Which tasks should you give AI? A stake-and-ambiguity matrix

Your list for the week has 20 lines. Which go straight to AI, which does AI draft for you to review, and which do you keep? I try an answer on a hypothetical case: Lan, operations manager at Chuyến Mẫu ("sample trip"), a delivery firm with 15 lorries in Can Tho. The firm, Lan and every figure are my invention.

The framework she uses is The AI delegate matrix. Its two axes have sources. The four cells do not: no source proposes this matrix.

Which tasks should you give to AI, and which two axes decide?

Ask two questions of each task: how much damage follows if AI gets it wrong, and is the brief clear or ambiguous? The answers put the task in one of four cells.

The stake axis borrows from Parasuraman, Sheridan and Wickens (2000), a paper written for designers of automated systems such as air traffic control. They define the risk of a decision as the cost of an error multiplied by the probability of that error, and write that low-risk decisions are strong candidates for high-level automation (Parasuraman et al., 2000, p. 292). I borrow that one idea.

"Ambiguity" is the site's word, not any paper's. What the research offers is the jagged technological frontier. Dell'Acqua and colleagues tested 758 consultants at Boston Consulting Group in spring 2023: one group had no AI, two had GPT-4. On 18 tasks inside AI's reach, AI users completed 12.2% more tasks, finished 25.1% faster, and human graders rated their work more than 40% higher. On one task built to sit outside its reach, AI users were 19 percentage points less likely to be correct (Dell'Acqua et al., 2023). The authors stress that tasks of similar apparent difficulty can fall on either side, and that even professionals struggle to see where the frontier runs.

So ambiguity predicts little about whether AI will fail. The experiment's inside-the-frontier tasks were fairly open, such as proposing at least ten ideas for a new shoe or drafting a press release (Dell'Acqua et al., 2023). The axis tells you how much judgement a person must supply: a clear task can be checked against a known answer, an ambiguous one has none.

How do you score a week's 20 tasks?

List all 20 and score each from 1 to 5 on both axes; 3 or above counts as high. The scale and threshold are mine, not any source's. Stake 1 means a fix takes minutes, 5 means losing a major customer or breaching a contract. Ambiguity 1 means you can write down what "right" looks like and check it in five minutes, 5 means many answers would be acceptable.

Here are 8 of Lan's 20 tasks. Every score is hypothetical.

Across all 20, Lan counts 7 tasks in "AI does it all", 6 in the drafting cell, 4 in "human does it, AI challenges" and 3 human only. Task 3 runs against instinct: ambiguity 5 does not mean AI is weak, because the creative tasks in the 2023 experiment sat inside GPT-4's reach (Dell'Acqua et al., 2023).

How do you run a one-trial test before trusting a cell?

Take a past case with a known answer, give it to AI exactly as you would give it to a person, and compare. The trial beats the cell. Four steps:

1. Pick a finished case whose answer you trust.
2. Give AI the information the person doing the job would have. The five-part context bundle lists five parts to keep ready.
3. Compare every figure and conclusion.
4. If it matches, keep the cell. If it is wrong in a way a skim would miss, move the cell towards the person. If the error is obvious, keep the cell and add a review step.

The frontier cannot be read off how hard a task looks, and in the experiment the people who did worse with AI tended to adopt its output and question it less (Dell'Acqua et al., 2023). One trial is not proof; where money is involved, I would run three cases.

Task 8 shows why. Lan scored it stake 2, ambiguity 1: clear figures, fifteen minutes' work. She gives AI last week's spreadsheet, which she has already analysed. Suppose AI picks the route with the most returns. But one line in the dispatcher's notes says 40 of those returns came from a single customer recalling a faulty batch, not from a trend. The figures match, so a skim would miss it. Lan moves the task to "AI drafts, human reviews" and logs: check the notes column too.

The outcome is invented; only the shape of the problem is borrowed. The outside-the-frontier task in the 2023 experiment was built from a spreadsheet and interview notes whose details changed the conclusion: 84.5% were right without AI, 60% and 70% with it (Dell'Acqua et al., 2023). The paper does not say what AI would do with Lan's spreadsheet today.

When do three quick questions do instead of the matrix?

When you have no time to score the list, ask three questions before each new task. They are Minimum viable judgement, this site's grouping of three more detailed frameworks; "minimum viable judgement" is the site's own label.

1. Is it worth giving to AI? Anthropic (2024) says agentic systems often trade latency and cost for better task performance, and that you should consider when the trade makes sense (Anthropic, 2024). For the person delegating, I read that as: how long to write the brief and check the result, against doing it yourself?
2. What does AI need to know? Anthropic (2025) advises aiming for the minimal set of information that fully outlines the behaviour you expect, and notes that minimal does not necessarily mean short (Anthropic, 2025). See The five-part context bundle.
3. What is the damage if it is wrong? The cost of consequences is one criterion for choosing a level of automation (Parasuraman et al., 2000, p. 292). See The five-rung verification ladder and the five-rung verification ladder post.

A hypothetical example: Lan needs a 40-page warehouse lease summarised for the director before renewal. It is worth giving, because reading it all takes two hours. AI needs to know the decision is whether to renew, and to look hard at the early-termination penalty. If the summary misses that clause, the firm renews on the wrong terms, so Lan reads the clauses the summary points to. With large consequences, three questions are not enough on their own.

Going deeper: what do the sources show, and what do they not show?

Dell'Acqua and colleagues (2023) is a Harvard Business School working paper dated 22 September 2023. The study was pre-registered. They used the version of GPT-4 available at the end of April 2023 and randomly assigned people to no AI, GPT-4, or GPT-4 plus a prompting overview. The inside-the-frontier experiment had 385 people and the outside one 373 (Dell'Acqua et al., 2023). Limits I draw from the text itself:

• One firm, one model, one moment, and a very high-skill group; the authors note that "higher-skill" is only relative. I have no data on the models of today.
• There is only one outside-the-frontier task, built by the authors so that GPT-4 would err. They write that it was difficult to design a task outside the frontier where skilled people would consistently beat AI. So the paper does not say how common such tasks are in real work.
• The "more than 40%" is the human graders' score. GPT-4 grading gave smaller gains, about 17% to 19%, and the authors note that GPT-4 is a more lenient grader.
• The paper does not measure damage in money, so the stake axis does not come from it.
• The version I read is a working paper that calls itself a draft. Some authors work at BCG, as the paper discloses.

Parasuraman, Sheridan and Wickens (2000) set out a model of four types of automation, each with graded levels, before generative AI existed. They say the model does not prescribe what should be automated (Parasuraman et al., 2000, p. 286), describe it as qualitative, and admit they have not said how to weigh the criteria against each other (Parasuraman et al., 2000, p. 294). The two Anthropic posts are a model developer's guidance to agent builders; I use one sentence from each. The four-cell matrix and the three questions are the site's own grouping, and no source proposes them.

I read the body of the Dell'Acqua paper from the MIT Sloan PDF, but only the captions of its figures; I did not read the tables after Table 1 or the appendices closely, nor any later version. From Parasuraman and colleagues I read the abstract, the model, the criteria (pp. 290–293) and the limitations (p. 294), and skimmed the rest. I read both Anthropic posts in full as they stand today.