What C-Suites Must Measure Before Treating AI Productivity as a Workforce-Reduction Mandate
By Roman Razuvayev and Mahesh M. Thakur
The AI conversation inside many executive teams has become dangerously compressed.
A company buys enterprise licenses for Copilot, Claude, Gemini, or another generative AI platform. Employees begin using it. Competitors announce layoffs. Headlines celebrate “leaner” organizations. Then comes the question:
“We invested in AI. Where is our 25% cost reduction?”
That question is understandable. It is also premature.
And it is no longer hypothetical. In February, Reuters reported that Block would cut more than 4,000 jobs — nearly half its workforce — as part of an AI overhaul. Reuters also reported that WiseTech Global planned to cut about 2,000 jobs, around 29% of its global workforce, over two years. Atlassian said it would lay off roughly 10%, or 1,600 employees, to push into AI and enterprise sales. By May, outplacement firm Challenger, Gray & Christmas reported that AI had led all reasons for U.S. job cuts for the third month in a row and had been cited in 87,714 cuts in the first five months of 2026. Round-number workforce targets – 10%, 20%, 25%, 40% – are now circulating in boardrooms as if they were benchmarks.
Generative AI is already creating meaningful value. It can accelerate research, documentation, customer support, software development, analysis, drafting, testing, and decision preparation. But a license purchase is not an operating-model transformation. Token consumption is not trust. Individual task acceleration is not enterprise productivity. And a workforce reduction is not proof that AI has created sustainable economic value.
The most dangerous mistake a C-suite can make in 2026 is to treat a visible AI investment as a mandate for an immediate, linear workforce cut.
The headlines, the competitor benchmarks, and the token dashboards all point in the same direction — and none of them answers the question that matters. More tokens consumed does not mean more work completed, more customer value created, or more trust earned. In some workflows, more tokens may mean the opposite: more uncertainty, more retries, more review burden, and less confidence in the first answer. AI will not, by itself, solve process, leadership, capability, data, or accountability problems that existed long before the model arrived.
The better question is not, “How many people can we remove?”
It is:
“Which work can now be done better, faster, safer, and at lower total cost – and what must change in our operating model before we redesign capacity?”
That is a harder question. It is also the one that separates responsible AI-led transformation from expensive organizational theater.
The pressure cascade is real
The pattern is becoming familiar.
The board reads that competitors are cutting staff while investing billions in AI. The CEO asks the executive team for a productivity plan. The CFO asks for a cost target. Business-unit leaders are told to deliver the same output with fewer people. Managers, afraid of being seen as resistant to AI, accept a number before they have validated the workflow.
The pressure then cascades downward:
· Leaders promise 25% productivity improvement.
· Teams are expected to sustain output with less capacity.
· Senior employees spend more time reviewing, correcting, governing, and integrating AI-generated work.
· Quality, reliability, customer trust, and employee trust become hidden costs.
The organization eventually discovers that it has not removed the bottleneck. It has merely moved it.
It would be a mistake, though, to dismiss this pressure as irrational imitation. Capital markets are actively rewarding it. When Cisco announced an AI-focused restructuring alongside strong results in May, Reuters reported that its shares surged 17% to a record high, set for the company’s biggest single-day gain in more than two decades, with market capitalization on pace to increase by about $70 billion. Boards see that. A CEO who declines to announce a workforce number is not resisting a fad; they are turning down a visible, immediate reward in favor of a slower, better-evidenced one. That takes conviction, and conviction requires evidence. Producing that evidence quickly is what the rest of this article is about.
This is not a warning against using AI aggressively. It is an argument for using AI intelligently.
One of us has led large, global engineering and business organizations where the gap between executive ambition and delivery reality is not theoretical. The other works with CEOs and senior leaders navigating high-stakes AI, talent, and organizational transitions. From both vantage points, the recurring failure is the same: leaders confuse a technology capability with a business outcome.
The cascade does not stop at your company’s walls
There is a second version of this pressure that most commentary misses, because most commentary is written from inside a single enterprise.
In global engineering services — the world one of us leads in daily — the 25% question does not arrive as an internal board memo. It arrives across the negotiating table: “You have deployed AI across your delivery teams. Why has your price not dropped 25%?”
AI-productivity expectations are now being written into vendor conversations: rate-card negotiations, contract renewals, managed-services agreements, fixed-price bids. The client CFO who has promised the board an AI dividend looks for the fastest place to collect it — and the services partner’s invoice is the most visible line item on the page.
