AI Productivity Is a Workflow Question
An impressive demonstration cannot tell a business whether its whole operation will become faster, better or cheaper.
Start with the unit of work
A model can draft a response quickly while the complete task remains slow. Someone still needs to obtain the right information, check the answer, secure approval and deliver it. Productivity belongs to that entire chain. Counting generated words or completed prompts creates an easy metric with an uncertain relationship to useful output. A business needs a definition of completion that includes quality and the consequences of error.
Evidence with a boundary
The NBER working paper Generative AI at Work studied the introduction of an assistant among customer-support agents and found uneven productivity effects, with larger benefits for less experienced workers. This is evidence from a particular working environment, not proof that every occupation receives the same gain. Its relevance lies partly in that variation: the same tool can perform different economic functions for different people.
The review bottleneck
Suppose a team can produce ten proposals in the time previously needed for three. If a specialist must review every claim, the bottleneck may move rather than disappear. An organisation can respond by improving source material, narrowing permitted claims and routing exceptional cases to experts. Simply asking the reviewer to work faster undermines the quality constraint that made the original output useful.
Separate assistance from authority
A draft, a recommendation and an executed action have different consequences. The approval needed for each should follow those consequences. Low-impact drafts can tolerate flexible experimentation. Actions involving customer commitments or irreversible changes require clearer boundaries and records. The point is to give people control where it matters while allowing routine work to move without unnecessary ceremony.
Measure the work after adoption
A useful pilot compares completed cases, rework, time to resolution and error severity. It should include representative difficult cases rather than only demonstrations selected by enthusiasts. Measure the burden on adjacent teams as well: a faster front office can create more exceptions for operations. Adoption is successful when the whole process improves, not when one department reports a flattering local result.
Define the pilot before the demonstration
Choose a bounded task and record its present completion time, error rate and review burden. Decide how output will be sampled and who can judge the difficult cases. Run the comparison over a period long enough to include ordinary variation. A result obtained from only simple cases is useful for those cases; it should not be presented as a general operating improvement. Where possible, compare similar work and separate the effects of the tool from simultaneous changes in staffing, training or customer demand. Otherwise the organisation may attribute an improvement to the wrong intervention.
Watch for work that disappears from the metric
A support agent may close cases sooner while customers need to contact the company again. A drafting team may deliver more documents while reviewers spend evenings correcting them. Include these displaced costs in the decision. The appropriate measure is useful work completed to a stated quality standard, not merely visible activity. If a pilot improves one metric and worsens another, explain the trade-off rather than compressing it into a single success claim. An honest mixed result can guide a better second iteration.
The management implication
The editorial inference is that workflow design can matter as much as model capability. A strong tool placed inside contradictory processes may create little value. A narrower tool with dependable context and a well-defined handoff can be useful immediately. Before expanding a pilot, identify the exact task, the source of truth, the reviewer and the standard of completion. Those details turn a technical capability into an operating advantage.
Sources & further reading
NBER · Generative AI at Work, Working Paper 31161 ↗
Editorial photograph for illustration.
Continue reading
The Most Valuable AI Skill May Be Knowing What to Check
As production becomes easier, organisations need a disciplined way to separate plausible output from reliable work.
When intelligence becomes cheap, judgment becomes expensive
Producing an answer and accepting responsibility for it are different economic activities.
The Real Cost of a World That Trades in Blocs
Geopolitical alignment changes more than trade routes. It changes who pays for resilience, who carries the adjustment and who can afford alternatives.