AI Productivity Gains Are Leaking Before They Reach the P&L
AI can save task time without creating business ROI. Operators need workflow capture, not generic tool usage.
AI productivity is not the same as AI ROI.
That distinction matters more than most tool demos admit. A new Okane Land explainer points to a harder version of the AI-at-work story: real workers can save time with chatbots, but the business often fails to capture much of the value. The underlying workplace research from Anders Humlum and Emilie Vestergaard covered roughly 25,000 workers across 7,000 Danish workplaces and found average time savings of 2.8% of work hours for users. Useful, yes. But the same research found no significant impact on earnings or recorded working hours, and only 3% to 7% of productivity gains passing through to wages.
That is the bit operators should sit with.
The easy interpretation is that AI is overhyped. That is too lazy. Other studies still show real task-level gains. The NBER and QJE customer-support study found that a generative AI assistant increased issues resolved per hour by about 14% to 15% on average, with especially strong gains for less experienced agents. The BCG and Harvard work on the “jagged frontier” found that consultants using GPT-4 completed more tasks, moved faster, and produced higher-quality work when the task sat inside the tool’s capability zone.
Both stories can be true.
AI can make a task faster and still fail to improve the business. That is not a contradiction. It is an operating-model problem.
Task Speed Is Not System Performance
A controlled study measures a task: write this response, solve this case, produce this draft, finish this analysis. A company measures a system: handoffs, reviews, customer queues, meetings, incentives, quality checks, capacity planning, and whether anyone changes what happens after the time is saved. The speed gain can evaporate inside that system.
Take a simple support workflow. If AI helps an agent draft replies faster, the company only captures value if something downstream changes. Maybe the agent handles more tickets. Maybe customers wait less. Maybe quality improves because the agent spends the saved time checking edge cases. Maybe the team keeps headcount flat while volume rises. Those are captured gains.
But if the saved minutes turn into more internal messages, longer review loops, vague “better work,” or a slightly calmer day that nobody measures, the P&L will not move. The person may feel helped. The business may still look the same.
This is where AI adoption has to grow up.
Too many teams still roll out AI as a tool-access project: buy seats, encourage usage, share prompt tips, wait for productivity to appear. That approach can create scattered convenience, but it rarely creates clean ROI. Convenience is not a strategy. Usage is not a result.
Build for Workflow Capture
The better approach is workflow capture.
Start with one workflow where the work is frequent, measurable, and painful. Customer replies. Proposal drafts. Internal reporting. Invoice exception handling. Sales research. QA summaries. Content repurposing. Pick something with enough repetition that small improvements compound.
Set the baseline first
Before AI touches the workflow, measure how long it takes, how many items move through it each week, the error rate, where review happens, what causes rework, and what a good outcome looks like.
Give AI a narrow job
Draft the first response. Summarize the call. Extract the fields. Prepare the checklist. Flag exceptions. Do not ask it to “make the team more productive.” That is not an instruction. That is a wish.
Assign the saved capacity
This is the missing step. If a workflow saves five hours a week, who owns those hours? Are they turned into faster turnaround, higher volume, deeper review, better customer follow-up, or lower external spend? If nobody can answer that, the gain is already leaking.
Measure the business result
Track the outcome that justified the change: shorter response times, higher throughput, lower cost, fewer errors, or better conversion. Tool usage is diagnostic data. It is not the result.
The Risk Is Vague Adoption
The research is a warning against vague adoption. It is not a warning against AI.
AI can absolutely improve knowledge work. But the business value does not arrive automatically because a worker finishes a draft faster. It arrives when the company redesigns the workflow around that speed, protects quality, and measures the result.
The operator takeaway is simple: stop asking whether AI saves time in general. Ask where the saved time goes in your company.
If you cannot trace the path from tool use to business outcome, you do not have an AI ROI problem. You have a workflow design problem.