Skip to content

How the Order of the Work May Shape What AI Takes On

  • Paper: Chaining Tasks, Redefining Work: A Theory of AI Automation
  • Authors: Mert Demirer (MIT), John J. Horton (MIT and NBER), Nicole Immorlica (Yale and Microsoft), Brendan Lucier (Microsoft), Peyman Shahidi (MIT)
  • Posted: NBER working paper, February 2026; arXiv preprint, June 2026. The arXiv listing says it was accepted to EC '26, the ACM Conference on Economics and Computation.
  • Who it's for: Leaders deciding which parts of a workflow to hand to AI, and how to reorganize roles around it
  • Link: arxiv.org/abs/2606.15960

The one-line takeaway

In this model, AI pays off most when the steps it handles well sit together, so one person checks the end result. Where AI-suitable work sits in a workflow matters too, not just how much there is.

What they did

Most economic models of AI treat a job as a bag of separate tasks, each judged on its own for what AI can do. This paper argues the order of the work matters.

It models work as a sequence of small steps. A firm groups neighboring steps into tasks, and tasks into jobs, with one worker per job. Each step can be done three ways:

  • Manual: a person does it.
  • Augmented: AI does it, and a person reviews the output.
  • Automated: AI does it and hands the output straight to the next AI step, with nobody checking.

A run of AI steps in a row is a chain. Every step but the last is automated. The last is augmented: a person checks the final output. The model assumes that check costs the same however long the chain is. If any step goes wrong, though, the whole chain has to be redone.

To check the theory against data, the authors built a dataset for 872 US occupations from four sources (Demirer et al., 2026):

  • O*NET, the US Department of Labor's database of the tasks each occupation involves (May 2023 release);
  • earlier researchers' ratings of which tasks AI could do in half the time at equal quality (what "AI-suitable" means below);
  • Anthropic's Economic Index, which shows which tasks people actually bring to Claude (these are the "AI-used" tasks below);
  • an order for each occupation's tasks, generated by GPT-5-mini, an OpenAI model.

In the data, each O*NET task stands in for one step of the model.

What they found

AI-used tasks sit next to each other more than chance predicts. The authors compared the real data with 1,000 random reshuffles of it. The real runs of AI-used tasks sat at the extreme end: longer than chance would produce. The average run was 1.45 tasks, counting a lone AI-used task as a run of one (Demirer et al., 2026).

Scattered AI-suitable work goes with less AI use, under one definition. Take occupations with the same share of AI-suitable tasks. Those whose suitable tasks are spread out through the workflow have fewer AI-used tasks. Under the broader definition of "AI-suitable," this held in all three ways the authors ran the test, and was statistically significant each time. Under their stricter definition, it was not statistically significant (Demirer et al., 2026).

A task is likelier to be AI-used when its neighbor is. To compare like with like, the authors matched tasks involving the same kind of work across occupations, 10,708 tasks in all. A task with an AI-used task right before or after it was 12 percentage points likelier to be AI-used. After accounting for more differences between tasks, in a smaller sample of 4,096, the gap was 4 to 5 points and still statistically significant. Once those differences were accounted for, tasks two places away made little difference (Demirer et al., 2026).

In theory, gains can arrive in jumps. In a worked example, early AI improvements change nothing, then cut costs gradually. Past a quality threshold, the best setup flips. Both steps become one chain, overseen by a single lower-skill worker who checks the final output. Cost then drops sharply. The authors link this to the "J-curve," where productivity dips after firms adopt a new technology and then climbs. In our reading, the model covers only the climb: in the worked example, cost never rises.

How much to trust it

Take the core idea seriously as a way to look at a workflow. The evidence fits the theory but doesn't prove it.

What it is. A theory paper by researchers at MIT, Yale, and Microsoft, circulated as an NBER (National Bureau of Economic Research) working paper and an arXiv preprint. The one sign of outside review is the EC '26 acceptance, and the listing doesn't say what that review covered.

The data is thinner than the abstract suggests.

