Local task productivity creates enterprise value only when the end-to-end process can absorb the gain.
McKinsey’s latest global AI survey creates a useful tension for anyone responsible for process improvement. Eighty percent of respondents said AI had improved their individual productivity, and half said it helped them make better decisions. Yet only 37% said AI had contributed positively to their organisation’s EBIT, essentially unchanged from the previous year. The survey was conducted from May 4 to June 8, 2026 and gathered 1,719 responses across 97 nations (McKinsey & Company, 2026).
Those figures do not prove that individual productivity gains are being wasted, and a self-reported survey cannot establish which management practice caused each financial result. They do show why task-level efficiency should not be confused with end-to-end process improvement. A person may prepare an analysis faster, draft a document faster or resolve a routine query faster while the complete process still waits for approvals, repeats checks, accumulates exceptions or sends work back for rework.
The practical question is therefore not whether AI makes a task faster. It is whether the organisation can convert that faster task into shorter lead time, higher throughput, better quality, lower cost, improved customer outcomes or reduced risk. That conversion requires process design, measurement and control. It is also where Process Improvement Training, Continuous Improvement Training and Kaizen Training become directly relevant to the way organisations use AI.
What McKinsey’s 2026 Survey Actually Shows
Individual productivity is not process performance
A process is a chain of work, decisions, queues and handoffs that produces an outcome. Improving one activity can help the process, but the effect depends on what happens before and after that activity. If AI reduces a ten-minute task to two minutes while the case then waits two days for review, the task has improved but the customer lead time has barely moved. If faster drafting increases the number of items reaching an already constrained approval team, the local gain can even make the queue more visible without increasing completed output.
This distinction matters because individual users naturally experience the part of the process they touch. They can feel the time saved. Enterprise value depends on the entire flow. The relevant baseline should therefore include total elapsed time, active work time, queue time, rework, first-pass quality, demand, capacity and the final business outcome. Measuring only prompts, licences, users or task minutes risks rewarding local optimisation instead of process performance.
High performers redesign workflows rather than insert tools
McKinsey reported that AI high performers represented about 6% of respondents. Nearly three-quarters of those high performers said they were fundamentally redesigning workflows because of AI, compared with about one-quarter of other respondents. High performers were also more likely to have defined processes for measuring the impact of AI initiatives (McKinsey & Company, 2026). These are associations rather than proof of causation, but they point toward a familiar operational principle: sustained value is more likely when technology changes the operating method, not merely the speed of an existing task.
Workflow redesign means deciding where an AI output enters the process, who owns the next action, what information is sufficient to proceed, which situations require human judgement, what happens when the tool is unavailable and how exceptions are recorded. It can also mean removing steps that existed only because information used to be difficult to obtain. If the old process remains untouched, employees may keep the old checks, add new AI checks and create a faster but more complicated route.
Measure the value chain from task to outcome
A useful benefits method links four levels of evidence. The first is the task effect: time saved, error reduced or decision support improved. The second is the process effect: lead time, throughput, first-pass yield, backlog, cost per completed case or service level. The third is the capacity effect: whether the released time is actually redeployed to valuable work or allows the same demand to be handled with less effort. The fourth is the business effect: revenue, margin, customer retention, working capital, risk or another outcome the organisation is trying to improve.
The links between these levels should be explicit. Saving one hour of labour is not automatically one hour of financial benefit, because the employee may remain fully paid and the freed capacity may not be used. Faster case preparation is not automatically faster customer service if downstream queues remain unchanged. Process improvement makes these assumptions visible so leaders can test them rather than count theoretical savings.
Convert AI Task Gains Into End-to-End Process Improvement
Find the constraint before automating more work
The first step is to identify what currently limits the process outcome. The constraint may be a specialist review, a policy decision, data quality, equipment capacity, customer response time, an approval threshold or a recurring defect. AI should be evaluated against that condition. Automating work upstream of the constraint can still be worthwhile, but the improvement case should explain how the gain reaches the constraint or frees capacity for another valuable purpose.
