Your team can prepare a draft faster than it could a few months ago. The first version arrives quickly, and there is less time spent staring at a blank page. Yet the manager reviewing it still has a full afternoon of checking details, resolving questions and following up.
That can leave a business owner with an awkward question. The team is using AI, and parts of the work are clearly moving faster, so why does the business still need so much attention to get the job finished?
One place to look is the boundary around what you are measuring. If you count the time it takes to produce the draft, you see the benefit at that step. Following the work through review and acceptance tells you more about what the business actually gained.
Follow the work past the first draft
Consider a hypothetical distributor using AI to help prepare customer quotes. A coordinator supplies the request and available product information, and the tool organizes a draft. The coordinator has less typing to do and can prepare more quotes during the morning.
Before a quote can go out, however, a sales manager checks pricing, availability and any unusual terms. Some drafts need corrections because the source information is incomplete. Others sit while the manager confirms a delivery commitment with operations, and a few return to the owner because nobody else is sure which exception is acceptable.
AI may still be helping this business. But the coordinator's experience tells only part of the story. The manager's review, the corrections and the owner's decisions are also part of getting a usable quote to the customer.
Last week's article focused on defining what finished means before assigning work to AI. The next question is how much effort that finished result requires. Once the team agrees on an acceptable quote, it has a sensible endpoint for examining the work.
The evidence gives us a reason to measure
There is credible evidence that AI can improve useful output. A study published in the Quarterly Journal of Economics, using data from 5,172 customer-support agents at one business-software company, found that access to an AI assistant increased issues resolved per hour by 15% on average. Results differed across workers, and that finding does not establish what a distributor or another business should expect. Brynjolfsson, Li and Raymond, “Generative AI at Work.”
What I find useful about that measure is its connection to the job: resolving customer issues. It gives the business a more meaningful question than how quickly an assistant can generate a response. For your own workflow, the equivalent might be an accurate quote ready to send or a completed order the next team can use.
There is also a reason to pay attention to the people receiving all that output. In its October 2026 report on its own workplace, Anthropic describes growing demands on human attention and supervision as employees work with AI agents. This is a vendor's internal observational account of an unusual, fast-growing company; it cannot show that AI causes the same pattern in other businesses. It does make the review burden a useful question to examine locally. Anthropic, “How employment is changing at Anthropic.”
METR makes another helpful distinction in its May 2026 research: the speed of doing work and the value of that work are different questions. Its survey of technical workers relies on self-reports from a convenience sample, so the results are not measured business returns. The distinction still helps frame an operating decision: faster work matters when it improves something the business needs. METR, “Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity.”
Start with five comparable cases
Choose one recurring workflow that already has an agreed standard for completion. Follow five comparable cases from the initial request to the accepted result, and ask the people involved to keep a simple log. Five cases are a manageable starting point for a conversation, not proof of an AI return on investment.
For each case, record:
-
The type of work: Note the complexity, missing information and unusual requirements so you can recognize important differences between cases.
-
Human preparation time: Include gathering inputs, giving instructions and preparing the first draft.
-
Human checking and correction time: Include everyone who reviews, repairs or clarifies the work, including the owner.
-
Elapsed waiting time: Record delays separately, along with what the work was waiting for.
-
The accepted result: Note whether it met the agreed standard, what needed changing and any later issue discovered.
-
Owner or manager interruptions: Record who was needed and why. Their active minutes belong in the effort log too, so do not count those minutes twice.
Keep human effort and elapsed time separate. If a quote waits two hours for a five-minute approval, that is a two-hour delay and five minutes of human work. If two people spend ten minutes checking it together, that is twenty person-minutes, even though only ten minutes pass on the clock. Agent runtime belongs in a separate record if you need it; it is not employee time.
Compare the cases with similar work from before the change, where you have reliable records, or establish a comparable baseline using your existing process. Keep the quality standard consistent and note differences in workload, staff experience and case complexity. If the baseline is uncertain, say so and collect more observations before making a larger decision.
Use the pattern to choose the next change
Return to the distributor. If the log shows less preparation effort but repeated corrections to the same pricing information, the next useful change may be fixing that source. If completed quotes wait for the owner's interpretation of routine exceptions, the team may need clearer decision rules and an agreed approval owner.
The log may also show that the workflow is working well. Perhaps the team is completing acceptable quotes with less total effort and fewer interruptions. That is a reason to keep observing the improvement and consider where the approach might fit next, while accounting for setup, training and tool costs before making a financial claim.
In the Execution Gap Framework, Operational Intelligence means gaining visibility into how work actually flows. This small exercise puts that idea into practice. It helps you see where time is being recovered, where work is moving to someone else and where one unclear decision keeps bringing the job back to you.
Recovered time also deserves a deliberate use. A manager might spend it coaching a newer employee, resolving a persistent customer issue or working on a project that has been waiting. An improvement becomes more useful when the team knows what the added capacity is for.
Start this week with one recurring job and five comparable cases. Review the record with the people who do and receive the work, then choose one next step: keep the useful change, address a source of rework or gather better evidence before expanding it.
If that conversation reveals broader dependence on you to keep work moving, the Execution Gap Assessment offers a structured way to reflect on those operating dependencies.