Choose Work You Can Check

Two Ways to Use AI

Rae, a fictional account manager, has client notes. The pilot is complete. Customer testing is blocked by missing access. The launch date is undecided.

She could give AI one of two assignments. Which would you pick?

Assignment A: handle the client update, including choosing a launch date.

Assignment B: write an internal summary of completed work, blockers, and proposed next steps for Rae to inspect.

Both put AI to work. Only one produces something Rae can actually check.

What Makes a Task Checkable

For Rae’s assignment, write down three things before asking AI to start.

First, the specific output AI will prepare. Second, how someone can assess it. Third, the human decisions that stay outside the assignment.

To assess this summary, Rae will compare its claims with the project notes.

For Rae: the output is the internal summary, the checker is Rae, and the launch date stays hers.

"Handle the client" leaves this assignment unclear: it does not say which output to return or which decisions Rae keeps.

Why the Summary Wins

Take each claim in Rae's summary. She can hold it against her notes and see if it matches.

Assignment A fails this test. A chosen launch date has nothing in the notes to compare against, because the notes say it is undecided.

So Rae keeps the launch decision and the client communication.

AI prepares the draft. Rae checks it, decides, and talks to the client.

One Success Is Not a License

Rae's summary came out clean. Does that mean AI can now run the whole client relationship?

No. Harvard Business School researchers describe AI capability as uneven across knowledge-work tasks.[[1]](https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/) They call this a "jagged technological frontier": AI excels in some tasks but falls short in others.

The accurate summary establishes what happened in that summary task. It does not settle the still-undecided launch date in Rae’s notes.

For Rae, summarizing the notes and choosing a launch date are two different tasks. Success at the summary supplies evidence about summarizing those notes. It supplies no evidence that the AI can make a sound launch decision. Treat a new task as its own delegation decision: it needs its own output and someone who can check that output against evidence. A launch date has no such evidence in Rae's notes, so it stays her call.

The Same Split in Support Work

This split shows up outside Rae's desk too. In one customer support deployment, AI suggested responses to human agents.

Agents could accept, edit, or ignore each suggestion. They remained responsible for the conversation.[[2]](https://doi.org/10.1093/qje/qjae044)

One caution: the support study covered a single firm.[[2]](https://doi.org/10.1093/qje/qjae044) Treat it as a parallel to Rae, not a guarantee.

Split Your Own Recurring Task

Pick a task you repeat. Before you hand it to AI, draw the boundary.

Name the output. Name who can compare it with evidence or requirements. Name the decisions you keep.

What if nobody can assess the output? Then narrow it, or bring in a qualified reviewer.

For Rae, the boundary is set: an internal status draft for her to check, and launch commitments that stay her decision.

An assistant accurately sorts supplier emails by topic. Its next assignment would choose which supplier to hire. What evidence is missing before trusting the new assignment?

• Evidence that it can perform the supplier-selection task, assessed by someone qualified to judge the decision.
• Evidence that it can sort the same emails into the same topics again.
• Evidence that it can present the topic labels in a polished report.
• Evidence that it can draft a short message to the chosen supplier.

Answer: Evidence that it can perform the supplier-selection task, assessed by someone qualified to judge the decision.

Sorting and choosing a supplier are different tasks. Repeating or polishing the demonstrated task does not establish decision quality on the new one. A qualified person must be able to assess the new output.

A manager asks AI to prepare a supplier comparison from three proposals. She can verify prices but cannot judge the technical claims, and no specialist is available. What is the best bounded assignment now?

• Choose the best supplier using prices and technical claims together.
• Extract the prices for her to check and leave the technical comparison for a qualified reviewer.
• Write a confident technical recommendation for her to approve.
• Summarize the technical claims as verified facts because they came from proposals.

Answer: Extract the prices for her to check and leave the technical comparison for a qualified reviewer.

The price extraction is checkable with the available evidence and reviewer. The technical judgment is not. Narrowing the assignment preserves useful work without treating an unassessable recommendation or supplier assertion as verified.

A manager has meeting notes that list completed tasks but leave the launch date undecided. Which output can she check from those notes?

• A recommended launch date supported by a polished explanation.
• A client promise inferred from the completed tasks.
• An internal summary of completed tasks with the date marked undecided.
• A prediction that the remaining work will finish this week.

Answer: An internal summary of completed tasks with the date marked undecided.

The summary can be compared with the notes. The other outputs add a date, promise, or prediction the supplied evidence does not establish. The assignment should separate checkable preparation from an additional decision.