AI Productivity: Find the Team’s Coordination Bottleneck

Imagine a small agency using AI to prepare client campaign drafts. A writer finishes much sooner, but the work still waits for a weekly approval meeting. The client receives it on the same day as before. The writer has saved effort; the delivery process has barely changed. This hypothetical example shows why an impressive task result deserves a second question: what happens to the work next?

For business owners, AI productivity needs to be examined across the complete job. Drafting, checking, waiting, deciding and delivering are different activities. Improving one can be useful without removing the delay that governs the customer’s experience.

Two field studies, two different questions

Shifting Work Patterns with Generative AI, by Eleanor Dillon and colleagues, reports a six-month field experiment across 66 firms and 7,137 knowledge workers. Workers were randomly selected for access to AI integrated into existing applications. The authors report email time savings among users and less work outside regular hours, but no detectable shifts in the quantity or composition of tasks from individual-level AI provision.

The NBER working paper was issued in May 2025 and revised in November 2025; the journal lists it as forthcoming in American Economic Review: Insights when checked on 23 October 2026. Some authors worked for Microsoft, whose tool was studied. The experiment concerns access within those workplaces. It does not establish what would happen after a deliberate redesign of an entire team’s workflow.

A different experiment, The Cybernetic Teammate, was published online in Organization Science on 12 June 2026. Fabrizio Dell’Acqua and colleagues studied 791 Procter & Gamble professionals working on product-innovation challenges. Participants worked individually or in cross-functional pairs, with or without AI. Individuals using AI matched the performance of pairs without it; AI-assisted proposals also better balanced technical and commercial perspectives.

That is evidence about the assigned innovation work, not proof that every ongoing team can deliver with fewer people. The paper also finds a continuing role for human judgement in selecting ideas. The two studies should not be combined into a single promised productivity percentage: their settings, interventions and outcomes differ.

Follow one piece of work through the system

My practical inference is to test AI at two levels: the task it changes and the complete workflow around that task. The following method is a management proposal, not a protocol validated by either experiment.

Choose a recurring job with a clear finish. For a creator, it might be a checked newsletter ready to send. For an agency, it could be a client-approved campaign. Describe the finish precisely enough that producing a draft cannot be mistaken for completing the job.

Then follow a few representative cases and record:

• Active work: time spent preparing the input and producing the output.
• Review and correction: time spent checking evidence, resolving issues and revising.
• Waiting: time when the job is ready for another step but cannot proceed.
• Completion: when the agreed result reaches the intended recipient, with any necessary approval.

Record why a wait occurs. “Waiting for review” could mean a missing file, an unavailable specialist, an unclear decision or a deliberately scheduled checkpoint. Those causes call for different changes.

Do not confuse effort with elapsed time

Here is an illustrative calculation, not a measured business result. Drafting takes 40 minutes, review takes 20 and corrections take 15: 75 minutes of active effort. If AI reduces drafting to 15 minutes while the other stages remain unchanged, active effort becomes 50 minutes.

That saving may be valuable. But if every completed draft still waits for the same weekly approval meeting, delivery can remain tied to that meeting. Measure both the effort saved and the elapsed time from an agreed starting point to completion. Report them separately.

Also inspect what the receiver gets. Ten additional drafts may create useful options, or simply give an already busy reviewer more material to assess. Ask whether the next person needs more versions, a clearer recommendation, better evidence or a smaller decision.

Change the dependency you have actually found

If work stalls because supporting information is missing, agree what must accompany a submission. If the relevant specialist is routinely unavailable, plan review capacity or adjust the delivery promise. If nobody knows who can decide, make that authority explicit; the MaryChuks guide to mapping decision rights provides a starting point.

Some waiting protects quality. A specialist check or a pause for the client to confirm an important condition may be necessary. Give it a clear purpose and realistic capacity before trying to remove it. A shorter queue is only useful if the work remains fit for its intended use.

Where AI helps someone explore another discipline’s perspective, invite the relevant colleague to test the assumptions. A commercially balanced proposal still needs someone who understands the actual customer, delivery constraints and available evidence.

Run a small comparison without overstating it

Keep the job and acceptance criteria reasonably consistent while trying one change. You might introduce a clearer submission brief or a predictable review window. Compare cases of similar difficulty and note changes in staffing, demand or deadlines that could explain the result.

• Did active effort fall across everyone involved?
• Did the recipient get an acceptable result sooner?
• Did correction work or unresolved issues increase?
• Did work accumulate at a different stage?
• Did the change preserve necessary checks?

Use a small local trial to learn where the process needs attention. It cannot establish a universal causal effect. If AI is helping organise the observations, give it recorded timings and notes; ask it to flag missing information rather than invent durations.

Decide what improvement would justify keeping the change before reviewing the results. Sometimes the outcome worth protecting is less overtime, more careful review or greater capacity for difficult cases. Faster delivery is one possible aim, and should be named when it is the aim.

The next useful AI investment may concern how people receive, check and act on work already being produced. Follow one job far enough to see what the team is waiting for, then improve that step with the people who own it.

Sources

• Dillon, Jaffe, Immorlica and Stanton: Shifting Work Patterns with Generative AI, forthcoming in American Economic Review: Insights. NBER Working Paper 33795, May 2025, revised November 2025.
• Dell’Acqua and colleagues: The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork, Organization Science, published online 12 June 2026.

Source versions and publication status checked on 23 October 2026.