AI Productivity: Stop Counting Drafts and Start Counting Finished Work

Your team can finish the first draft faster and still leave work later. The missing measure is what happens after someone presses generate.

Longer workdays are a warning, not a verdict

An AI assistant can shorten a task without shortening the working day. That distinction matters when a business owner sees a fast demonstration and assumes the team has recovered capacity. A draft that takes seconds can still create an afternoon of reviewing, comparing and chasing approvals. The question is not whether the tool generates work quickly. It is whether useful work reaches its destination with less total effort.

ActivTrak's 29 October 2026 analysis makes the distinction difficult to ignore. The workforce analytics supplier examined digital work patterns from 9,140 knowledge workers across 467 enterprises, collected between 1 April and 25 September 2026. It reported workday spans of 8.4 to 9.1 hours among AI users, against 7.6 hours among non-users. That is a difference of up to 90 minutes, not evidence that every AI user works an extra 90 minutes.

Treat this as an observational finding from a supplier with a commercial interest in workforce measurement. People with heavier workloads may be more likely to adopt AI. Roles, responsibilities and working patterns can differ between groups. A workday span is also not the same as uninterrupted working time. The study does not establish that AI caused longer days or lower productivity.

For a UK business, the useful response is therefore neither to cancel AI nor to celebrate adoption. It is to test whether the particular workflow is expanding. Are staff producing more versions because producing versions became cheap? Has the manager become the bottleneck? Are people finishing customer work sooner, or simply carrying a larger pile of almost-finished material? Those questions turn an alarming headline into a practical management check.

More AI activity does not prove that your team has recovered capacity.

Count accepted work, not generated work

Start with a definition of finished that someone outside the AI pilot can recognise. For a quotation, finished might mean a correctly priced proposal approved for sending. For support, it might mean a resolved case that does not reopen within your chosen review period. For a management report, it might mean a decision-ready document with checked figures, not a polished first draft.

This avoids a common accounting mistake. If a salesperson produces ten draft proposals instead of four, output appears to have risen dramatically. If only four can be reviewed, the customer-facing result has not changed. The other six may be useful future work, unnecessary inventory or a new source of confusion. Until the business decides which, they should not be booked as realised productivity.

Worklytics' September guide, updated in October separates access, activity and outcome. It also explicitly labels its example dashboard figures as illustrative rather than audited client benchmarks. That distinction is worth copying. Your own pilot should label estimates, observed measures and commercial outcomes separately instead of presenting them as interchangeable proof.

What this means in practice is a small shared record, not a new analytics platform. Use the CRM, ticketing system or spreadsheet you already have. Record the number of requests received, outputs accepted, items returned for correction and items still awaiting approval. Add total handling time across the people involved, plus elapsed time from request to acceptance. Keep the definition stable through the comparison.

An increase in accepted work with stable quality and reasonable hours is a meaningful gain. An increase in drafts with a longer approval backlog is a different result. Both can be true at once. Looking at the second result honestly gives you a chance to fix the constraint before buying more licences or promising customers faster delivery.

A draft is inventory until someone can use it.

Look for work being moved to the next person

The easiest saving to see belongs to the person pressing generate. The hardest cost to see often belongs to the colleague receiving the result. That person must reconstruct missing context, verify claims and decide whether the material answers the original question. When the cost moves between departments, both teams can report apparently sensible numbers while the business becomes slower.

RAD Intel's October analysis makes this downstream burden the centre of its argument. It is a commercial commentary, not an independent trial, but its practical distinction is useful: faster creation and faster completion are different outcomes. A report may arrive earlier while the decision it supports arrives later because reviewers now face more material of uncertain quality.

Consider an illustrative account-management workflow. Before AI, preparing one customer update takes 30 minutes and review takes ten. After AI, preparation takes ten minutes but review takes 25 because figures and recommendations need checking. The gross drafting saving is 20 minutes; the net handling saving is five. If five additional versions are generated and circulated, the team may lose that saving altogether. These are example numbers, not research findings or a client result.

What this means in practice is to ask the receiving team to help define the pilot. Give them a short rejection checklist: missing source, wrong customer, unsupported figure, unclear action or unnecessary length. Let them return work that fails it. Do not make the recipient silently absorb the verification burden to protect an adoption target.

Set a work-in-progress limit as well. If the reviewer can approve five proposals today, generating fifty is unlikely to help unless there is a clear reason to build stock. Agree the priority order and require new drafts to have an identified user. Cheap generation should make work easier to finish, not easier to dump on someone else.

The person saving time and the person paying for the mistake may be different people.

Protect focus without turning the pilot into surveillance

Workflow friction can appear between tasks as well as inside them. ActivTrak reported that heavy AI users had average uninterrupted focus sessions of 37.8 minutes, compared with 48.2 minutes for light users. It also reported lower application switching among people using embedded AI than among those using browser-based tools. Neither comparison proves that moving AI into an existing application will solve your team's problem.

The sensible test is narrower. Watch the workflow with the people doing it. Does someone copy a customer request into a chatbot, paste the answer into a document, retrieve a missing fact from the CRM, return to the chatbot and then reformat everything for email? A shorter answer-generation time can conceal a longer sequence of handovers. A template or a simpler process might remove more effort than a different model.

Do not respond by installing intrusive employee tracking by default. UK employers still need to consider data protection when measurement involves identifiable staff. The ICO guidance on monitoring workers says employers must be clear about their purpose and choose the least intrusive means of achieving it. It discusses lawful basis, fairness, transparency and impact assessments. The ICO also flags this guidance as under review following the Data (Use and Access) Act, so check the current position before implementing monitoring.

A practical first measurement can be a time-limited, openly explained task sample using existing work records. Collect only what you need to understand completion time and rework. Restrict access, set a retention period and avoid capturing prompt contents or private communications without a justified need. Team-level reporting can reduce intrusion, but small groups can still reveal individuals.

Ask staff where the interruptions occur and what would remove them. Their explanation may be more useful than another dashboard. Measure the process to improve the process, not to turn AI usage minutes into a performance ranking.

Measure the handovers before measuring the people.

Run a pilot that names where the saved time will go

The strongest counterargument is straightforward: a longer working day does not mean AI has failed. A growing business may deliberately use the tool to serve more customers, investigate more options or improve quality. That can be valuable. The mistake is calling every expansion a time saving and then adding the supposed saving to the next workload forecast.

Before a pilot begins, choose the intended benefit. Is it reduced overtime, more accepted work within normal hours, faster customer turnaround or better quality at the same cost? Pick one primary result and a small number of safeguards. If quality improves while speed stays the same, describe it as a quality gain. If more work is completed but hours increase too, calculate the cost honestly rather than presenting additional effort as free capacity.

For a first comparison, choose a bounded workflow with enough repeatable work to observe. Record a baseline, then test AI on comparable tasks. Separate easy cases from difficult ones and note changes in demand, staffing and approval availability. A short pilot is a management signal, not a universal causal claim. Extend it if the sample is too small or the work is seasonal.

Review accepted output, corrections, total handling time and the oldest unfinished items together. Include tool costs and an appropriate value for the human effort involved. Do not count saved staff minutes as a cash saving unless they reduce actual expenditure. Capacity that is genuinely redeployed to worthwhile work is a different benefit and should have a named destination.

Finish with a decision: continue, change the process, narrow the use case or stop. Agree who owns the next action. If the pilot works, protect some of the recovered capacity instead of immediately filling it with additional drafts. AI earns its place when the business can explain what improved, who benefited and what stopped consuming time. The number of prompts submitted is not that explanation.

Recovered capacity needs a destination, not another pile of drafts.