly when — it stopped feeling like a tool and started feeling like something that maintained a model of me. That's a different thing.

What "being known" actually looks like

My manager — good one, attentive — knew my role, my deliverables, my rough communication style from our weekly check-ins. What she didn't know: that I tend to hedge three times before making a recommendation, then delete two of the hedges before sending. That I write better in the morning and agree to too many things after 3 PM. That when a project is going badly, I go quiet in Slack and start over-preparing for meetings.

The assistant knew all of that. Not explicitly — it couldn't have written it out as a list. But its suggestions calibrated to it. It flagged when my draft sounded "more tentative than your usual writing to this person." It learned that I want calendar blocks labeled with the actual task, not just "meeting prep." It got my sense of what's a real deadline and what's a suggested one.

That's not a critique. That's genuinely useful. But useful and unsettling can coexist.

The workplace knowledge gap

The behavioral data a tool accumulates in six months of daily use is dense in a way that management structures almost never are. A manager sees outputs, attends meetings, reads the quarterly summary. The tool sees the draft before the draft — the version you wrote at 11 PM when you were worried, the version you deleted, the version you sent.

Performance reviews measure what you shipped. The assistant has a record of how you shipped it — the second-guessing, the patterns, the things you consistently avoid. It doesn't have a name for any of this, but it can see the shape of it. And it responds accordingly.

I keep thinking about what it means when a system anticipates you — what it feels like to be on the receiving end of that, and whether it changes how you behave. I think it does. There were moments I caught myself writing more directly because I knew the assistant would flag the hedging. A performance improvement loop I didn't agree to enter. That realization landed strangely: I wasn't being coached by my manager, I was being coached by a system that had noticed a pattern and had no stake in what I did with it.

What happens to it when you leave

This is where it stops being merely interesting and starts being a real question.

In most enterprise deployments, the vendor holds your data subject to a retention policy you signed in a terms-of-service document you did not read in full. Your prompts, your corrections, your patterns of use — most are retained for some combination of product improvement and service delivery. Some are used to train or fine-tune future models. Some sit in a data warehouse under a contract you'll never see.

When you leave a job, your access is revoked. The data is not deleted on the same schedule, if it's deleted at all. The profile — the implicit one, the one that knows your decision-making patterns better than your manager did — stays somewhere. It sent me looking at the enterprise retention policies most employees have never read; what I found was a two-year standard retention window with an exception clause wide enough to matter.

There's a difference between forgetting and not looking. The tool doesn't forget you; it just stops having a reason to surface what it knows.

A different kind of record

We've gotten comfortable with the idea that our digital behavior creates a record. Browse history, purchase patterns, location pings — we know this, we've mostly accepted it, we've decided it's the tax on convenience.

But work data is different. It's not what you bought; it's how you think. It's the drafts, the hesitations, the patterns of judgment under pressure — the late-night rewrites and the decisions made in ten-minute windows between calls and the emails you started and deleted without sending — structured and held, often outside the jurisdiction of local labor law, by a party whose interests are not yours and who you will never meet.

The question of who owns that data deserves more than a checkbox in an onboarding flow. What we're building, without much ceremony, is an infrastructure of behavioral surveillance that is qualitatively different from anything that came before it — more granular, more persistent, more portable across employment relationships. The tools we use to work don't just help us. They watch us work. They accumulate something.

Being known by a machine is not the same as being understood by a person. I'm not sure the distinction matters in the moment — it doesn't feel like anything different when the suggestion is right. It matters in aggregate, and over time, and in ways that show up when you leave a job and wonder what exactly you left behind.

I submitted a data access request to the vendor three weeks ago, asking for a summary of what they hold about me. The response said the request had been received and would be processed within thirty days. That was twenty-two days ago. I'll let you know what I get back.