Don't Book the AI Labour Saving Yet

AI is beginning to affect workforce plans before most organisations can show where the operating return has landed. S&P Global reports a modest negative net employment effect from AI over the past year, even though only 37% of the initiatives in its enterprise use-case research were live and delivering value. Just 46% were assessed as on track for positive return within 12 months. Those measures come from different survey components and do not prove that AI caused each job reduction. They do expose the sequencing risk: a cost action can arrive before the organisation has converted capability into a durable result. ¹

The decision for boards, CFOs and CHROs is immediate. Keep an AI-related labour saving out of the base plan until a named workflow has produced a measurable outcome with acceptable quality, risk and service. Treat time saved as potential capacity, not cash. Treat usage as activity, not return. Treat a workforce change as the final step in a conversion chain, not the opening assumption.

This is not an argument for slowing useful automation. It is a way to fund the deployments that genuinely change unit economics while preventing an optimistic productivity forecast from becoming an irreversible workforce decision.

Stage one is time saved. Establish the workflow, the baseline and the full task time before and after AI—including prompting, checking, escalation and rework. A European Commission survey across 18 member states found that employed AI users who reported time savings estimated an average of 7.4 hours a month. That is useful, but it is self-reported and applies to users reporting a saving, not every employee. ²

Stage two is capacity released. Ask whether the saved minutes combine into usable capacity or remain fragmented across people and days. Ten minutes removed from a task may improve experience without creating a schedulable hour, a shorter cycle or an avoidable vacancy. Managers need to redesign queues, roles and hand-offs before local efficiency becomes operating capacity.

Stage three is value captured. Name the destination: more cases resolved, faster revenue, lower external spend, fewer errors, better retention, shorter lead times or additional strategic work. The Commission's illustrative economy-wide productivity estimate assumes that reported time savings fully convert into output and explicitly describes the result as an upper bound. That caveat belongs in every operating plan. ²

Stage four is the workforce decision. Only after demand, quality, resilience and workload are visible should leaders decide whether to redeploy people, slow hiring, leave vacancies unfilled or change roles. A broad headcount target imposed earlier can remove the expertise needed to make the workflow work.

A survey of nearly 6,000 senior executives in the United States, United Kingdom, Germany and Australia found that 69% of firms actively used AI, while nine in ten reported no impact on employment or productivity over the previous three years. The same executives forecast that AI would raise productivity by 1.4%, lift output by 0.8% and reduce employment by 0.7% over the next three years. ³

The forecast may prove correct. The management error is using it twice: first to justify technology spend, then again as a booked labour saving before the operating change exists. Plans should show which part is observed, which part is forecast and which decision is reversible if conversion stalls.

There is strong counter-evidence against a blanket freeze. A separate survey of nearly 750 corporate executives found positive productivity gains concentrated in finance and high-skill services, associated with innovation and demand-oriented channels rather than capital deepening alone. ⁴ The implication is selective expansion: back the workflows with measured commercial outcomes and stop averaging them together with experiments.

Released capacity does not allocate itself. It can become customer growth, faster service, improved quality, innovation, lower contractor spend, reduced overtime, learning time or lower employment. Each choice has a different owner and a different financial result. If leaders do not choose, the dividend is likely to dissolve into more messages, more output of uncertain value and more supervision.

Use one outcome per workflow for the first 90 days. A service process might target resolved cases per paid hour while holding reopen rates and customer satisfaction. A sales process might target qualified pipeline per representative while holding complaint and discount rates. A software workflow might target lead time while holding escaped defects and operational incidents. The numerator and the guardrails should be agreed before rollout.

Then align the commercial terms. If a vendor charges by seat, token or action while the organisation carries integration, supervision and failure costs, activity can rise faster than value. Renewal should depend on the customer's unit economics, not the provider's usage dashboard.

S&P Global reports that larger firms increasingly expect negative employment effects while smaller and medium-sized firms still forecast positive effects. Its regional discussion describes more consolidation pressure in advanced markets and more labour complementarity and output growth in markets including Brazil, India and parts of Sub-Saharan Africa. ¹⁸ One global headcount assumption would erase the economics that matter.

High wages can strengthen the automation case, while scarce skills can make released capacity more valuable in growth. Labour law changes the timing and reversibility of role decisions. Language coverage changes error and supervision cost. Informal work, unreliable connectivity, weak integration capacity and public-service obligations change what can be captured safely. Global policy should specify the conversion stages; local management should supply the assumptions.

Australia is part of the four-country executive survey, but it is not the centre of this story or a proxy for the region. Australian leaders should test the same workflow economics under local employment, skills and service conditions. Leaders elsewhere should do the equivalent rather than importing a US, EU or Australian productivity rate.

For each material deployment, record the baseline task cost, adoption threshold, full supervision and rework time, capacity released, destination of that capacity, business outcome, quality guardrails and accountable executive. Separate observed results from forecasts. Review the record monthly until the workflow is stable.

Fund more when the outcome improves and the guardrails hold. Redesign when time is saved but capacity remains fragmented. Renegotiate when provider charges rise faster than customer value. Stop when supervision, errors or integration consume the gain. Delay a workforce reduction when the business has not yet proved where the released capacity went.

The organisations that gain will not be those that report the most AI hours or announce the largest productivity ambition. They will be those that can trace a faster task through operating capacity into a result the enterprise actually keeps—and make the workforce decision only after that chain is visible.