AI agentsAI governanceLeadershipWorking with AI

Before you let an AI act on its own in your team chat

AI agents are moving into team chat tools and can now act unprompted. The two calls a leader should make before letting one into the firm's shared channels.

Good Transformer24 June 20269 min read

The newest AI tools no longer wait in a separate window for you to open them. They are moving into the chat your team already works in, where they can see the conversation, remember it, and in some cases act without being asked. On 23 June, Anthropic launched Claude Tag, which puts an AI agent into Slack as a standing member of a channel that anyone can tag for help. It follows Microsoft switching on an autonomous agent inside Microsoft 365 the week before. The shape of the thing has changed, and it changes what a leader has to decide. Two calls now sit on your desk, and neither is technical: what an always-present AI may see and keep across your shared channels, and where it is allowed to act on its own rather than only when asked.

Start with what is actually new, because the word "agent" has been doing a lot of work lately and the differences matter. Until now, the agent story has mostly been about delegation. You give an AI a defined job, it works through the steps, and it hands back a result you check. That is what Microsoft's Copilot Cowork does, and we wrote about the decisions it forces in an agent that finishes work you assign. Claude Tag is a different kind of colleague. In Anthropic's description it is "multiplayer", meaning one shared Claude sits in a channel and everyone can see what it is doing and pick up where someone else left off. It "learns over time" by building context from the channels it follows, and it can draw on other channels it has been given access to. It "works asynchronously", taking a task away and even scheduling work for itself over hours or days. And if an optional "ambient" mode is switched on, it "takes initiative", flagging things it thinks you need to know and chasing threads that have gone quiet, in its own words, "without being asked". It is in beta for business customers now, and the same pattern is arriving across the tools most firms already run on.

That is genuinely useful. A great deal of what clogs a professional week is the chasing, the summarising and the pulling-together that an always-present assistant can do well. But an AI that lives in the room where your team talks about clients, remembers what it hears, and occasionally speaks up on its own is not the same as a tool you open, use and close. It asks two questions a tool never did.

Decision one: what is it allowed to see and keep?

The first decision is about memory, not just access. Granting an AI a login is a familiar idea. Granting it a permanent seat in the channels where your team thinks out loud, with the ability to remember what it sees there and carry that context forward, is a different proposition, and it lands hardest in exactly the firms that handle other people's confidential business.

Picture the channels as they really are, not as a tidy diagram. In an accountancy or advisory firm, someone pastes a client's figures into a thread to ask a colleague a quick question. In recruitment, a channel carries candidates' real names, current employers and salary details. In corporate finance, a deal channel names the target and the price weeks before anything is public. In a legal or risk-aware practice, a matter channel holds exactly the material a client trusts you to keep close. These are not edge cases. They are the normal texture of the work. An always-present AI that learns across those channels is being handed a standing view of your most sensitive conversations, and the question is no longer "can it log in" but "what should it be in the room for at all".

Anthropic has built controls for this, and they are the right controls: an administrator decides which channels and tools the model can reach, sets up separate scoped identities so a setup built for one use does not carry its memory into another, and Claude does not draw from private channels. The point is that those settings are a decision someone has to make on purpose, before the team starts tagging, not a default to discover later. The safe instinct is the same one you would apply to a new hire on day one: give it access to the channels where it genuinely helps, keep it out of the ones that carry the most sensitive material until you are sure, and decide what it is allowed to retain rather than letting that build up unseen. Your duties to clients and candidates over their information do not change because the new colleague is software.

Decision two: where does it act on its own?

The second decision is the sharper one, and it is the part most leaders have not yet clocked, because it hides behind a single setting. There is a real difference between an AI that acts when you tag it and an AI that acts on its own initiative. The first is a faster way to get help you asked for. The second is a colleague that watches the channel and decides, by itself, when to step in.

Used well, proactive behaviour is a relief. An AI that quietly chases your own unanswered admin, reminds the team that a quote has gone cold, or pulls together a status note from a busy thread is doing the work nobody enjoys. Used without thought, it is a colleague with confident judgement and no instinct for the room. An AI that follows up on a quiet thread is helpful when the thread is your internal to-do list and a problem when the thread is a delicate client matter or a conversation about a member of staff. The machine does not feel the difference. It does what its instructions allow, completely and literally, which is the same reason every agent needs a clear owner, a point we made in a dozen agents and no one in charge.

So draw the line on purpose. Decide where the AI may only respond when tagged, and where, if anywhere, it may act unprompted. A reasonable starting position for most firms is to keep it reactive on anything client-facing or personnel-related, and allow proactive help only on internal, low-stakes work where a wrong nudge costs nothing. This is the same judgement that sits behind what to keep off an AI's desk: the gathering and the chasing can be shared, the calls that carry your firm's name stay with a person.

The controls are there. The job is to set them.

It is worth saying plainly that none of this requires you to become technical. The tools now ship with the knobs a leader needs: which channels the AI can see, what it may remember, whether it can act on its own, a cap on what it can spend, and a log of everything it did and who asked for it. The work is not to understand the engineering. It is to make four or five deliberate choices, write them down, and tell the team what you decided, in the same way you would set out what a new starter can and cannot do in their first week.

If your firm wants help making those calls in a way that fits your actual work rather than a generic policy, that is what our AI Lessons for Leaders coaching is for: practical sessions built around the decisions in front of you, not a tool demo.

There is also a human point that is easy to skip. The people who work in these channels are now working alongside a colleague that sees what they say and may act on it. They deserve to know it is there, what it can see, and what it is allowed to do, before it arrives, not after they notice it reading along. An AI teammate introduced openly, with clear limits, is something a team can trust and use well. One that appears by stealth in the channels where they speak freely is something they will quietly work around, and you will have spent money to make your people more guarded.

The move this week is small and worth making before the habit forms. If a tool like this is on your radar, decide two things in advance: which channels an always-present AI is allowed into and what it may keep, and where it must wait to be asked rather than act on its own. Set those before the first person types @ and tags it in, and the always-on colleague is an asset. Leave them to default, and you have given a confident, literal-minded newcomer a seat in your most sensitive conversations and the freedom to speak up whenever it likes. If you would value a second view on where those lines should sit, book a discovery call and we will work through it against your real channels.

What is Claude Tag, or @Claude?

It is a way to add Anthropic's AI to a team's Slack as a shared member. Once an administrator gives it access to chosen channels and tools, anyone in those channels can tag @Claude to delegate a task, and it responds in the thread. It can remember context from the channels it follows and, if enabled, act on its own initiative. It launched in beta for business customers on 23 June 2026.

Is it safe to give an AI access to our team chat?

It can be, if the access is scoped on purpose. The risk is not the tool itself but giving an always-present AI a standing view of channels that carry confidential client, candidate or deal information, and letting it retain that context. Limit it to the channels where it genuinely helps, keep it out of the most sensitive ones, and use the administrator controls for memory, spend and an activity log. Your confidentiality duties to clients and staff still apply.

What does proactive or ambient AI mean?

It means an AI that acts without being asked. Rather than only responding when you tag it, an ambient agent watches the channels it is in and decides for itself when to flag something or follow up. That is useful for chasing internal admin and risky on client-facing or personnel matters, so it is worth deciding where it is allowed and where it must wait to be asked.

We are a small firm. Do we need to think about this yet?

Even if you do not adopt one of these tools today, the pattern is arriving in the everyday software you already use, so the decisions are coming either way. Knowing in advance what an always-present AI may see and where it may act means you choose deliberately rather than accept a default, which is cheaper and safer than unwinding it later.