How Far Should Workflows Be Redesigned After AI Adoption?

In Dillon et al.’s randomized experiment across 66 firms and 7,137 people, users of generative AI spent about two fewer hours on email each week, but there was no statistically detectable change in work volume or composition. [1] Email tasks under individual control became shorter, while meetings, approvals, and role allocation requiring coordination did not automatically change.

Decompose the workflow into intake, creation, verification, exception judgment, approval, confirmation, and transfer to the customer or next process. Track arrivals, completions, waiting, and returns in every step under the same case identifier. If one step becomes faster while confirmed volume and end-to-end time do not move, report a local improvement rather than a system improvement.

A static workload does not prove that the AI is weak. Demand may be fixed, downstream capacity may be unchanged, verification may become stricter, or saved time may be absorbed elsewhere. [1] Keep these competing explanations at each step and redesign the first location where congestion moved.

• FLOW: Map steps and owners from intake to confirmation
• QUEUE: Measure arrivals, completions, waiting, and age at every step
• REWORK: Record return origin, revision reason, and repeat passes
• OUTCOME: Use confirmed volume and end-to-end time, not local task time, as primary measures

Dell’Acqua et al. randomized 791 P&G professionals across AI and no-AI conditions and individual and two-person teams. Individuals with AI matched teams without AI, and AI reduced functional bias between R&D and commercial professionals. [2] Process decomposition also indicated that AI mainly improved idea generation, while human judgment retained value in evaluation and selection. [2]

This does not show that AI can replace every team. The study used a short product-development task in one firm and did not measure long-term implementation or customer outcomes. [2] It does support testing roles by candidate generation, evidence verification, comparative evaluation, exception judgment, and final approval instead of fixing them by job title.

For each sub-process, test human-only, AI-only, and combined arrangements where feasible. Measure quality, time, verifiability, failure impact, and handoff load. Where work remains human-owned, define what evidence must be reviewed, what triggers rejection, where work stops, and who receives the return—not merely a responsible-role label.

• GENERATE: Create candidates, drafts, and exploratory options
• VERIFY: Check evidence, numbers, rules, and missing inputs
• SELECT: Compare options against objectives and constraints
• EXCEPTION: Decide on nonstandard, high-impact, and unsupported cases
• APPROVE: Confirm the result and accept residual risk

Brynjolfsson et al. studied 5,172 customer-support agents and found an average 15% increase in issues resolved per hour with AI assistance. Gains were larger among less experienced and lower-skilled workers, while top performers had small speed gains and a slight quality decline. [3] The value of the same workflow placement can differ by worker and case.

Do not deploy the same arrangement to everyone. Segment by case difficulty, worker experience, availability of prior examples, verifiability, and failure impact. AI can transfer effective patterns to less experienced workers while adding little—or even interfering—when experts handle exceptions. [3] The average is not a mandate for uniform placement.

Update the boundary from errors and exceptions. Separate cases confirmed with little revision, held for missing evidence, substantially rewritten by experts, and associated with severe error. Review where the workflow arrangement held each month instead of using adoption rate as the success measure.

As an operating hypothesis, greater creation volume may increase arrivals to reviewers if the share sent for review is unchanged. Across three developer experiments and 4,867 people, the June 2025 author version reports an estimate, using assigned access as an instrument for actual use, of 26.08% more completed weekly pull requests from use, under additional assumptions; this is not the average effect of offering access, and firm-level estimates varied and experimental conditions differed. [4]

The current publisher abstract retains the reported number, but its full methods were not accessible for comparison with that author version. Follow work through review, testing, and merge rather than equating output with completion.

Put missing evidence, conflicting rules, severe impact, customer-specific conditions, and human-AI disagreement into an exception queue. As routine cases accelerate, harder cases can concentrate on experts and increase time per review. Track not only counts, but also difficulty, age, recurrence, and concentration on the same people.

When the bottleneck moves, test several responses: narrow upstream use, batch outputs, auto-confirm only low-impact segments, clarify review criteria, or improve information sent to experts. If throughput is sustained through reviewer overtime or unpaid work, record the result as workload transfer rather than improvement.

• ARRIVAL: Measure the volume and difficulty reaching the next step before and after AI
• SERVICE: Measure available review and approval capacity
• BACKLOG: Track exception count, age, recurrence, and reviewer concentration
• TRANSFER: Return load shifted to other teams, overtime, or vendors to the total

Time savings do not become organizational outcomes until they are reallocated. The absence of detectable work-composition change in Dillon et al. suggests that access at the individual level may not decide where capacity goes. [1] Brynjolfsson et al., by contrast, show conditions where assistance can connect to throughput and quality in a stable support process. [3]

Choose a destination in advance: unmet demand, customer interaction, quality verification, improvement work, learning, or workload reduction. Not all time must become throughput. Lower overtime and previously unaffordable quality checks can be outcomes. But do not count the same time simultaneously as labor-cost reduction and additional output without observing both realizations.

For example, a 30-minute reduction in drafting that is absorbed by another meeting does not shorten case lead time. If the time is reassigned to a larger verification sample and severe errors decline, the value is avoided quality loss rather than speed. Put destination, beneficiary, and confirmation method on the workflow change record and verify realization later.

• CAPACITY: Record the amount of time released and the role where it appears
• DESTINATION: Choose throughput, quality, customer work, learning, or workload reduction
• REALIZATION: Confirm that the destination outcome actually changed
• NO DOUBLE COUNT: Do not book one capacity gain as both cost reduction and added throughput

Workflow redesign is not completed by publishing a new procedure. Record the eligible segment, AI and human steps, inputs, verification, exceptions, approval, reallocation, and stop conditions as one version. Compare old and new procedures for representative cases, and move only the boundary that meets floors for end-to-end outcomes and quality and ceilings for exception age and workload.

The evidence includes broad access with no detectable work-structure shift, throughput gains in a bounded support process, and changed expertise boundaries in a short innovation task. [1] [2] [3] These findings are not contradictory; the work, technology, coordination, and unit of evaluation differ. Compare conditions rather than transplanting numerical effects.

At monthly workflow reviews, inspect confirmed volume, end-to-end time, quality floors, rework, exception age, expert load, and realized reallocation—not adoption rate. When improvement fails, vary one competing explanation at a time: eligible segment, handoff information, review capacity, or reallocation destination. Do not begin with a change that cannot be safely reversed.

• BASELINE: Preserve flow, quality, exceptions, and workload in the existing process
• VERSION: Record step allocation and boundary conditions as a workflow version
• PARALLEL: Compare old and new procedures on representative cases in the same period
• ROLLBACK: Restore segments that fall below the quality floor or exceed ceilings for exception age or expert load