  • It sees only Claude. "AI-used" means tasks people brought to Claude, as Anthropic classified them. Use of other AI tools is invisible to the study.
  • Some real AI use is counted as manual. About a third of the Claude-linked tasks had every conversation filtered out, either because it couldn't be classified or to protect privacy. The authors counted them as manual. They argue this makes their results understate the pattern (Demirer et al., 2026).
  • The workflow order is a model's guess. Nobody watched real work. GPT-5-mini ordered each occupation's tasks. Ten differently worded prompts gave broadly similar orders, but the paper reports no check of the orders against observed work (Demirer et al., 2026).
  • It can't see unchecked AI steps. The model's key idea is automation inside a chain. Anthropic does label some chats as automation, but its labels ignore where a task sits in a workflow. The main tests count both labels as AI use. Using the automation label alone, the effects were smaller and often not statistically significant.
  • It measures individual AI use, not productivity, profit, or firm decisions. The data are people's chats with one tool. The model is about firms redesigning jobs, which the data never observe.
  • A simpler story isn't ruled out. This is our read, not the authors'. GPT-5-mini may place similar tasks side by side, and similar tasks tend to attract AI for the same reasons. That would create clusters without any chaining. The authors address this for tasks within the same kind of work, but not in general.
  • The authors don't claim the relationships are causal. The abstract's "supports" and the introduction's "direct evidence" read stronger than that.

Do the conclusions follow? The theory's results follow from its assumptions. Two do most of the work: the flat cost of checking a chain, and unseen AI steps costing no human time. So the headline claims show what can happen, not what does. In our view, where rules or audits force human checks mid-chain, the case for long chains weakens.

Interests to know about. Two authors work at Microsoft, which sells AI. On the NBER release page, John Horton discloses that he is a paid advisor to Anthropic, the source of the paper's AI-use data. The arXiv version carries no such statement. Horton also co-founded Expected Parrot, the software the authors used to run GPT-5-mini; neither version says so. This brief was drafted with Claude, which Anthropic makes. None of this makes the results wrong.

So what, for you

Most of these come from the model, not the data. Use them as ways to look at your own workflow.

  1. Map the workflow in order before you pick what to hand to AI. Mark the runs of AI-suitable steps. In the model, a step that sits alone between human steps gains less, because it needs its own check.
  2. Where a final check reliably catches errors, check at the end of a run. In the model, a longer chain adds risk but no extra checking. If your reviewer can't spot an error made three steps back, keep the checks closer together.
  3. Judge a step as part of its run. A step a person handles more cheaply alone can still belong in a chain, if taking it out adds another checkpoint.
  4. Read a flat pilot as early evidence. In the model, gains can wait for a quality threshold and then jump, so map the workflow again when the tools improve. This covers flat results, not losses. The model doesn't explain early losses.
  5. Check whether a job boundary cuts through a chain. The model also lets firms redraw roles. Handing a chain to one person to oversee can lower the skill a role needs; merging duties can raise it. Neither is tested against data.

For the wider debate on rebuilding a process around AI, see the research note Rethink the Process, or Just Add AI?.

The fine print

  • It can't size your redesign. The model needs each step's time, skill level and AI success rate. Those aren't in the data, so there's no estimate of where real thresholds sit.
  • The model holds the steps fixed. It regroups existing steps and moves the checkpoints. It doesn't model reordering steps or inventing new ones, so "rethink the whole process" goes beyond it.
  • Failed chains are simply rerun. The model assumes every AI error is caught and redone, so it can't capture a mistake that slips through.
  • Hand-offs between people are held fixed. The model assumes AI doesn't make passing work between people any cheaper. The authors say that in reality it might.
  • Chains in the data are short. At 1.45 tasks on average, the long chains the theory is about are mostly not yet visible.

Sources

  • Demirer, Horton, Immorlica, Lucier & Shahidi (MIT, Yale, Microsoft) — Chaining Tasks, Redefining Work: A Theory of AI Automation, 2026. Empirical evidence supports the model's key predictions that (1) AI-executed steps co-occur in chains, (2) dispersion of AI-exposed steps lowers AI execution at the job level, and (3) adjacency to AI-executed steps increases the likelihood that a step is AI-executed. ... the average AI chain length in the original dataset is 1.45, and the average number of AI chains per occupation is 2.10 ... while we do not claim that these relationships are causal. View source · verified 2026-09-15 · primary