AI productivity: faster output is only half the business case

The easiest AI productivity gain to demonstrate is a faster first draft. The harder question is whether anyone finishes the job sooner.

A report produced in seconds can still spend hours in review. A support reply can arrive instantly and need three corrections. A team can generate more proposals without having the capacity to price, approve or deliver them. Speed at one stage is useful, but it does not settle the investment case.

For a CIO defending an AI budget to the finance director, the useful question is what happens to completed work. Erik Brynjolfsson, Danielle Li and Lindsey Raymond’s 2023 working-paper findings put that distinction into practice: customer-support agents with an AI assistant resolved 14% more issues per hour on average, with gains concentrated among less experienced workers. That is a result for a particular deployment and workforce, rather than a transferable percentage for every AI purchase.

The implication for an enterprise pilot is to measure the outcome and its distribution. A service-desk manager needs to know whether new starters close cases faster, whether experienced staff inherit more escalations and whether customers return with the same unresolved problem. A single average can conceal each of those changes.

The work has moved. Has it shrunk?

Consider a hypothetical team preparing customer reports. Before introducing AI, each report takes 40 minutes to draft and 20 minutes to check. With AI, drafting falls to 10 minutes, but checking rises to 35 because the reviewer must reconcile claims with the underlying records.

The drafting stage is 75% faster. The complete task is 25% faster: 45 minutes instead of 60. Both figures are mathematically correct, but only the second describes the labour released across that workflow.

Now suppose difficult cases need a further ten minutes of rework. Whether the pilot still looks attractive depends on how often those cases occur and who handles them. A specialist reviewer can become the new bottleneck even while the team’s average handling time falls.

This is why the pilot boundary matters. Starting the clock when someone opens the AI tool and stopping when text appears excludes the work most likely to challenge the result.

Quality belongs inside the productivity calculation

Compare completed work against the same acceptance standard. If an AI-assisted report passes with missing qualifications that would have sent a manually written report back for revision, the comparison has changed the product as well as the process.

That does not mean every output needs perfection. It means the business must decide what is good enough for the task. A draft for internal discussion and a customer-facing recommendation need different treatment. The AI evaluation set should reflect that difference.

Separate common cases from exceptions. An apparently strong average can conceal a workflow that handles routine requests well but transfers the most demanding work to a small number of experienced staff. Those people’s time is part of the cost, even if it sits in another department.

Released time is an opportunity, not a cash saving

In our example, 100 reports release 25 hours before additional rework. That is capacity. It becomes a financial saving only if something changes in expenditure, such as reduced overtime or a genuinely avoided hire. It becomes additional output only if there is demand and the next stage can absorb it.

The benefit may instead be a shorter queue, more attention to complex customers or less evening work. Those can be worthwhile outcomes. Present them on their own terms rather than converting every recovered minute into a theoretical salary saving.

A useful question for the programme sponsor is: what will we do differently if this time becomes available? An answer such as “clear the backlog of renewal reviews” can be tested. “Improve efficiency” leaves the benefit unresolved.

A stronger pilot follows the whole task

Use comparable work, retain the original quality threshold and include time spent preparing inputs, checking outputs and correcting failures. Observe adoption as well as availability: a licensed tool that people avoid has a different business case from one that becomes part of daily work.

Keep the commercial costs beside that workflow evidence. Subscription fees, integration and ongoing support can outweigh a small handling-time improvement; a modest time gain can still be compelling at sufficient volume. The cost per completed AI task is more revealing than either licence cost or generation speed alone.

The best result is not necessarily the largest percentage on a slide. It is a repeatable improvement that survives ordinary workloads, ordinary users and the full cost of getting the work accepted. That is the point at which a promising demonstration becomes a business case.