The AI Worked. Does the Evidence Support the Proposed Restructuring?

AI can deliver real productivity and economic value. But when a productivity result is used to justify a permanent workforce reduction, leaders face a separate question: does the evidence support this particular restructuring decision?

Suppose an AI rollout produces a credible 20% productivity improvement.

The business case then proposes removing 200 roles.

The 200 may be exactly right. It may reflect detailed analysis of workload, demand, service requirements and the future operating model.

But the productivity result and the restructuring decision are not the same proposition.

The productivity result establishes that some work can be performed more efficiently. The restructuring decision has to establish that the business will no longer need the capacity provided by those 200 roles, and that it can remove that capacity on the proposed timetable.

The approval question is: what evidence connects this productivity result to removing this much capacity permanently, on this timetable?

That means showing how the work and demand will change, and how much capacity the business will still need. It also means showing why the proposed amount and timing make economic sense.

A practical way to ask that question is:

Have we seen the evidence for how much capacity the business will actually need, not just the productivity result?

The answer to that question can support a substantial reduction just as readily as it can challenge one.

What a real productivity result does, and does not, tell you

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the introduction of a generative-AI assistant to 5,172 customer-support agents. Access to the assistant increased productivity by about 15% on average, measured as issues resolved per hour. 1

The effect was not uniform across the workforce, which matters if an average gain is being used to estimate capacity. Gains were much larger among less experienced and lower-skilled workers, while the most skilled workers saw little significant productivity improvement. 1

If demand for customer support remained unchanged, the researchers calculated that the same volume of work could be handled with roughly 12% fewer worker-hours. 1

Fewer worker-hours do not, by themselves, mean the same percentage of roles can be removed. The released time still has to translate into positions the business can actually remove, on the proposed timetable.

The calculation therefore moves through several different quantities before any workforce conclusion can be reached, and even its worker-hours estimate depends on demand remaining unchanged.

The authors also note that longer-run labour demand could move in different directions as service improves, workers take on different activities, or new work emerges around the technology. 1

The useful question is therefore not simply whether the productivity result is real. It is what the evidence says about workload and labour demand after that gain.

The workforce conclusion still needs its own evidence

Commonwealth Bank of Australia illustrates how that distinction can matter in practice.

In July 2025, CBA confirmed that 45 roles in its Customer Service Direct business were being cut alongside the introduction of an AI voicebot. 2 The Finance Sector Union challenged the decision through Australia's Fair Work Commission. 3

CBA later reversed the redundancy decision. The bank said its initial assessment had not adequately taken all relevant business factors into account and that the roles were therefore not redundant. It also said it should have been more thorough in assessing the roles required. 3

That does not establish that the voicebot failed or that AI-linked restructurings are generally unreliable. It establishes something narrower: the workforce conclusion was a separate management judgement, and CBA later said its initial assessment had not adequately supported that workforce conclusion.

AI can still support real reductions

That caution does not mean AI cannot remove substantial amounts of human work. IBM reports a 94% containment rate for common employee HR inquiries through AskHR, in a support model where AI handles routine inquiries and human advisers handle more complex needs. 4 CEO Arvind Krishna has also said AI agents replaced work previously performed by several hundred HR employees as IBM increased hiring in other areas. 5

These are company-reported outcomes, not independent proof that a productivity gain translates into a particular headcount reduction. Nor do they show how IBM derived or validated the amount of capacity it removed. They support a narrower point: AI can coincide with substantial real workforce reduction. Requiring evidence for a restructuring decision is not an argument against reducing capacity.

The evidence needs to show how the productivity result changes the work, how much demand remains, and how much capacity the business still needs. It also needs to support the proposed amount and timing. Without that connection, the evidence does not yet support the permanent reduction.

The productivity result is evidence for the restructuring case. It is not a substitute for the restructuring case.

Related resource

AI Restructuring Business Case Stress Test
A structured method for examining the evidence behind a live AI-linked restructuring case.

References

  1. Erik Brynjolfsson, Danielle Li and Lindsey Raymond, “Generative AI at Work”, The Quarterly Journal of Economics, 140(2), May 2025, pp. 889–942. ↩ ↩ ↩ ↩
  2. ABC News, “Commonwealth Bank replaces dozens of call centre jobs with AI chatbot”, 29 July 2025. ↩
  3. ABC News, “Commonwealth Bank backtracks on AI job cuts, apologises for 'error' as call volumes rise”, 21 August 2025. ↩ ↩
  4. IBM, “IBM AskHR”. ↩
  5. The Wall Street Journal, “IBM CEO Says AI Has Replaced Hundreds of Workers but Created New Programming, Sales Jobs”, May 2025. ↩