AI Productivity Is Not a Headcount Calculation

Steven Gabris

Executive summary

AI is producing measurable time savings across operational and knowledge work. That is useful evidence of efficiency, but it does not prove that an organisation has realised a financial or operational benefit.

A saving spread across selected tasks, roles and different points in a workflow does not automatically create an equivalent amount of removable headcount. Organisations need to establish where usable capacity has actually been released, decide what they intend to do with it, and then demonstrate the resulting change in cost, throughput, workforce demand, delivery or quality.

The productivity evidence is becoming harder to ignore

BCG's 2026 AI at Work research found that 42 per cent of regular frontline AI users reported saving at least eight hours each week. The same research found that 66 per cent received limited or no guidance about what to do with that time, while more than half were not redirecting it into more strategic work.[1]

The investment results are less convincing. BCG reported in August that only 6 per cent of companies were seeing meaningful value from AI measured through reduced costs or increased revenue. Its analysis identifies a familiar gap: organisations measure activities automated or hours saved without connecting them to a financial outcome.[2]

Gartner found a similar imbalance in finance. In a survey of 204 finance leaders, 45 per cent said their AI investments leaned towards productivity, while only 20 per cent said they leaned towards decision quality. Gartner argued that CFOs should measure enterprise impact rather than pilots launched, activity completed or hours saved.[3]

There is an additional problem between measuring efficiency and claiming value. An organisation can accurately measure a task-level time saving and still misunderstand the capacity it has created.

The task is usually the unit of AI efficiency. The role is the unit used for workforce planning. The financial or operational outcome is the unit that matters for benefit realisation.

The benefit chain

I would separate four stages when assessing the value created by AI.

Stage	What should be demonstrated
Task-level time saving	A defined activity consistently requires less effort or elapsed time.
Usable capacity	The saving creates time that can be used elsewhere in the role, team or process.
Management action	The organisation decides whether to retain, redirect, consolidate or remove that capacity.
Realised benefit	The intended financial, operational or delivery outcome occurs and can be evidenced.

These stages require different evidence.

If a task that previously required five hours can consistently be completed in three, there is reasonable evidence of an efficiency improvement. It does not follow that the employee has two hours available every week, that the team can absorb more demand, or that the cost base has changed.

The saving may occur during a monthly reporting cycle rather than throughout the month. Verification and rework may consume part of it. Another step in the process may still constrain throughput. The person may also use the time to complete work that had previously been delayed or neglected. None of those outcomes makes the original time saving false. They determine what the saving is worth.

Why fractional productivity does not equal removable headcount

The calculation looks simple on paper. If 100 people each save 10 per cent of their working time, the aggregate saving appears equivalent to ten full-time employees.

Operationally, AI usually changes activities rather than removing complete roles. It might reduce the time required to prepare a report, analyse a document, produce an initial draft, reconcile information or collect status updates. The same people may still spend most of their week on judgement, negotiation, problem solving, stakeholder management and decisions requiring organisational authority.

The capacity can also be fragmented. A project manager may save several hours during monthly reporting without gaining the same capacity during planning, delivery intervention or stakeholder management. An operations employee may complete one process step faster while remaining constrained by work upstream or downstream.

A project or program team makes the problem particularly visible. Project managers, business analysts, engineers, architects, testers, change specialists and subject-matter experts will not all experience the same improvement. Even within one role, AI will not affect every activity equally.

Before treating an efficiency estimate as removable capacity, leaders need answers to some practical questions:

1.       Which activities have disappeared or materially reduced?

2.      When and where does the released time become available?

3.      Is it concentrated enough to reorganise work or roles?

4.      What remaining work, constraints or controls still determine capacity?

5.       Can the proposed operating model maintain the required service, throughput, quality and risk outcomes?

Without that analysis, an aggregated productivity percentage is not a workforce plan. I have seen both sides of this distinction in my own work.

Capacity redirected towards better project management

One trial used an AI project-support agent across three projects involving 28 project resources and 14 other stakeholders. A recurring problem was the effort required to collect status and RAID information, review what had been submitted, and follow up where responses were incomplete or too weak to support effective project management.

The agent helped collect updates and identify where stakeholders had not provided enough detail. It produced better RAID submissions and could potentially have saved around two hours each week for many stakeholders outside the core delivery team.

That estimate should not be treated as a headcount saving. The time was dispersed across people with responsibilities that remained necessary, and the trial was not designed to remove complete roles.

The more credible benefit was an improvement in how the portfolio operated. Less effort was spent assembling information, while project managers could direct more attention towards weak RAID items, targeted quality discussions and issues requiring intervention.

If the objective had been to support a larger portfolio without increasing project-management resources at the previous rate, the measure would have been different. The organisation would need to demonstrate that additional demand had been absorbed without a proportional increase in cost.

When AI creates more work

In another example, we introduced a delivery framework with a more structured front door for proposed initiatives. A bespoke agent helped people answer the required questions. Because it also contained the relevant framework and process information, we used the same agent to review submitted business cases.

The reviews were detailed, but often too detailed. The agent generated granular feedback on points that did not warrant the level of interrogation it applied.

That output still required somebody to reread the business case, assess the feedback, reclassify it using more realistic Must, Should, Could and Won't priorities, and then provide the relevant comments to the people who had prepared the proposal.