The demand is rational. It is also where value gets destroyed fastest, on both sides, when it is linear and unmeasured. A provider forced to concede a blanket productivity discount before workflows are redesigned will protect its margin the only ways it can: thinner teams, more junior staffing, compressed review, deferred quality. The client collects the discount — and, two or three quarters later, collects the rework, the escalations, and the quiet attrition of the senior engineers who carried the account’s institutional knowledge. The bottleneck was not removed. It was moved across a contract boundary, where it is harder to see and slower to fix.
The same discipline that should govern internal workforce decisions applies between companies. Where AI demonstrably improves a workflow, the gains can and should be shared — transparently, and measured against the same evidence a CEO should demand internally. A client and a provider who jointly instrument a delivery pipeline and share the verified gain will both outperform the pair who settle the question with a percentage in a procurement meeting.
Four Q2 2026 signals that leaders should read carefully
The second quarter of 2026 offered several high-profile examples of companies reshaping their workforces while investing heavily in AI. They are important signals — but none of them is a formula for cutting 25% of a workforce.
1. Microsoft: Adoption is not the same as realized value
In April, Reuters — citing a CNBC report of an internal memo — reported that Microsoft was planning the first voluntary employee buyout in the company’s 51-year history. The same report noted that adoption of Microsoft 365 Copilot stood slightly above 3% of the company’s 450 million Microsoft 365 customers. That does not mean Copilot lacks value. It means even one of the world’s most capable AI companies must contend with the gap between product availability, customer adoption, workflow redesign, and realized commercial value.
The lesson is straightforward: license counts — even enablement counts — are not evidence that productivity has been captured. Before AI investment becomes a basis for workforce decisions, leaders need to know which teams use it, in which workflows, with what quality, and with what measurable business outcome.
2. Meta: AI-related workforce change is organizational redesign, not a simple productivity equation
Meta’s 2026 restructuring shows why timing and causality matter. A spring memo reported by The Wall Street Journal described a plan to lay off roughly 10% of employees – about 8,000 people – in May. A later Reuters report detailed the other half of the story: about 7,000 employees were to be transferred into new AI-related initiatives, managers eliminated, and organizational structures flattened. Reuters also reported that layoffs and transfers together would affect about 20% of the company’s workforce, and that an additional 6,000 open roles had been closed as part of the process.
That is not a clean “AI replaced 20% of employees” story.
It is a company redesigning its structure, reallocating talent, flattening management, building AI-native workflows, and shifting capacity toward a different strategic future. Those are very different decisions from declaring that AI has made one in five jobs unnecessary.
Workforce reduction, talent redeployment, management redesign, and AI workflow transformation are related — but they are not interchangeable. A company can eliminate roles while still investing aggressively in people, skills, governance, data quality, and change leadership.
3. Cisco: The strongest AI moves are connected to a growth thesis
Cisco announced in May that it would reduce fewer than 4,000 roles – less than 5% of its workforce – as part of a restructuring to shift investment toward AI, silicon, optics, security, and related growth areas. At the same time, Cisco had taken $5.3 billion in AI infrastructure orders from hyperscalers and raised its full-year order expectation to $9 billion from $5 billion. Its restructuring was expected to cost as much as $1 billion.
Cisco’s CFO was explicit that this was “really not a savings-driven restructure” but a reallocation of resources toward where demand is moving.
Cisco’s example is useful because it is not only a cost story. It is a strategic allocation story. The company is placing a bet that AI infrastructure demand will create future value in networking, optical systems, silicon, security, and adjacent growth markets. The workforce action is connected to a defined growth thesis — not to a broad claim that “AI means fewer people.”
That is the standard: if a workforce decision cannot be explained in terms of a stronger growth engine, customer advantage, faster time-to-market, or more strategic capacity, it may be managing optics rather than transformation.
4. Oracle: A complex restructuring rarely fits a single AI headline
In June, Reuters reported that Oracle’s workforce had declined by approximately 21,000 people, or 13%, in fiscal 2026. Oracle had 141,000 employees as of May 31, 2026, compared with about 162,000 a year earlier. Reuters also reported that the company spent $1.84 billion in severance and other exit costs in fiscal 2026, up from $374 million the prior year, and that Oracle’s filing pointed to multiple drivers: management and product changes, performance issues, strategic shifts, acquisitions, and AI adoption and deployment.
And the same fiscal year produced record results: total revenue of $67.4 billion, up 17%, with cloud revenue up 39% to $34.0 billion. Oracle cut from strength, not from weakness — which is precisely why the simple story does not hold.
