The product I wanted to build already exists, watches the wrong place, and its headline feature is the one your lawyer would strike. On 2026-10-06 I floated an idea: an AI chat for employers that notifies management when staff misuse it. Inappropriate content, client data going out, someone looking at things they shouldn’t. I did a day of homework before writing a line of code. The finding matters more than the product, so here it is.
It already exists, at every size
Microsoft already sells this to anyone on its top business plan. Its Purview compliance tooling inspects Copilot prompts and responses for jailbreak attempts, requests for sensitive data such as card numbers or salaries, and inappropriate content. Flagged items go to a reviewer queue with usernames pseudonymised by default, role-based access and an audit log.
Below that tier, a whole category of startups watches prompts across the browser, the desktop apps and the command line, including the tools your company never bought. Above it, the big data-loss-prevention vendors have bolted prompt inspection onto what they already sell. Zscaler’s 2026 threat report counted 410 million policy violations from ChatGPT alone in a year, up 99 per cent. Vendor-sourced, but the direction isn’t in doubt.
The demand is real. The gap isn’t in the software.
It watches the wrong place
The person you’d want to catch isn’t using your chat. They’re on personal ChatGPT on their phone. An employer-run assistant only sees the people who chose to use it, which is the group you worry about least. Every product that has won in this space sits at the browser, the device or the network, not inside its own chat window.
Snooping detection has the same problem. Whether a question is legitimate depends on the person’s role, what data they already have access to, and why they’re asking. The enterprise tools answer that with the identity and access logs they already hold. A standalone chat would have to rebuild all of it.
The reporting is the part that gets it switched off
A tool that pings a manager because someone asked about a severance letter lasts one quarter. Or a harassment complaint. Or a medical leave. The vendors learned this already and moved to pseudonymised review queues and in-the-moment nudges to the user, rather than a report to the boss.
There’s a legal side too. In Ontario, employers with 25 or more staff must have a written electronic-monitoring policy that says whether, how and why they monitor, and AI reading staff communications sits inside that. Ontario has no private-sector privacy statute and Canada has no AI act in force, so counsel works from principles, and “notify management” as the headline is the part they push back on. The federal regulators’ principles for organizations using generative AI are on the site already. This is information, not legal advice.
And the culture cost. An assistant framed as surveillance suppresses exactly the use you bought it for. The adoption pitch and the monitoring pitch pull against each other.
What a small firm should want instead
Flip it from catching bad employees to giving good ones a safe place to work. A 10 to 200 person firm with no security team and no top-tier Microsoft plan has nothing today. What it needs is a sanctioned AI workspace with the rules built in:
- Guardrails in the moment, not reports afterwards. Client names and personal information redacted before they leave. Tagged documents can’t be uploaded. The user gets a warning as they type, not a meeting next week.
- Aggregate, pseudonymised reporting for the owner: what kinds of data are flowing, which tools are in use. A named person is escalated only at a threshold you wrote down in advance, under the written policy Ontario already asks for.
- An audit log that answers a client’s vendor questionnaire or an insurer’s form. That’s the question small professional-services firms actually get asked.
- The written AI-use and electronic-monitoring policy as part of setup. The one-page template is most of it.
The model vendors’ own team plans keep adding admin controls, so the software part of this list gets cheaper every quarter. The policy and the setup are the work. That’s where I landed: not a product, a piece of the readiness work I already do.
The decision I’d ask you to make
Decide whether your AI tool is a safe place or a trap. You can’t have both, and the trap doesn’t work. Your staff are already using AI on their own. The sanctioned workspace brings them in. The monitoring tool pushes them further out, onto their phones, where nothing you buy will see them.