Your AI is emailing my AI—and nobody’s in charge

More than half of U.S. employees now use artificial intelligence at work, with 51% using it for writing and editing, according to Gallup, and AI agents are increasingly communicating on behalf of humans, as seen in a Bloomberg profile of Tyler Cadwell, founder of an Arizona glassware business, who uses an AI agent to triage his email and respond to supply chain problems. This trend raises governance questions, prompting experts to urge organizations to decide what must never be delegated, make delegation deliberate, and rebuild ownership for AI-mediated exchanges.

Not long ago, a colleague of mine received a courteous and professionally worded email. It informed him about the current state of a shared project and listed the next steps required and who was accountable for each. A perfectly ordinary email—except it was, from start to finish, the work of an AI https://www.fastcompany.com/section/artificial-intelligence agent that had been set up to act on behalf of the person it represented. My colleague’s experience is still relatively unusual, but it is also the canary in the coal mine. In May, Bloomberg profiled Tyler Cadwell, the founder of an Arizona glassware business who has built an AI agent he calls his “first AI employee.” This agent does things like “ triaging his email inbox and even responding on its own to supply chain problems https://www.bloomberg.com/news/articles/2026-05-11/why-claude-ai-agents-are-driving-record-mac-mini-demand .” And even when agents aren’t independently sending emails, people are certainly using AI’s help more often to write those emails. Last month, Gallup reported https://www.gallup.com/workplace/712736/organizational-adoption-jumps-six-points.aspx that more than half of U.S. employees now use artificial intelligence at work, and the single most common use, cited by 51% of them, is in writing and editing—or, in other words, communicating with each other. For example, that email my colleague received from an agent? He pasted it into his chatbot and sent the reply it drafted. On the surface, it looks like humans talking to humans. Underneath, however, AI is talking to AI. This is how business leaders need to respond. There is no way to put the email-writing AI genie back in its bottle. And I don’t think we should even want to. The machine version is sometimes simply better. When the insurance firm Allstate handed the drafting of its claims correspondence—roughly 50,000 messages a day—to generative AI, the machine-written emails were clearer, less jargon-laden, and more empathetic https://www.wsj.com/articles/turns-out-ai-is-more-empathetic-than-allstates-insurance-reps-cf5f7c98 than the ones its human reps had been sending. Moreover, delegating to AI is increasingly the only rational response left to employees drowning in electronic communications. Microsoft’s 2025 report on work trends found that the average worker received 117 emails and 153 Teams messages a day https://www.microsoft.com/en-us/worklab/work-trend-index/breaking-down-infinite-workday . It makes sense—and may even make workers more productive https://www.fastcompany.com/section/productivity —to delegate some of this load to AI. But delegated conversation is not the same thing as a free-for-all. In fact, as the use of AI for interpersonal communications rises, governing this conversation becomes increasingly important. The rise of AI makes organizations confront questions their policies never anticipated: Which conversations may be delegated? What may a machine commit us to? Who owns what gets said? Three pillars matter most for organizations to meet this challenge: Decide what must never be delegated, make all other delegation deliberate, and rebuild ownership for the exchanges you hand over entirely. Video killed the radio star, and AI is killing the performance review. An ex-Dropbox manager told Axios in 2024 that she used ChatGPT to write her appraisals https://www.axios.com/2024/02/12/chatgpt-human-resources-performance-reviews . More recently, The Wall Street Journal reported in February that more and more managers were using AI to do their performance reviews https://www.wsj.com/business/c-suite/job-performance-reviews-ai-326cb4eb . This is not merely individual initiative. It is also organizationally driven: As discussed in a recent piece in the Harvard Business Review , Citi, JPMorgan, and Boston Consulting Group have all built AI tools that support drafting performance evaluations https://hbr.org/2026/05/gen-ai-could-fix-performance-reviews-or-make-them-even-worse . I’ve argued before https://www.fastcompany.com/91512817/why-some-decisions-need-human-room-even-when-ai-gets-right that some leadership activities—the performance review being one of them—depend on full human engagement for their value. An appraisal matters because your manager actually weighed your year. Bad news lands humanely because someone chose to deliver it. Mentoring works because a person you respect spent their scarcest resource on you. And delegating such activities to AI doesn’t make your workflow more efficient; it destroys the value the work was supposed to generate. So every organization needs to draw its human line: Name the conversations whose value depends on the human element, and make sure that they remain 100% human. Everything else may be delegated, so long as you do it deliberately. Which is where the second pillar comes in. Delegation runs along a spectrum. At one end sits AI-assisted work: You create, and the machine polishes. Further along, AI-delegated: The machine drafts, and you skim and send. At the far end, AI-represented: Your agent conducts the exchange, and you may never see it at all. It is easy to drift along this spectrum without realizing it. For example, an estate agent profiled by the Financial Times https://www.ft.com/content/216edf7e-6fe8-47e6-96f2-bf2d59bd560a? uses an AI tool to run nine inboxes. This saves her hours every week; but “sometimes,” she admits, “I am guilty of letting it think for me.” The problem is not that the AI thinks for her—that’s simply what delegation is, and delegation is often the right call. The problem is that it happens without a decision being made. Chosen delegation comes with a handover: You know what you’ve given up, so you know how it needs to be managed. Drifted delegation comes with no handover at all. Multiply that across a workforce, and it becomes an organizational condition. When individuals aren’t quite connected to what they say, the organization no longer knows what is being said in its name or how much human judgment is behind it. For everything outside the human line, the task of governance is not to restrict delegation—it is to make it visible. Give your teams a simple norm: Know precisely where you are on the delegation spectrum, and make sure the organization knows it, too. The Bloomberg story tells the tale of a startup executive whose household agent ran amok, ringing up $100 charges hour after hour https://www.bloomberg.com/news/articles/2026-05-11/why-claude-ai-agents-are-driving-record-mac-mini-demand —he was saved only because he happened to notice the billing alerts in his inbox. The remedy is not to review every message an agent sends—that would defeat the purpose. It is instead to rebuild what the human safeguard used to provide. Specifically, that takes three things. First, a mandate: an explicit decision about what the agent may commit you to—a meeting, perhaps a delivery date, never a price. Second, containment: Do not rely on the agent obeying its instructions. Instructions are just more text to an agent; it follows them the way it does everything else—usually, not always. The controls that count live outside the agent, in the systems it touches: a payment card with a hard cap, credentials that open the calendar but not the contract folder, approval gates that route consequential exchanges to a human before anything is agreed. And third, an owner: a named person who answers for the agent regardless. The pillars above are the core principles that should inform governance of the new communication layer in organizations. Here are four moves that you can make right away to start translating those principles into actual governance in your organization. 1. Audit where AI is already talking to AI. Map the exchanges in your organization that are AI-drafted on both ends. The results might surprise you. 2. Draw your human line. Convene your leadership team and name three to five conversations whose value depends entirely on human presence. Declare them human-only, and explain why. 3. Label the modes. Pick one team, and have members tag their communications for a week: 100% human, assisted, delegated, or represented. Your real baseline is almost certainly further along the spectrum than you think. 4. Give one agent a mandate. For a single agent use case, write down what it may commit to autonomously, put the real controls outside the agent—spending caps, limited credentials, approval gates—and name the person who owns its output. This becomes the template you can iterate for everything after. The communication layer of work is being rebuilt in real time by vendors shipping inbox agents, by banks building review-writing tools, by a Realtor in Chester and an engraver in Arizona, one delegated exchange at a time. And the organizations that thrive will be neither those that ban delegated conversation nor those that surrender to it. They will be the ones that decide, deliberately, which conversations still require human presence, and on what terms machines may speak in the rest.