The first draft arrived in minutes. Why did the work still take all afternoon?
A service team uses AI to draft a proposal. It includes the right broad service, a polished introduction and a plausible timeline. Then an account manager notices that the proposed delivery date was never agreed. Finance corrects a price. A colleague removes a promise that does not match the client's actual request. The draft helped them begin, but they still needed to produce something the client could rely on.
That distinction is easy to lose when teams measure “time to first output” rather than “time to useful result.” It applies to proposals, support replies, software changes, intake summaries and reports. Speed at one step may save time, move work to a reviewer, or create new work when a confident-looking mistake is difficult to spot.
Measure the complete task
In a January 2026 survey reported by Workday, nearly 40% of AI time savings in the study were lost to correcting, rewriting and checking outputs. The study covered 3,200 respondents across several regions. It is a survey finding, not a universal loss rate for every workplace.
The implication is to measure locally. If a task used to take four hours and a first draft now takes thirty minutes, the difference is not automatically three and a half hours. Count time spent finding source information, checking claims, seeking approval, repairing mistakes, communicating the result and resolving follow-up questions.
Where the work can move
Before generation: Someone still has to locate current facts. If policy, price and customer history disagree, no prompt can make the source reliable.
During review: The reviewer must distinguish accurate details from plausible inventions and decide which suggestions fit the real case. A polished format can make uncertain content feel finished.
At approval: An AI suggestion is not a business decision. The person with authority may need to approve a price, deadline, refund or customer commitment.
After delivery: A customer may ask a question that reveals the first response was incomplete. The team needs the earlier context and a way to correct the record.
A team can benefit from AI at several of these stages. It still needs to know which stage is actually slow and whether the final result improves.
The person reviewing needs support
Workday's separate October 2026 workforce report found that 79% of surveyed workers said they knew which skills they needed, while 66% said their employer helped them develop those skills. This is a different study, not another measure of rework. It raises a practical question: when a role changes from creating to checking more output, has the organization trained and equipped that reviewer?
Checking a client-facing proposal requires access to approved terms. Reviewing code requires understanding the application and its failure paths. Reviewing a summary requires the source record. Without those, “human in the loop” can mean one person carrying the uncertainty alone.
A small measurement exercise
Choose ten comparable instances of one recurring task. For each, record when information arrived, when a first draft was ready, time spent correcting it, time awaiting a decision and when the recipient received a useful result. Note the reason for each correction and whether the task reopened later.
Then ask what changed. Perhaps AI made drafting easier but approval remains the bottleneck. Perhaps people repeatedly search for the same policy. Perhaps the output is good enough for internal planning but needs much closer review before a customer sees it. Those are design findings, not a verdict on AI as a whole.
A better system might put approved information next to the task, show source evidence, mark a proposed decision as pending, or give the next person a complete handoff. My engineering portfolio includes working examples of reliability and workflow design. The goal is validated progress through the complete task.
Work with Massoda
I build software around the decisions, reviews and handoffs that make a result useful. To examine one workflow before adding automation, email hello@massoda.me.