Before you turn on AI in your team chat: a ten-question readiness checklist

Turning on AI inside your team chat is not like turning on dark mode. An assistant that can summarize, draft and answer questions has to read the conversations behind them. Those conversations were written by people who assumed a certain audience, in rooms with a certain level of privacy, and some of them involve clients. The assistant will also write things in your company's voice, and whatever it says carries your company's authority whether you meant it to or not.

None of this is a reason to wait. Teams that use AI well in their chat catch up faster, write clearer updates and find decisions that would otherwise be buried. It is a reason to ask ten questions first, and to roll it out in a way that earns trust rather than spending it.

Start with the reading question

Every other question depends on this one: what can the assistant read, and on whose behalf? The right answer is that it reads exactly what the person asking could already open, and nothing more. An assistant that answers from a channel you are not in has just leaked that channel to you. An assistant that gives a client a summary that includes your internal room has done something far worse.

There is a simple test. Pick a channel you are not a member of, and ask the assistant what was discussed there this week. The only acceptable answer is that it cannot tell you. If it can, stop evaluating and ask the vendor how that happened.

The ten questions

1. Does it read only what each person can open? Not what an admin can open, and not the whole workspace: the specific person who asked.
2. What happens to prompts and outputs? Are they stored, for how long, and are they ever used to train models? Get the answer in writing.
3. Can an admin turn it off, for everyone, immediately? Not per feature, not by ticket: one switch, and it is off until you turn it back on.
4. Do clients and guests get AI? If they do, what can it read on their behalf? A client should only ever reach their own rooms, with or without an assistant.
5. Can people tell when something was drafted by AI? The tool may not mark it, so your policy has to say when a person must review and own what goes out.
6. Where does it run? Which region, which providers, and under which agreement. Your security and legal people will ask, so ask first.
7. Does it work from your real history, or only from new messages? An assistant that starts from zero is far less useful than one that can read the years of context you bring with you.
8. What does it cost at your size? Per seat, included on certain plans, or metered? Model the number for the team you will have in a year, not the one you have today.
9. What does it do when it does not know? The right behaviour is to say so. The wrong one is a confident paragraph with no basis in anything your team wrote.
10. Who owns the output? Summaries, drafts and answers generated from your conversations should belong to your business, with no claim on them from anyone else.

Where AI in chat earns its keep, and where it does not

Being clear about this up front sets expectations and stops the two failure modes: a team that never uses it, and a team that trusts it with things it should not.

It earns its keep on:

• Catching up. A summary of what you missed overnight, after time off, or in a channel you only skim. This is the most common use and the most valuable.
• The morning brief. What changed, what needs you, what is coming today, assembled before you have opened anything.
• First drafts. The weekly client update, the announcement, the reply you have been putting off. You edit; it gets you past the blank box.
• Finding the decision. "What did we agree about the launch date?" answered from the thread where it was agreed, with a link to it.
• Language. Replying in a colleague's or client's language, with the meaning intact.

It does not earn its keep on:

• Making decisions. It can lay out what was said; a person decides.
• Anything legally or financially binding that goes out without a human reading it first.
• Replacing the conversation. A summary of a hard discussion is not the discussion.
• Judging people. Activity, tone and productivity are not things an assistant should be asked to score.

A rollout that earns trust

How you introduce the assistant matters as much as which one you choose. People are rightly wary of something that reads their messages. A rollout that respects that wariness gets adopted; one that ignores it gets quietly worked around.

1. Start with one team that wants it, for two weeks, with three named uses: summaries, drafts and the daily brief.
2. Tell the whole company, before the pilot, exactly what the assistant can read, in one sentence: it reads only what the person asking could already open.
3. Publish the admin off switch. Knowing it exists is most of what people need to relax.
4. Collect the awkward moments in the pilot: the summary that missed the point, the draft that sounded wrong. These become your policy's examples.
5. Write the policy from what you learned, not from a template, and keep it to one page.
6. Expand team by team. Each new team hears the one sentence and the off switch before anything else.

The one-page policy

A workable AI-in-chat policy fits in a handful of lines. Yours will differ, but it should answer the same things.

• What the assistant is for: catching up, drafting, finding decisions, the daily brief.
• What it is not for: decisions, unreviewed external messages, evaluating people.
• Who reviews what: anything sent to a client or published outside the company is read and owned by a named person.
• What it can read: only what the person asking could already open. Clients only reach their own rooms.
• Who can turn it off: workspace admins, for everyone, at any time.
• Where to raise a concern, and who answers it.

Pair the policy with your broader approach to communication ownership. If you have not written that down either, our guide to who owns your company's conversations is the place to start, and what an AI-native workspace means covers the difference between AI bolted on and AI built in.

How SphereOne answers these questions

SphereOne Assist was designed around the reading question. It only reads conversations the person asking can already open. It can't surface a channel you aren't in, and a client only ever reaches their own rooms, with or without Assist. Admins can switch AI off for the whole workspace, and back on, at any time, from the admin settings.

Assist does the work on the list above: summaries with action items, drafts and rewrites, replies in your language, answers from your workspace, and a daily brief every morning. Because SphereOne imports your existing history, it works from the context your team has already built rather than from a blank page.

Assist is included on every paid plan at different levels, with the fullest version on Enterprise; the Free plan does not include AI. Pricing lists exactly what each plan includes, so you can answer question eight before you start.