The agent had completed a review task quickly, but it had not necessarily reduced the end-to-end effort. It had moved work into verification, prioritisation and feedback. Unless the review criteria were recalibrated, reporting the agent's processing time as a productivity saving would have overstated the capacity released. This is why measurement needs to follow the workflow rather than stop at the AI-supported activity.

Workforce reduction is one possible outcome

Cost reduction is a legitimate objective. If an AI investment was approved to reduce the cost base and redesigned work can genuinely be delivered with fewer people, workforce reduction may form part of the benefit.

The workforce decision should follow the changed operating model, not the productivity percentage. EY-Parthenon's 2026 survey of 60 Australian CEOs illustrates the range of responses organisations are considering. Forty-four per cent ranked large-scale reskilling and upskilling as a leading workforce priority, while 41 per cent prioritised redesigning roles around human and AI capabilities. Twenty per cent identified reducing hiring in particular roles as a priority.[4]

The Richmond Fed and Duke University surveyed 748 financial executives and senior financial decision-makers across two samples. They found little aggregate impact from AI on employment during 2025 and similarly modest expectations for 2026. Large organisations, however, expected AI to reduce employment by around 0.8 per cent.[5]

PwC's 2026 Global AI Jobs Barometer provides another perspective. Companies with greater exposure to AI experienced stronger headcount growth than those with lower exposure, although the research does not establish that AI caused that growth.[6]

Current evidence does not support one universal workforce response. Growth, role redesign, avoided hiring, reallocation and cost reduction are occurring at the same time. For an individual investment, the relevant question is narrower: has the work changed enough to support the proposed workforce model?

Management needs to decide what the capacity is for

In a growing operational function, the best outcome may be absorbing additional demand without hiring at the previous rate. Another function might reduce overtime or contractor expenditure. A project organisation may support more delivery activity or give managers more time to address decisions and dependencies that had previously received insufficient attention.

Executive leadership and finance or benefits-management functions should define the enterprise outcome expected from a material AI investment. Portfolio leaders can identify where released capacity would address the most important demand or bottleneck. Line managers can test whether the theoretical capacity exists in practice, when it is available and whether the work can be reorganised.

This is where an aggregate model can become misleading. It may calculate the equivalent of twelve full-time employees across a function while the underlying time is dispersed across dozens of people, varies by role and appears only at particular points in the working cycle.

The opposite problem also occurs. Teams naturally absorb every efficiency by accepting additional work. That may be valuable, but it should not be presented as a cost reduction if the cost base and workforce demand remain unchanged.

Benefit realisation needs operational evidence

Does AI is exposing a weakness that predates the technology? Organisations are often disciplined when approving investment. Business cases contain financial models, governance forums test assumptions, executives approve expenditure and expected benefits are recorded.

After implementation, the original estimate can survive longer than the assumptions supporting it. Productivity benefits are particularly vulnerable because an efficiency percentage looks precise even when the path to financial value is unclear.

AI provides better opportunities to test that path. Tasks can be timed before and after implementation. Workflow volumes, cycle times and quality measures can be compared. AI-supported processes can generate detailed usage and activity information.

The evidence should match the benefit being claimed:

·         Avoided hiring: Demonstrate the change in demand, the previous workforce requirement and the recruitment genuinely avoided.

·         Higher throughput: Compare output or completed demand against a credible baseline without hiding deterioration in quality or control.

·         Faster delivery: Show an observable change in cycle time, backlog, milestones or comparable delivery measures.

·         Lower expenditure: Test the effect on overtime, contractors, external providers or another identifiable cost.

·         Improved quality: Define the expected change and measure outcomes such as defects, rework, incomplete submissions or control failures.

·         Workforce reduction: Show how activities and roles were redesigned so that a smaller workforce can maintain the required outcomes.

The measurement does not need false precision. It does need a defensible chain between what AI changed and the value the organisation says it received.

Conclusion

AI is making some forms of work measurably faster. The mistake is allowing an activity-level efficiency measure to travel through an investment case until it is treated as a financial or workforce outcome.

Leaders should follow the benefit through the organisation. Establish what became more efficient, determine whether usable capacity emerged, decide what will happen to that capacity, and measure the result against the outcome that justified the investment.

Hours saved can prove that a task became more efficient. Where the chain to cost, throughput, workforce demand, delivery or quality cannot be demonstrated, the business benefit remains unproven.

About Steven Gabris

Steven Gabris is an enterprise delivery and transformation leader with experience across financial services, payments, government and consulting. His work has included major transformation, delivery governance, operating-model improvement and the practical introduction of AI into project and program environments.

Through Gabris Group, Steven works with organisations to align strategy, operations and execution, with a focus on delivery confidence, governance and the practical application of AI in complex enterprise environments.

Sources

1.       Boston Consulting Group, "AI at Work: Why Strategy Matters More Than Tools", 3 June 2026.

2.      Boston Consulting Group, "Look Past Productivity to Get Real Value from AI", 31 August 2026.

3.      Gartner, "Gartner Survey Shows 45% of CFOs Say Their AI Investments Lean Toward Productivity, While 20% Say These Investments Lean Toward Decision Quality", 20 July 2026.

4.      EY Australia, "Australian CEOs prioritise reskilling over job cuts as AI adoption accelerates", 14 May 2026.

5.       Federal Reserve Bank of Richmond and Duke University, "How Might AI Change the Workplace? Evidence From Corporate Executives", 25 March 2026.

6.      PwC, "2026 Global AI Jobs Barometer", 15 June 2026.