What you can actually hand to an AI teammate

Not everything, and not nothing. The honest list of what comes back useful, what comes back wrong, and how to tell which is which before you delegate.

The UnaTask team, Building UnaTask. Aug 2, 2026. 5 min read.

More on AI teammates

Nadia has an AI teammate on her team and a nagging feeling she is not using it properly. She has given it two things this month. One came back genuinely good and saved her most of a morning. The other came back confident and wrong, and she has not tried since.

This is the normal experience, and the normal conclusion drawn from it is the wrong one. The lesson is not that AI teammates are unreliable. It is that nobody tells you the shape of the work that suits them, so everyone finds out by donating a task and seeing what happens.

So here is the shape, written down plainly, from watching what comes back well and what does not.

The test: does it end in words or a file?

An AI teammate in UnaTask produces two things: a comment on the task, and files attached to it. That is the whole output surface. So the question to ask before you delegate is simply whether the finished job could be handed to you as writing or as a document. If yes, it is a candidate. If the job ends in a decision only you can make, or a conversation with a person, or a change inside a tool you have not connected, it is not.

  • A first draft of something you will rewrite anyway. The blank page is the expensive part, not the editing.
  • A summary of something long, with the bits that matter pulled out.
  • Research pulled together into one place, so you start from a page instead of a search box.
  • A checklist or a plan broken into steps, which you then argue with.
  • Three options to choose between, when your real problem is that you cannot get started.
  • A file: a spreadsheet of the numbers, a document in the format you asked for.

And the other side of the line, just as usefully:

  • Anything where being 90% right is worse than not doing it. A legal clause, a price, a number that goes to a client.
  • Work that needs a fact only a person in your company holds and nobody has written down.
  • Anything that has to go out under your name without you reading it. That is not delegation, that is a hope.

The instructions matter more than the task

The other half of Nadia’s problem is not what she asked for, it is what her teammate knew. An AI teammate has a job description and a set of instructions you write once when you add it to the team, and it inherits the context of the workspace it lives in. A teammate told it writes support replies in your tone, with your product names, will do that job forever. A teammate told nothing will produce something generic and you will blame the model.

Then it gets better on its own terms. Every piece of work comes back with three buttons under it, and if you paste in the version you actually used, that correction becomes a rule the same teammate carries into the next job. The second draft of anything is noticeably better than the first, because by then somebody has told it something true.

Same month, different Nadia. She has stopped thinking of it as a clever thing that might do her work and started thinking of it as the colleague who does the first version. She hands it the draft, the summary, the sweep of research. She keeps the judgement, the pricing and the conversation with the client. Nothing about that division has changed in twenty years of working with junior colleagues, which is probably the point.

Give it the blank page, keep the decision. That line has not moved just because the colleague is software.

See how AI teammates work

Read next

  • Assign a task to an AI the way you assign it to a person
  • AI agents fail one task in three. We built for that.
  • Teach your AI teammate once, and it remembers

More from the blog

  • UnaTask
  • Features
  • Mac app
  • Pricing
  • Blog
  • Comparisons
  • For your team
  • Updates