The Office Graph Needs Manners

The new workplace AI does not want to sit in a box anymore. It wants the calendar, the meeting notes, the spreadsheet, the chat thread, the directory, the deck someone abandoned after the offsite, and the half-finished planning document with three polite comments because the budget meeting was already tense. It wants to know who works with whom, which files keep resurfacing, and which conversation happened in the corridor-shaped part of the company now known as Teams, Slack, Gmail, or Drive.

This is not a bug in the category. It is the category maturing. Microsoft is giving developers Work IQ APIs so agents and apps can be grounded in organizational context: relationships, content, meetings, messages, permissions, and the strange sediment of daily collaboration. The redesigned Microsoft 365 Copilot is less a chatbot than a work surface stitched across search, notebooks, agents, and app-native actions. Google is making Workspace more proactive and more file-aware. OpenAI keeps moving ChatGPT nearer to connected files and spreadsheets.

The old promise was that AI would understand your prompt. The new promise is that AI will understand your workplace. That is a much stranger claim. A prompt is a thing you hand over on purpose. A workplace is a weather system. It contains formal authority and private doubt, stale documents, political camouflage, and the thing everyone knows but nobody has written down because writing it down would summon a meeting.

If agents are going to reason from that material, the question is not only whether they have permission to see it. The question is whether they have manners.

Manners are not etiquette theater. They are a way of handling other people's context without acting like access equals intimacy. A meeting transcript is not consent to be quoted in a performance review. A shared folder is not a worldview. A comment thread is not always a durable position. A calendar invite can mean responsibility, obligation, avoidance, ambition, or simply that somebody forgot to remove a name after the reorg.

An AI system with broad work context may not feel that voltage. It may only see a graph.

This is where the hype gets too hygienic. Product language treats context as nutrition for intelligence: more context, better answers, faster work. That sounds sensible until everyone starts receiving uncannily informed summaries from a system that appears to know too much and understand too little.

Context is not just fuel. It is social material.

A manager asking an assistant to summarize project risk is not simply querying documents. They may be pulling in dissent from a draft, uncertainty from a private note, or a compromise that only made sense inside one meeting's emotional weather. A sales lead may receive guidance built from support tickets, renewal chatter, legal edits, and a customer's annoyed aside. A product director may get a polished roadmap brief that flattens every unresolved disagreement into the calm voice of organizational memory.

Anyone who has worked in an actual company knows this voice. It is the voice that makes a mess sound decided.

That may be useful. It may also be dangerous. The workplace already has a habit of laundering uncertainty into slide format. AI can accelerate that habit by turning scattered context into confident prose before the people involved have finished arguing honestly.

The humane version of the office graph would not simply expose more context to more agents. It would make context behave differently depending on use. Some knowledge should summarize easily. Some should arrive with friction. Some should lead back to the original room. Some should decay. Some should remain searchable but not quotable.

Imagine an agent that prepares your Monday brief by reading every meeting, chat, file, and customer note it can reach, then presents the organization as if it were a single coherent creature with stable preferences and a fresh haircut. It tells you the top priorities, hidden blockers, and decisions that seem to be emerging. It may be right often enough to become indispensable. It may also train you to treat workplace life as something consumed in summary form.

There is a health cost hiding in that convenience. The modern workplace already asks people to live inside too many secondhand realities: dashboards, recaps, status posts, executive summaries, recordings watched at 1.5x speed. AI can make those realities richer and harder to distrust. A fluent assistant can feel like relief in the morning and like surveillance by lunch.

The problem is not that AI should avoid inference. Work runs on inference. The problem is that inference inside a workplace carries power. It changes who gets understood, who gets summarized, who gets interrupted, who gets blamed, and whose ambiguity survives contact with software.

Good interface design should treat that power as a first-class material. A workplace assistant should show the shape of its context, not just the shine of its answer. It should distinguish official sources from drafts, commitments from loose signals, public project knowledge from proximity noise. It should let people say, "do not use that kind of context for this kind of task," without needing a priestly tolerance for policy dropdowns.

Awkwardness is underrated in software because it does not look like efficiency. But a little awkwardness is how people notice that a boundary exists. The warning before turning a transcript into a performance narrative. The confirmation before moving from "summarize this project" to "infer what each person should do next." These are reminders that context belongs to people before it belongs to systems.

A better-grounded assistant can reduce repeated explanations, recover lost decisions, connect related work, and spare people the ritual of hunting for the one document whose title nobody remembers. The point is not to lock every useful signal behind suspicion.

The point is to remember that a workplace is not a database with feelings accidentally attached.

It is a social system that happens to produce data. Its files and meetings are traces of people trying to make things together under constraints: fatigue, ambition, care, politics, budget anxiety, pride, and the ordinary wish not to look foolish in front of colleagues. If AI tools are going to operate inside that system, they need more than retrieval. They need tact.

The next frontier of workplace AI may look, from a distance, like better grounding: more connectors, richer graphs, deeper personalization, smarter agents, cleaner handoffs. From inside the workday, it will feel like something more intimate: software learning the texture of the room. That can make work calmer. It can also make work feel haunted by a competent intern with access to everyone's notebook and no instinct for when to stop talking.

The office graph will become part of how organizations think. Before it does, we should teach it some manners: when to summarize, when to cite, when to ask, when to forget, when to keep quiet, and when to admit that the graph does not know what the room meant.