// Opening

AI probably arrived in your team through a chat window. Someone typed a real question, got a genuinely useful answer, and a habit was born: when something is awkward, open the window and start typing.

That's a good habit. It holds up right until the work changes shape.

// Part 2

Two shapes of work

Conversation-shaped work: - You're not yet sure what you're asking. - The next move depends on the last answer. - What you want is a decision or a direction, not an output you could have described in advance.

Folder-shaped work: - There are many similar items, and you could count them. - The first item and the hundredth are handled the same way; only the contents differ. - You could write down what "done" means for one item before starting, clearly enough that a colleague could mark a result pass or fail.

The mismatch is folder-shaped work done one exchange at a time.

That's not a criticism of chat. A conversation earns its keep because nothing has to be settled in advance. On a folder, you're paying for flexibility you don't use, and supplying the repetition yourself: one paste, one wait, one read-back, item after item.

Opening the window was the right first move; a conversation is how you find out what a job needs. The cost comes later, if the job was a folder all along.

// Part 3

Three checks on one job

Before looking at any tool, pick one job you did more than ten times last month. Name it as an operation on a single item, a sentence with one verb. If it needs "and" joining two verbs, that's two jobs. Pick one.

Then ask three things about the job.

1. What does this job actually need?

Go back to the last item you did by hand and find the moment you paused. Were you making a decision that could reasonably have gone the other way, or applying a rule nobody has written down?

Then take one item from the middle and write what "done" looks like, in two to four bullets. Could a colleague mark a result pass or fail against them without asking you a question?

  • If you can't write "done", part of this job is still a conversation, and chat may be the right tool for that part.
  • If you can, you already have a specification.

Watch for "it depends on the item". Count the recent items that genuinely needed a decision. Those are the judgement; the rest were an unwritten rule.

2. Is this material allowed to go there?

Who owns what's in the folder: you, your client, or someone who gave it to your client? Open the contract or privacy notice and read the actual sentence.

A permission given for a single paste can quietly extend to everything behind it. So the check isn't whether this item may go there; it's whether all of it may, every week. Where do the copies live afterwards? If a client asked, could you answer with a date and a scope?

3. What does it cost at volume?

  • How many items are in the folder, how many arrive each week, and does it come round again?
  • Per item, how long does the human part take: opening, preparing, waiting, reading back, deciding?
  • Per item, what does the tool charge, in whatever unit it charges?
  • What fraction comes back thin and needs a second pass, and how long does it take to spot which ones?

Multiply by the real count. That usually settles the question.

We won't put a figure on this. Every number is yours: your count, your stopwatch, your bill.

// Part 4

Why volume changes the picture

A single item never tests a way of working. One item is a demo.

Hypothetical example, not a client of ours: a small research team finishes a study with a folder of a couple of hundred interview transcripts. Each needs the same short note: three themes, one quote, a flag if the participant mentioned price. By the fortieth, someone has a rhythm (paste, wait, skim, correct, next) and nobody has called it a project, because no single item was big enough to be one.

At two hundred, the effort becomes visible. It was never the item that was expensive; it was the count.

Cost depends on how you pay. Most chat plans are a flat monthly fee per seat, with usage limits, so the folder mainly costs people's time. Usage-billed tools and APIs charge per amount of text processed, so the bill grows with the folder (see, for example, OpenAI's and Anthropic's pricing pages). It's worth knowing which you're on.

// Part 5

What fits folder work

Once you've done the checks, there are usually three sensible options:

  • The same assistant, pointed at the whole folder. Use a project or file access and give it your written "done", rather than pasting items one at a time.
  • For real volume, a scripted call per item. One fixed set of instructions, run over every item, with results collected in one place. This is a simple, legitimate pipeline.
  • A person checks the exceptions and a sample. Someone reviews the items that were flagged, plus a regular sample of the rest.

And sometimes the right answer is "write the rule down", "change who does it", or "leave it as it is".

// Part 6

The scoping is yours

All three checks resolve into facts inside your business, which nobody outside can answer for you.

The written description also outlasts any tool. A team with a clear page on one job (the operation, the definition of done, the permission, the count) can change tools cheaply and judge them against the real job, not a demo.

// Part 7

If it's folder-shaped

Do the checks on one job this week. A page, a folder, nothing bought.

If it turns out you've been putting conversation-shaped effort into folder-shaped work, and you'd rather not build the replacement yourselves, start with The Catalyst: one real problem, scoped with you, built with you, and handed back for your team to run.

And if the job turns out to be a conversation after all, good. Now you know which one it is.

← All posts