The lesson is not that Oracle’s choices were right or wrong. The lesson is that headlines simplify. “AI caused a 13% workforce reduction” is easier to repeat than the truth: enterprise restructuring is multi-causal. Strategy, capital allocation, acquisitions, product decisions, customer demand, leadership changes, and AI adoption often happen at the same time.
The wrong conclusion to draw from Microsoft, Meta, Cisco, and Oracle is also the simplest one: “They cut, therefore we should cut.” Each of these companies made a structural decision inside its own strategy. Copying the percentage without the strategy imports the risk and leaves the logic behind.
A token is not a unit of value
One of the worst productivity metrics now appearing in executive dashboards is token usage.
Token consumption may tell a company that people are engaging with an AI model. It does not tell the company whether the model improved customer outcomes, reduced cycle time, lowered rework, increased quality, reduced risk, or created economic value.
In fact, more tokens can sometimes signal the opposite:
· Employees are prompting repeatedly because they do not trust the first answer.
· Teams are using AI to generate work that must be extensively reviewed.
· Users are exploring without a clear workflow or decision purpose.
· Managers are measuring activity because they cannot yet measure outcomes.
· People are learning to “look AI-active” rather than becoming more effective.
More usage can be positive. But usage without outcome measurement can become performative.
Trust is even more important.
Google Cloud’s 2025 DORA research found that 90% of surveyed technology professionals reported using AI at work and more than 80% believed it increased their productivity. Yet 30% reported little or no trust in AI-generated code. The same research found that AI adoption was positively associated with delivery throughput and product performance, while still negatively associated with delivery stability.
That is the actual leadership challenge.
AI can make teams faster. It can also expose weak architecture, poor testing, unclear ownership, fragmented data, weak product requirements, poor feedback loops, and brittle operating processes.
AI does not fix a weak system. It amplifies it.
Faster tasks are not a faster enterprise
The distinction between local improvement and system-wide value matters most in complex work.
A developer may complete a draft faster. A customer-support agent may summarize a case faster. A product manager may create a first-pass requirements document faster. A finance employee may produce a scenario analysis faster.
Those are valuable gains. But a company cannot assume that those gains translate directly into fewer people required across the entire workflow. The constraint may sit elsewhere.
A faster coding step may create more review work. Faster customer-response drafting may create more escalations. Faster analysis may require more verification. Faster content production may create a heavier legal, brand, or compliance burden. Faster software development may overwhelm testing, security, integration, release management, or customer onboarding.
In safety-regulated engineering domains, this gap between generated output and verified value is not a metaphor – it is a certification requirement. In automotive software, work does not create value when code is written; it creates value when it passes the functional-safety obligations of standards such as ISO 26262, tool-qualification requirements, and the validation gates that stand between a code change and a vehicle on the road. AI can compress the first step dramatically. If verification capacity and methodology do not change with it, the organization has produced faster inventory, not faster value — and no OEM program director will mistake one for the other.
A controlled study by METR illustrates why leaders should be cautious about relying on perception alone. In a randomized trial involving 16 experienced open-source developers completing 246 real tasks in familiar codebases, developers using AI tools took 19% longer to complete work, even though they had predicted a 24% speedup and still believed afterward that AI had made them 20% faster. The study does not prove AI slows down all developers or all work. It does prove that self-reported productivity is not sufficient evidence for a workforce decision.
The critical question is not: “Did an individual task become faster?”
It is: “Did the end-to-end system create more customer value, at lower total cost and acceptable risk?”
The Five Proofs: what every C-suite needs before cutting capacity
Before treating GenAI productivity as a workforce-reduction mandate, leadership teams should establish five categories of evidence. We call them the Five Proofs, because each replaces an assumption with evidence.
1. Flow: Did work move faster from demand to customer value?
Measure the full journey, not one task. Depending on the business, this could include:
· Lead time from customer request to resolution
· Time from idea to release
· Quote-to-cash cycle time
· Time from claim intake to decision
· Product-development cycle time
· Time to onboard a new customer
· Time to close a material operational exception
If AI accelerates one activity but the overall cycle time does not improve, then the value has not yet been captured.
2. Quality: Did the output improve, hold steady, or deteriorate?
Measure rework rates, defect escape rates, customer escalations, compliance exceptions, security incidents, rollbacks, error correction, and AI-output acceptance without material human rewrite.
A faster first draft that creates more downstream correction is not productivity. It is deferred cost.
3. Economics: Did the company improve the cost of a verified outcome?
Measure cost per resolved customer case, per compliant decision, per product release, per qualified lead, per validated design, per shipped feature, per dollar of incremental revenue.
Cost reduction measured only through headcount is incomplete. Measure the cost of a verified, reliable, customer-valued outcome.
