Your AI agent is already on the org chart

The hardest part of deploying an AI teammate is deciding what kind of teammate it is allowed to become.

Consider this hypothetical.

On Monday, a leadership team adds an AI agent to a busy project channel. Its job is straightforward: summarize discussions and update the project record. It will also flag blocked decisions and follow up on overdue work.

By Thursday, the agent is doing all four, but it has also quoted a tense exchange in its daily summary; treated an abandoned draft as the current plan; tagged a senior leader on a decision the team expected to resolve alone; driven one employee to stop thinking out loud; sent another into private messages; and prompted the manager to create an unofficial record the agent cannot see.

The agent did not fail a technical test. It changed the room.

This is what many leaders miss about agentic AI. An agent does not enter an organization as neutral software. The moment people change their behavior around it, the agent has a place in the social system.

It is already on the org chart. It has a role and a voice. It also has some rank and a reach across the work. The only question is whether anyone designed those things.

Every agent gets a role

A new field study from Google Research and Google DeepMind makes this problem unusually concrete. The researchers placed a persistent, proactive AI teammate across more than 20 teams inside one large technology company. Over five months, the system participated in 41,000 conversational turns and produced 11,000 responses. The research team then interviewed 17 people from 11 of those teams. The paper was published on September 24, 2026.

The agent collided with parts of work that rarely appear in a process map.

It treated stale documents as legitimate authority. It sometimes added to what participants called document pollution. The same friendly behavior felt supportive to one person and intrusive to another. People had to spend mental energy deciding when the agent understood the situation, when it had crossed a line, and whether correcting it was worth the effort. When people felt the agent had been imposed on them, willingness to use it fell.

The study is qualitative and narrow. It took place in one technology company. The interviews capture reported experience, not audited productivity or long-term performance. Seventeen participants cannot tell us how every workforce will respond.

But the mechanism deserves attention. The agent was not merely completing tasks. It entered relationships and redistributed attention. It changed who felt safe saying what in front of whom.

That is an organizational design issue.

Technical permissions do not settle social authority

Most implementation teams ask sensible permission questions.

Can the agent read this channel? Can it update this record? Can it contact a customer? Can it approve a transaction?

Those controls matter. They do not answer a different set of questions.

Should the agent speak for the team? When may it expose a disagreement? Whose version of events counts as authoritative? Can an employee contest an agent's interpretation without looking resistant to AI? Who decides that the agent has earned more independence?

A system prompt cannot resolve questions leadership has avoided.

The Google researchers found that teams often made sense of the agent by treating it like a new hire. They gave it low-stakes work and watched how it behaved. Its freedom expanded only after it demonstrated useful judgment. That was not anthropomorphism for its own sake. It was a practical way to make trust progressive instead of automatic.

The useful principle is simple: access can be provisioned in a day. Authority should be earned in public.

Work redesign now includes relationships

Two other reports published this week reinforce the point from different angles.

The Conference Board's September 24 framework for agentic AI and work redesign draws on sessions with more than 75 senior leaders. Its evidence also includes 12 interviews and four focus groups. More than 100 members used its workflow exercise. Its guidance begins with the desired outcome. Teams then redesign and allocate the work before deployment. The report also calls for one accountable human for every workflow.

Boston Consulting Group's September 24 study interviewed leaders at 50 AI front-runners across 10 countries. The firms were changing decision rights and team structures. Their performance systems were changing too. New roles included workflow designers and domain anchors, along with AI guardians and agent shepherds.

These are not controlled experiments, and the organizations studied are not representative of every employer. The evidence does not prove that copying their structures will produce better financial results.

My inference is narrower. Once an agent participates across a workflow, implementation cannot remain a software installation. Someone must design how the agent relates to human authority, expertise, consent, and accountability.

If nobody owns that work, employees will do it informally. Private channels and shadow records will appear. So will unwritten rules. The official workflow may look more automated while the real workflow becomes harder to see.

Run the meeting before you run the agent

Before placing an agent inside a shared workflow, bring together the people who own the outcome and the people who live with the process. Use one real case, not a polished demonstration.

Ask:

What may the agent observe, and what should remain outside its view?

When can it speak, decide, or act without invitation?

Which source outranks the others when records conflict?

How can a person correct, pause, or appeal its action?

What evidence would justify giving it more independence?

Who remains accountable when the agent performs exactly as instructed and the result is still wrong?

Then say the uncomfortable part out loud: people are allowed to have a different relationship with the agent.

One employee may welcome a persistent assistant. Another may experience the same behavior as surveillance. A team that cannot discuss that difference honestly is not ready to hide it inside an adoption metric.

That silence is data.

This does not mean every deployment needs months of deliberation. A low-risk agent can start in a bounded area where actions are reversible. Its work should be visible. Stopping it should be easy. Trust can expand as evidence accumulates.

What leaders should not do is confuse technical availability with organizational consent.

The job is to define the relationship

Return to the hypothetical project team.

The fix is not only a better summary prompt. Leadership must decide whether the agent is a recorder, a coordinator, an adviser, or an enforcer. It needs an owner. The team needs to know what the agent can expose, which records it should trust, how people can challenge it, and what would cause its authority to shrink.

Without those decisions, the agent will still acquire a role. Defaults will shape it, along with whoever has enough power to work around it.

That is the commercial risk. A technically successful agent can increase supervision work, push important conversations out of sight, and weaken the very coordination it was meant to improve.

It is also the opportunity. Leaders who treat the relational contract as operating infrastructure can use agents to clarify authority, make hidden work visible, and return attention to people instead of creating another system people must carry in silence.

If everyone responsible for an agent can clearly explain its purpose, boundaries, escalation path, and route of appeal, you may not need outside help. If the answers change depending on who is in the room, start with our free 15-question AI readiness read.

Use the free 15-question AI readiness read to surface the first disagreements. If an agent is about to enter a workflow that touches customers, money, employees, or public services, we can examine that workflow with you in a 30-minute diagnostic conversation.

The question is no longer whether AI can join the team.

It already has.

The leadership question is what kind of teammate you are willing to make it.