Business communication is undergoing a quiet transformation. According to a Gallup report cited by writer Faisal Hoque, more than half of U.S. employees now use artificial intelligence at work, and the single most common application, named by 51% of those workers, is writing and editing. In practical terms, that means the emails, messages, and replies flowing through organizations every day are increasingly composed, shaped, or sent by machines. The question most organizations have not yet answered is: what happens when nobody is watching?
When AI Talks to AI Without Anyone Noticing
The scenario that illustrates this shift most clearly is not dramatic. It is mundane. Someone receives a professionally worded email, reads it, pastes it into a chatbot, and sends back the reply the chatbot drafts. On the surface, two humans are communicating. Underneath, two AI systems are exchanging text, with the humans functioning more as relay stations than as authors.
This is not a hypothetical. It is already happening. Tyler Cadwell, the founder of an Arizona glassware business, built an AI agent he describes as his “first AI employee,” one that triages his inbox and responds autonomously to supply chain issues. The agent acts on his behalf without requiring him to review each exchange. Multiply that pattern across a workforce, and the communication layer of an organization begins to operate at a level of abstraction that no existing policy was designed to govern.
The volume pressure driving this shift is real. Microsoft’s 2025 work trends report found that the average worker receives 117 emails and 153 Teams messages per day. At that scale, delegating some portion of communication to AI is not laziness. It is a rational response to an impossible inbox.
The Spectrum of Delegation, and Where It Goes Wrong
Not all AI-assisted communication is the same, and this is where most coverage of the topic misses the point. There is a meaningful difference between using AI to polish a draft you wrote, having AI draft something you review before sending, and deploying an agent that conducts exchanges you may never see at all. These represent different points on a delegation spectrum, and organizations are drifting along it without making deliberate choices.
An estate agent profiled by the Financial Times uses an AI tool to manage nine inboxes simultaneously. She acknowledges that sometimes she lets it “think for her.” That admission captures the core risk: not that delegation happened, but that it happened without a conscious decision being made. Chosen delegation comes with clarity about what has been handed over. Drifted delegation comes with none.
The consequences of unmanaged delegation can be concrete. A startup executive described in Bloomberg found that a household AI agent ran up repeated charges of $100 per hour before he noticed the billing alerts. The agent was following its instructions, more or less, but the controls that would have contained the damage lived nowhere in the system.
The insurance firm Allstate offers a counterexample worth examining. When it handed the drafting of roughly 50,000 claims correspondence messages per day to generative AI, the results were measurably better: clearer, less jargon-heavy, and more empathetic than what human representatives had been producing. Delegation, done deliberately and in the right context, can improve outcomes. The variable is not whether AI is involved. It is whether anyone has thought carefully about the terms.
Why Governance of Communication Is Now a Leadership Problem
The deeper issue is not technological. It is organizational. As AI takes on more of the communication workload, organizations face questions their policies were never designed to answer: which conversations can be delegated, what a machine is authorized to commit the organization to, and who is accountable for what gets said.
Some conversations, by their nature, cannot be delegated without destroying their value. Performance reviews are the clearest example. An appraisal carries weight because a manager actually evaluated someone’s year. Mentoring works because a person invested their time. Citi, JPMorgan, and Boston Consulting Group have all built AI tools that support drafting performance evaluations, as noted in a Harvard Business Review piece. The tools exist. The question is whether using them for certain conversations eliminates the very thing that made those conversations matter.
The practical governance challenge is not to restrict AI use. It is to make delegation visible and bounded. That means naming which conversations must remain fully human, establishing clear norms about where on the delegation spectrum any given exchange sits, and building controls that live outside the agent itself: spending caps, limited credentials, approval gates that route consequential decisions to a human before anything is finalized. It also means naming a specific person who is accountable for what an agent does, regardless of whether they reviewed each message.
In Short
More than half of U.S. workers already use AI for writing and communication, and AI agents are beginning to conduct exchanges autonomously on behalf of the people they represent. The risk is not that AI communicates. The risk is that organizations have not decided what AI is authorized to say, commit to, or decide. Governance of the communication layer is no longer an IT question. It is a leadership question, and most organizations are behind.
Based on reporting from Fast Company - Tech.