The curious confinement of the AI colleague

Consider a company that recruits an unusually capable analyst, then confines the new hire to a small office. Questions are slipped under the door. Answers return by the same route. A manager carries them to the meeting and reads them aloud.

Colleagues have objections. The budget has changed; a customer has raised a concern; someone remembers that a similar proposal failed last year. The manager makes another trip. Eventually, after the meeting has ended, the analyst receives an account of the discussion in which they were unable to participate.

Most businesses would recognise this as an eccentric way to employ intelligence. In our dealings with AI, it can pass for progress.

The small office is a private chat window. The notes are prompts. The manager is anyone who has told a room full of colleagues, “I asked the agent, and this is what it said.” The technology may be novel; the arrangement has the unmistakable character of an internal mail service.

Private exchanges with AI have a useful place in research, drafting and preparation. But when the work becomes collective, the means of collaboration ought to change with it. An agent whose answer is being debated can have something useful to contribute to the debate itself.

The cost of the intermediary

The inconvenience is easy to see. The loss of context is less conspicuous. A person consulting an agent on behalf of a group must decide what to include in the question, which parts of the reply to present, and how to convey the objections that follow. Every exchange depends on that person’s selection and interpretation.

An answer can be entirely responsive to the prompt and poorly suited to the problem. The changed deadline was never mentioned. The customer’s objection seemed incidental. The failed experiment was known to a colleague who had no part in composing the question. These details often emerge during discussion, as people discover which assumptions they do not share.

Meanwhile, the group is left interpreting an absent contributor. Why did the agent recommend that course? What evidence did it use? Would the recommendation survive a different assumption? The person holding the answer is expected to speak for a system that could, in principle, be asked directly.

Adding more agents behind the door does not necessarily resolve this. A researcher, planner and critic may exchange messages and review one another’s work. The organisation has acquired a more elaborate advisory department, while retaining the same narrow opening through which its advice must pass. The back office has grown more sophisticated. The delivery arrangements remain curious.

Capability is only part of the question

Research gives substance to the claim that AI can make a useful contribution. In a study of 5,172 customer-support workers, access to an AI assistant increased issues resolved per hour by 15% on average, with the largest benefits accruing to less experienced workers. This was a specialised support setting, rather than a general test of office work. Even so, it showed a measurable benefit in an operating business.

A preregistered experiment involving 791 P&G professionals considered a question closer to teamwork. On product innovation challenges, individuals using AI produced work comparable in quality to two-person teams without it. AI also helped participants develop proposals that balanced technical and commercial perspectives, bridging some of the differences between their professional backgrounds.

Neither study tested Faceport or established that a speaking agent improves meetings. Their relevance is more specific: AI can supply useful assistance, and the way people work with it merits serious attention. Faceport’s proposition is that, where a task is being worked through in conversation, an agent should be able to participate while the discussion is still open.

As agents become more capable, the question of how their contributions reach a team becomes harder to separate from the question of what they can do. Improving the analyst is a considerable achievement. It is still worth reconsidering the locked door.

A place in the discussion

Faceport brings AI agents into live online meetings, where they can listen to the human discussion and respond by voice. Participants can put questions to an agent and hear its replies together. When several agents are present, people address the one they want by name.

This changes who can take part in examining an answer. The colleague who knows the customer can challenge an assumption. The person responsible for delivery can correct a date. Others can ask for evidence or pursue a point that the original questioner had not considered. The agent’s contribution becomes available for direct scrutiny by the people whose knowledge is needed to assess it.

Consider a hypothetical product meeting. An agent is asked to examine the assumptions behind a launch plan and identifies dependence on an integration being ready. The project lead explains that the delivery date moved that morning. The group can then ask the agent to reconsider the sequence in light of the change, while the people responsible for the work are present to assess its reply.

An agent may still miss an important detail. But the person who knows it can supply it directly, before the discussion moves on. There is less need to reconstruct the exchange later from memory or to consult the agent after a decision has effectively been made. A written recap remains useful. It need not be the agent’s first encounter with the discussion.

The terms of admission

Participation brings responsibilities for the people doing the inviting. Fluency can make an answer sound more authoritative than its evidence warrants. A seat at the table should confer no exemption from the questions put to anyone else.

The distinction matters. In an experiment involving 758 BCG consultants, AI users worked 25.1% faster on tasks within the model’s capabilities. On a task outside those capabilities, they were 19 percentage points less likely to reach the correct answer. The findings concern particular tasks and an earlier generation of AI, but illustrate why usefulness and reliability must be assessed separately.

An agent needs a defined role. Its assumptions should be made explicit and its sources checked. People with relevant knowledge must be able to challenge it, and people must remain accountable for the decision. Bringing an answer into a shared discussion creates an opportunity for scrutiny; the group still has to exercise judgment.

The same restraint applies to attendance. An agent should join because its contribution is relevant, and speak when that contribution is useful. The familiar meeting in which too many participants say too much has little need of artificial assistance.

Reconsidering the door

The prospect of more capable AI invites ambitious questions about the future of work. It also invites a rather practical one: are we arranging the work so that this capability can be used sensibly?

When a team needs an agent’s analysis, allowing its members to examine that analysis together is a reasonable place to start. They can supply missing information, test an explanation and consider alternatives before the discussion moves on. The agent contributes; the team retains responsibility for what follows.

That is the case behind Faceport’s AI agents: give useful intelligence a place in the conversation, alongside the people doing the work.

We have devoted considerable ingenuity to what happens behind the door. Opening it should not be the difficult part.