Every project tool now claims AI. Almost none of them will show you what the AI actually did. You get a summary, a cheerful confirmation, and a quiet hope that it did the thing. For an agency, where the work belongs to a client who is paying for it, that is not good enough — and it is the reason most agencies use AI for drafting and nothing else.
The useful question is not what an AI teammate can do. It is what you can see afterwards.
What does an AI teammate actually mean here?
A member of your workspace that happens not to be a person. An agent has a name and an avatar, appears in your member directory, can be @mentioned in a comment, and shows up in the assignee picker beside everyone else.
It differs from a human member in two ways that matter. It has no email address and no password, and it cannot log in through any path — an agent exists only inside your workspace and cannot be phished, borrowed, or left behind at an ex-employee's mailbox. And it does not read notifications, because it is invoked directly rather than nudged.
That is the whole model. Not a chat window bolted onto the side of the product, and not a separate tool you copy work into. A teammate you can assign something to.
How do you give it work?
You assign the task. That is the entire mechanism — pick the agent in the assignee field the way you would pick a person, and the agent starts work on that task.
There is no separate queue, no prompt box, and nothing to copy across. The brief is already on the item, the client is already linked, the due date is already set, so the agent works from the same context your team does rather than from whatever you remembered to paste.
Reassigning the task to a person stops it, exactly as you would expect.
What do you see when it is done?
The trace. Every step the agent took, in order, with what it did and how long it took.
| What you see | Example |
|---|---|
| The step, in plain language | Updated a page |
| A summary of the result | Q4 campaign — status rolled up |
| How long it took | 1.4s |
| Whether it needed approval | Waiting on you, or auto-run |
| An undo button | On page edits and page creations |
Most tools tell you the AI finished. This tells you what it touched, in what order, and lets you put a page back the way it was.
The undo is deliberately narrow and worth being precise about: it applies to pages the agent created or edited, not to everything it did. A page edit is reversible because the previous content is recoverable; a message that has already gone out is not, and pretending otherwise would be worse than not offering undo at all.
What can it do without asking?
That depends on what the action is, and the distinction is built into the product rather than left to the model's judgement. Every tool an agent can call declares a side-effect class, and the runtime enforces it.
| Class | What it means | Behaviour |
|---|---|---|
| Safe | Reads something | Runs |
| Mutating, reversible | Changes something you can put back | Runs, and appears in the trace |
| Sensitive | Reaches outside your workspace or is hard to undo | Stops and asks you |
Running an external integration is classed sensitive. So an agent can read your sheets, roll up a status and draft a page without interrupting you, and it cannot fire something at a third-party system on your behalf without a person saying yes.
This is the part that decides whether an agency can use AI on client work at all. Delegation is only safe when the boundary is enforced by the system rather than by the phrasing of a prompt.
What does it cost?
It is metered per model, so the cost of a piece of work is attributable to the work rather than smeared across a monthly seat price. That matters more for an agency than for most businesses, because an hour spent on a client is a billable fact and an AI cost on that client's work is the same kind of fact.
The practical consequence: you can answer "what did AI cost us on this account last month" instead of estimating it. Most tools that bundle AI into a per-seat uplift cannot.
What should you actually delegate?
Start with work that is well specified, repetitive, and easy to check.
- Good first candidates. Rolling up status across a client's deliverables. Drafting a weekly client update from what changed. Turning a messy brief into a structured one.
- Poor first candidates. Anything where the judgement is the deliverable, anything client-visible that nobody will read before it goes out, and anything where you could not tell a good result from a bad one at a glance.
- The test that works. If you cannot verify the output in under a minute, do not delegate it yet. Not because the agent will fail, but because you will not notice if it does.
How does this compare with an automation?
An automation runs a rule you wrote. An agent decides which steps to take. Both have their place, and the honest split is unglamorous: use an automation when the steps never change, and an agent when the input is messy enough that writing the rule is harder than doing the work.
| Automation | AI teammate | |
|---|---|---|
| You specify | Every step | The outcome |
| Handles messy input | No | Yes |
| Predictable | Completely | Mostly |
| Auditable | The rule is the record | The trace is the record |
| Good for | Repeated identical work | Work that varies each time |
We wrote about that split in more depth in AI agent vs automation, and about what an agent can and cannot see in what can an AI agent see.
Where should you start?
Pick one recurring piece of work you already do by hand every week — the status roll-up is the usual one — and assign it to an agent on a single client. Then read the trace rather than the result. You are checking whether the steps it took are the steps you would have taken.
If they are, give it the same job on the other clients. If they are not, you have learned something specific and cheap, in an afternoon, on one account.
Assign one weekly status roll-up to an agent on one client, then read the trace instead of the summary. Start free