A colleague gets a polite, well organized email about a shared project. It turns out an AI agent wrote it, on behalf of a real person, without that person reading it first. This is no longer rare, and it is becoming the normal way work gets done.
Gallup found that more than half of US employees now use AI at work, and the single biggest use is writing and editing, which mostly means emails, reports, and messages to colleagues. Microsoft's own data shows why: the average worker now gets 117 emails and 153 chat messages a day. At that volume, handing some of it to a machine is not laziness, it is survival.
Some companies have already gone further than an employee quietly using a chatbot. Allstate now lets AI draft nearly all of its claims related emails, and reports the machine written messages are clearer and more caring than the ones its own staff used to send. JPMorgan built an internal AI tool for drafting performance reviews that reached more than 200,000 employees within eight months of launch.
Even small businesses are handing over full control. One Arizona shop owner built what he calls his first AI employee, an agent that reads his inbox and replies to supply chain problems on its own, with no review before it hits send.
What most business leaders have not clocked yet is that this is turning into infrastructure, not just a personal habit. Google has already released an open standard that lets AI agents from different companies find each other, prove who they are, and hand off tasks without a person in the loop. Once that becomes normal, an email between two companies could be two pieces of software talking, with humans only skimming the summary afterward.
The part everyone skips is who is responsible when this goes wrong. Legal experts already warn that letting an agent act on its own does not reduce a company's liability, it just makes that liability harder to see. An agent that overspends, promises a delivery date it cannot meet, or agrees to terms nobody approved is still the company's problem, even if no employee ever read the exchange.
The fix is not to ban AI from writing on your behalf. That fight is already lost, and the machine version is often better anyway. The real fix is deciding, on purpose, which conversations must stay entirely human, such as a real performance conversation or bad news delivered to a client, and putting hard limits on everything else: spending caps, limited system access, and a named person who owns what the agent says.
Most companies have none of this written down yet. They have adoption without any governance, and that gap is exactly what turns a helpful shortcut into a very expensive surprise.