When AI Joins the Team

Anthropic have just introduced Claude Tag: a new way for teams to work with Claude inside Slack.

At one level, it is exactly the kind of feature many people have been expecting. Instead of opening a separate chat window, you can tag Claude into a channel. It can follow the thread, use selected tools and data sources, work asynchronously, remember relevant context, and help the team move work forward.

In Anthropic's own framing, Claude Tag is not simply a chatbot inside Slack. It is closer to a team participant. There is one Claude in the channel that everyone can interact with. It can build context over time. It can take initiative if ambient behaviour is enabled. It can work on tasks over hours or days.

That is impressive. It is also a much bigger shift than it first appears.

Most of the public conversation about AI agents still focuses on capability. Can the agent write code? Summarise a meeting? Pull data? Draft a response? Chase a task? Work while we are doing something else? Those questions matter. But for leaders, the deeper question is this:

When AI joins the team, who still owns the decision?

Formally, the answer seems obvious. The human leader still owns the decision. The executive team remains accountable. The board, regulator, customer, employee, or investor will still expect human beings to explain what happened.

But decisions are rarely shaped only at the final moment of choice. They are shaped earlier by what gets noticed, remembered, summarised, ignored, framed as urgent, treated as evidence, escalated, or quietly allowed to disappear. In other words, decisions are shaped by context. That is why Claude Tag is interesting. And it is why some people have reacted strongly to it.

Slack is not just a messaging tool. In many organisations, it is where work actually happens. It is where questions are asked, assumptions are revealed, exceptions are handled, relationships are maintained, decisions are clarified, misunderstandings are repaired, and half-finished threads either move forward or quietly vanish.

Much of organisational intelligence lives there: the "we tried that last quarter" context; the "ask Priya before you move this" context; the "this customer is sensitive because of what happened in March" context; the "that looks simple, but it always touches three other teams" context.

So when an AI agent enters that environment, remembers relevant channel history, connects to tools and data, and begins helping work move forward, it is not merely producing output. It is beginning to participate in the organisation's context layer.

Ashwin Gopinath described Claude Tag as a "Trojan horse." I would use slightly less dramatic language, but I think the underlying concern is right. The issue is not that Anthropic is doing anything malicious. The issue is incentives, dependency, and ownership.

Mathew Berman, discussing the same concern in a recent video, sharpened the point. If a shared AI coworker can read across people, tools, documents, conversations, decisions, and tasks, it can begin to form a working graph of the company. That is the real shift. The agent does not only answer questions. It may begin to know how work gets done: who gets asked, which documents matter, which customer issues are sensitive, which projects are behind, which arguments keep recurring, and which interpretation of the past has become the team's working version of reality.

That can be extremely useful. An AI agent inside the flow of work could reduce coordination drag, catch unresolved threads, help new team members understand what is happening, draft better responses, pull together scattered information, and keep work moving across functions and time zones.

For large organisations, that is not a small benefit. The danger is that usefulness is often how dependency begins. At first, the agent saves time. Then it saves effort. Then it saves attention. Then people begin to assume it knows the context. Then it becomes easier to ask the agent than to ask the person. Then the agent's summary becomes the practical version of what happened. Then its interpretation becomes the starting point for the next decision. No single step feels dramatic. But the centre of gravity moves.

This connects to a broader issue I have been thinking about: the handoff of trust - when reliance on AI exceeds the human or organisational capacity to evaluate, contest, or correct it, while responsibility for the consequences remains human.

With AI agents, that risk becomes sharper because the output may not stay as output. It may become action, workflow, follow-up, or the thread everyone else picks up from.

Claude Tag adds yet another layer. If handoff of trust is about judgment slipping away, then handoff of context is about the conditions for judgment slipping away. And that matters because leaders do not only need better answers. They need better judgment. They need to understand what has been framed for them. They need to know what the AI has seen, what it has not seen, what it has inferred, what it has compressed, and what it has treated as important. They need to know when convenience has become authority.

Anthropic has clearly thought about some of the operational issues. Its agent identity model gives Claude its own identity, permissions, tools, and access controls. That is important. It is much better than pretending an agent is just acting as whichever user tagged it. But identity and permissioning are not the same as judgment. Knowing what an agent is allowed to access does not settle how its contributions should be interpreted, challenged, verified, corrected, remembered, or acted upon.

For senior leaders, the practical questions are therefore not simply: Can we deploy this? Can we control access? Can we keep the data safe? Those questions matter, but they are not enough. The deeper questions are: What context is the agent building? Who can inspect it, correct it, or export it? Can another model, provider, or internal system use it? Where might the agent's interpretation become quietly authoritative? Which decisions can it shape, and which decisions must it never make? When does its output need evidence, review, escalation, or human sign-off? How will teams avoid treating fluency, memory, or convenience as authority?

These are not anti-AI questions. They are the questions serious AI adoption now requires. The right answer is not to keep AI agents outside the organisation. That would be unrealistic and, in many cases, self-defeating. The right answer is to become much more intentional about the relationship between AI agents, human judgment, organisational context, and accountability.

I like the following phrase from the online discussion: "Rent the intelligence. Own the context." I would add:

Own the judgment.

Own the accountability.

Own the decision.

That does not mean building every model yourself. Most organisations will not and should not do that. But the more AI agents become embedded in work, the more important it becomes to distinguish between the intelligence layer, the context layer, the workflow layer, and the accountability layer. If those all collapse into one convenient agent experience, leaders may find that decisions are still formally theirs, but increasingly shaped by a system they do not fully understand, inspect, or govern.

So what should leaders do?

First, treat shared AI agents as organisational actors, not clever utilities. They may not be human or conscious, but they can still affect work, attention, interpretation, decision quality, and organisational memory.

Second, create a team or channel AI charter before the habits harden. Be explicit about what the agent is for, what it is not for, what information it may access, what it should remember, what must remain human, when its output must be checked, and when decisions need escalation.

Third, design for portability and plurality. That may mean multi-provider thinking, open standards where possible, clear export paths, separable memory layers, and enough internal understanding to avoid confusing a convenient agent interface with the organisation's own knowledge of itself.

Fourth, train the human disposition, not just the tool use. The same AI agent can become an answer machine, an oracle, a shortcut, a mirror, a challenger, or a disciplined thinking partner depending on the stance humans take toward it.

Executives and teams need to learn how to work with AI agents in ways that strengthen judgment rather than bypass it. That means curiosity without credulity. Challenge without defensiveness. Speed without surrender. Memory without dependence. Usefulness without hidden transfer of responsibility.

The next phase of AI in organisations will not be defined only by better models. It will be defined by the quality of the human-AI working systems we build around them. Some will become convenient dependencies. Some will become opaque operating layers. Some will make organisations faster while making them less able to explain themselves. But some, if designed and used carefully, may help teams think more clearly, surface assumptions earlier, preserve context better, and make more accountable decisions.

The difference will not come from the tool alone. It will come from the discipline of the relationship.

That is the work of AIchemy: helping leaders and teams build the questions, habits, boundaries, and evaluability they need to use AI agents as disciplined thinking partners, strengthening decision quality without handing off judgment, context, or accountability.

Because when AI joins the team, the aim is not to hand over the decision. It is to make the human decision better.