Which work is fair game — the triage

Agents at Work — CC BY 4.0

The most expensive mistake with agents isn’t a bad build. It’s building the right agent for the wrong job — automating work that should have stayed with a person. So before any of the how-to, a decision: which of your work is fair game for an agent, which wants an agent and a person, and which stays human.

You don’t need a new framework for this — the Village Small Business Course already has one, the Process Triage Worksheet, and it’s worth running properly rather than re-inventing. Here we’ll use it, and lean on the one part that matters most for agents.

The questions, quickly

Run each piece of repeated work past six questions:

1. Rule or judgment? Could you write down the rule, or does it “depend”?
2. Reversible? If it goes wrong, can it be undone?
3. What does a mistake land on? No one much — or a client, a colleague, a member of the public, money?
4. Does it need the picture? Foreseeing a consequence, or knowing this particular customer?
5. And the person? If an agent takes it, where does the person who did it go?
6. Whose data, and whose call? Does it handle other people’s personal information, or decide something about a person?

The first five place the work on a spectrum: mostly-rule, reversible, lands-on-no-one, no picture needed → an agent can do it, logged, with a human reachable. A mix → augment: the agent prepares, a person decides. Heavy “depends”, not reversible, lands on a person → human.

The sixth is different. It’s not a score; it’s a gate. If the work handles someone else’s personal information or makes a call about a person, it routes to a human whatever the other five say. “Low-stakes” and “reversible” don’t clear it — the Privacy Act reaches even what you type into the tool, and a wrong or unfair call about a person isn’t undone by deleting a row.

Why this gets harder, not easier, as agents improve

Here’s the trap the fast pace sets. As agents get more capable, more of your work becomes something an agent could technically do. The temptation is to read rising capability as permission — “it can handle this now, so let it.” But capability answers question one (can it?), never question six (should it?) — and “should it?” is Anchor 4, the purpose test: does handing this over actually serve the person on the other side, or just serve me? The rights gate doesn’t move because the model got better. If anything, a more capable agent handed a personal-data decision is more dangerous, because it’s more convincing and you’re more tempted to trust it.

So the triage isn’t a one-off. Re-run it as the tools change: the answer to “can an agent do this?” drifts; the answer to “should it decide about a person?” holds still.

Pick your first slice deliberately

For your first agent — the one you’ll learn on — choose work that lands cleanly in the agent bucket: a real rule you could write down, easily reversible, a mistake that lands on no one much, and no other person’s data in play. Keying a category of routine transactions, sorting inbound mail into folders, drafting a standard reminder. Dull, repeatable, checkable. That’s not a lack of ambition; it’s how you build the judgment to take on the harder work later, with your eyes open.

Save the “augment” jobs — where the agent prepares and you decide — for once you’ve seen a simple one behave. And treat the “human” and gated work as exactly that: not yet, and maybe not ever, whatever the tools can do.

Next

You’ll notice question six kept coming up, and that it can override everything else. That’s not an accident — it’s the single most important habit in deploying agents, and it deserves its own lesson. Next: the data question, first.