Buying an AI tool does not create productivity. Productivity appears when the organisation changes how work moves, who makes decisions and how quality is controlled.

That distinction matters because AI investment is accelerating while many organisations are still measuring access, licences or time saved. Those figures describe activity. They do not show whether customer response, management decisions, cost or delivery reliability have improved.

The OECD Economic Outlook published on 23 September 2026 makes the economic case carefully. It says AI may produce substantial productivity gains once it is effectively integrated into production processes. It also warns that returns may disappoint or take longer than expected.

For leaders, the practical message is simple: the operating model sits between the technology and the result.

What recent evidence does and does not tell us

On 29 September, Microsoft published a regional summary of its 2026 Work Trend Index. Among the AI users surveyed, 66% said AI allowed them to spend more time on higher-value work and 58% said they were producing work they could not have produced a year earlier.

The same research found that organisational factors were more strongly associated with reported AI impact than individual factors. Microsoft is clear that this is an association, not proof of causation. The underlying survey also covers people who already use AI, so it should not be treated as a measure of adoption across the whole workforce.

Even with those limitations, the findings point to a useful management question. If capable people have access to capable tools but the result is inconsistent, what in the surrounding system is holding them back?

Usually, the constraint is not prompting technique. It is one or more of these:

  • the process has no agreed outcome or owner
  • teams automate different versions of the same work
  • data is incomplete, inaccessible or poorly defined
  • nobody has set the boundary between automated action and human judgement
  • quality review happens informally and cannot keep pace with volume
  • savings in one task create extra checking or rework elsewhere
  • leaders count usage without measuring the operational result

Move from tool deployment to workflow redesign

A licence rollout asks, “Who should have access?” Workflow redesign asks, “What should happen differently from request to outcome?”

Take management reporting. An assistant may draft commentary or summarise a dashboard. That saves little if teams still argue about metric definitions, submit data late and maintain competing spreadsheets. The better sequence is to agree the management question, define the data and ownership, redesign the reporting cycle and then decide where AI reduces effort or improves judgement.

The same principle applies to customer service, project reporting, document review and internal approvals. Follow the whole workflow, including exceptions and handovers. Improving one task can make the end-to-end process worse if it sends more low-quality work to the next stage.

Before implementation, use the AI workflow readiness test to check the process, data and ownership. After implementation, review the full workflow again. The operating problem may have moved.

Five decisions leaders need to make

1. Define the business outcome

Choose an outcome a leadership team would care about without mentioning AI. Examples include shorter case resolution time, fewer reporting errors, faster approval of viable projects or more capacity for customer-facing work.

Record the baseline before the pilot. If no useful baseline exists, that is the first management information problem to solve.

2. Redesign the work, not only the task

Map the current process and the proposed process. Remove steps that exist only because information is copied between systems. Decide what the AI will prepare, recommend or execute and what people will stop doing.

Do not preserve every control automatically. Some controls exist to compensate for manual error. Others remain essential because they protect customers, money, safety or accountability.

3. Set decision rights and exceptions

State which outputs require review, who can approve them and when the workflow must escalate. A vague instruction to “keep a human in the loop” is not a control model.

Use thresholds. A routine, reversible and low-risk action may be automated. An unusual case, weak-confidence result or material commitment should move to a named decision-maker with the evidence needed to decide.

4. Give the workflow an operational owner

Technology can own the platform. It cannot own the business outcome alone. The operational owner should monitor performance, review exceptions, approve changes and decide when the workflow needs to be paused or redesigned.

This is also where local experiments become organisational capability. Teams need one place to record what failed, what changed and which version of the workflow is approved.

5. Measure net value

Measure the result across the process, not the speed of the AI-enabled step.

Weak measure Better operational measure
Number of licences Active use in an approved workflow
Prompts submitted Cases completed to the required standard
Minutes saved on drafting End-to-end cycle time
Outputs generated First-time-right rate and rework
Number of agents Cost, capacity or service improvement after control effort

Include the cost of review, correction, integration and governance. A faster first draft is not a productivity gain if it creates more work downstream.

A practical review for the next leadership meeting

Ask each AI initiative owner to bring one page covering:

  1. the business outcome and current baseline
  2. the workflow before and after the change
  3. the roles of the system, the user and the approver
  4. the most important exception and escalation route
  5. the measure that will decide whether to scale, change or stop

If the team cannot answer those points, it is too early to discuss enterprise rollout. The next step is process clarification, not another tool demonstration.

Where LUKiN fits

LUKiN helps organisations turn digital ambition into a workable operating system. That includes following the real process, clarifying ownership, designing governance and building reporting that shows whether the change is producing value.

The work is deliberately practical. A recent engagement connected tools, teams and workflows so that delivery could scale with clearer visibility and control. Read the digital transformation case study for the problem, approach and result.

If several AI initiatives are already running without one view of outcomes, dependencies or ownership, digital transformation delivery can provide the structure needed to move from experiments to reliable execution. If the constraint is still unclear, start with the Operational Performance Diagnostic.

Sources