4. Capacity: What will the company do with the time AI creates?
This is where many leadership teams make the wrong choice. If AI frees 15% of capacity, the company has options: reduce cost, increase innovation, improve service levels, shorten time-to-market, build new products, reduce employee overload, strengthen controls and quality, or redeploy talent toward growth.
A company that automatically converts every capacity gain into a cut may weaken its ability to grow precisely when the market is changing fastest.
5. Trust: Are employees, customers, and managers willing to rely on the system appropriately?
Trust should be measured through behavior, not slogans. Track override rates, escalation rates, human review time, repeat prompting, rework after AI assistance, user confidence by workflow, customer satisfaction, employee sentiment, and attrition among critical senior talent.
The question is not whether employees “like AI.” It is whether they trust it enough to use it appropriately — and whether the organization has built the controls required to earn that trust.
A better operating rule: redesign work before reducing workforce
The sequence matters. And every CFO who hears “measure first” will ask the fair question: how long? The honest answer is quarters, not years. Run per workflow, the sequence below fits inside a single fiscal year — and it is not a twelve-month freeze. Workflows enter the pipeline continuously, and capacity decisions are made workflow by workflow as evidence matures, not in one monolithic gate at the end.
First, establish the baseline (quarter one)
Measure current throughput, quality, cost, delay, and customer experience before deploying AI widely. Without a baseline, every later productivity claim becomes storytelling.
Second, pilot in defined workflows (quarters one to two)
Choose a workflow with a clear owner, measurable output, known pain point, and manageable risk. The starting point is not “use AI everywhere.” It is: “improve this specific workflow by this specific measure without compromising quality or trust.”
Third, redesign the workflow (quarters two to three)
AI cannot simply be added to an old process. Clarify roles. Remove redundant approvals. Improve data access. Strengthen knowledge management. Define escalation paths. Decide where human judgment remains mandatory. Update controls. The work itself must change.
Fourth, prove sustained value (quarters three to four)
One good month is not a new operating model. Require evidence over multiple cycles. Validate the gains across different teams, managers, customer segments, and workload conditions. Confirm that gains did not come from hidden overtime, deferred maintenance, quality erosion, or the overuse of senior employees.
Fifth, decide what to do with capacity (workflow by workflow, as evidence matures)
Only then should leaders decide whether the capacity created should be converted into growth, quality, innovation, customer experience, redeployment, or workforce reduction.
The workforce decision should be the result of a measured transformation – not the starting assumption.
A leadership team that starts this quarter can stand in front of its board within two quarters with pilot evidence, and within four with a defensible capacity plan. That is not slower than a headline cut. It is faster than announcing a number, missing it, and rebuilding trust afterward — with employees, with customers, and with the board itself.
The leadership test
There is a simple test for every CEO, board member, CFO, CHRO, and business-unit leader:
If AI use disappeared tomorrow, could you clearly identify which customer outcomes, workflows, economics, and decisions would become worse — and by how much?
If the answer is no, the organization has not yet proven value. It may have enthusiasm. It may have tokens. It may have licenses. It may have a competitor benchmark. It may even have a workforce target. But it does not yet have evidence.
The companies that win the AI era will not be those that cut fastest. They will be the companies that learn fastest: the ones that measure the full system, protect trust, redesign work, redeploy talent intelligently, and convert AI capability into durable customer and economic advantage.
Do not cut 25% of your workforce yet.
First prove that your operating model can create 25% more value. Then the workforce decision – whatever it turns out to be – will rest on your own evidence rather than on someone else’s headline.
Roman Razuvayev is a global technology, engineering, and operations executive with more than 20 years of experience leading complex transformation across automotive, industrial, MedTech, and digital-platform environments. As SVP Engineering at GlobalLogic, a Hitachi Group company, he has led global portfolios of up to $260M and organizations of more than 3,000 professionals across Europe, India, Japan, and the United States. Roman brings deep operating experience in global engineering services, enterprise delivery, cross-border scale, safety-regulated software environments, and AI-enabled productivity transformation.
Mahesh M. Thakur is a CEO and C-suite coach, AI leadership strategist, and former Silicon Valley technology executive who helps leaders navigate high-stakes transitions involving AI, culture, strategy, and enterprise performance. He has held senior product and business leadership roles at Microsoft, Amazon, Intuit, and GoDaddy, and now advises executives on how to lead with clarity, judgment, and measurable impact in moments of disruption. Mahesh is a Master Certified Coach, a Marshall Goldsmith MG100 coach, and a Stanford GSB Certified Board Member.