This is where a current-state process map is more useful than a list of AI use cases. Map the trigger, major activities, decisions, queues, handoffs, exceptions and completion point. Add actual timing and volume data where available. The map often reveals that the largest delay sits between tasks rather than inside them. That observation changes the question from ‘Where can AI save minutes?’ to ‘Which delay or failure prevents the process from producing more value?’
Redesign handoffs, decisions and exception rules
AI frequently changes information faster than it changes authority. A model may produce a recommendation immediately, but someone still needs permission to act. If every recommendation follows the old approval route, the process can remain slow. The redesigned process should specify which decisions can be made automatically, which can be made by the frontline role, which require specialist review and which require senior approval. The level of control should reflect risk, reversibility and consequence rather than organisational habit.
Exception design matters just as much. Teams should define what constitutes a normal case, what evidence triggers review, how uncertain outputs are handled, who resolves conflicts between AI and direct operational evidence, and how recurring exceptions become improvement opportunities. Without this structure, people create local workarounds. Variation then grows across teams even though everyone is using the same technology.
Build an end-to-end measurement chain
Before implementation, define the process measure that should move and the protections that must not deteriorate. A service process might target shorter end-to-end resolution time while protecting first-contact quality and complaint rates. A finance process might target faster close activities while protecting control failures and adjustment rates. A manufacturing process might target greater throughput while protecting scrap, safety and maintenance stability. This avoids celebrating speed that creates downstream cost.
End-to-end measurement follows the work beyond the faster task to the process result and verified outcome.
After implementation, compare the new process with the baseline under comparable conditions. Look at variation, not only averages. Confirm whether queues moved, whether rework appeared elsewhere and whether demand changed. When a task becomes faster, the next process step should be deliberately checked for congestion. That simple discipline prevents organisations from calling a local productivity gain a complete improvement.
If you want to practise this type of end-to-end analysis on a real process, STEP Bootcamp provides practical Process Improvement Training through live classes, applied assignments and personalised feedback. The programme is designed to help you move from problem definition and analysis to improvement, value measurement and control.
Make Continuous Improvement Part of AI Operations
Standardise the proven method, not the first method
A successful pilot should produce a better operating method, not only a successful demonstration. Once evidence shows that the redesigned process works, document the roles, decision rules, inputs, outputs, escalation paths and measures required to reproduce it. Train the people who operate the process and make the standard easy to access. The purpose of standard work is to create a reliable baseline from which the next improvement can be judged.
Standardising too early can freeze weak assumptions. Teams should first test the method under ordinary variation, including peak demand, incomplete data, unusual cases and tool downtime. They should also confirm that the intended control is proportionate. A low-risk, reversible decision may not need the same review structure as a safety, legal or major financial decision. Process discipline should remove unnecessary complexity, not formalise it.
Use Kaizen to improve the human-AI process itself
AI-enabled work will continue to change after launch. Models change, prompts evolve, policies are updated, users develop new habits and demand shifts. Kaizen provides a practical way to improve the combined human-and-technology process in small increments. Teams can review recurring exceptions, override reasons, rejected outputs, repeated delays, rework and customer feedback, then adjust the standard method when evidence supports a better approach.
This also creates a more useful conversation about adoption. Instead of asking whether employees are ‘using AI enough’, leaders can ask where the process still loses time or quality and what change would address it. Usage may be part of the answer, but so might decision rights, missing data, training, role design or a redundant control. Continuous improvement keeps the focus on the process outcome rather than the novelty of the tool.
Build capability to challenge local optimisation
McKinsey’s survey suggests that organisations are deepening AI use even while enterprise financial impact remains concentrated. The operational response should not be to slow useful experimentation. It should be to raise the standard of evidence. Professionals need to be able to distinguish a faster activity from a better process, identify where the constraint really sits, redesign the surrounding workflow and measure whether value survives from task to outcome.
That capability is useful far beyond AI. The same logic applies to automation, new software, reorganisations and ordinary Kaizen activity. A local change becomes process improvement only when the end-to-end system performs better and the gain can be sustained. For professionals who want applied capability rather than tool vocabulary, STEP Bootcamp combines Process Improvement Training, Continuous Improvement Training and Kaizen Training in one 28-day programme with live application and feedback.
References
McKinsey & Company. (2026, August 25). The state of AI in 2026: On the road to ROI. View source
Leave a Reply