In Brief
In our assessment, an AI project aimed at cutting staff still struggles to deliver even a small positive economic effect. Pointing the same AI at the constraint, the one step that limits how much your company can deliver and sell, can be worth considerably more. Three recent studies inform this view, though none of them tests it directly.
- CFOs expect AI to reduce total employment by only about 0.4% in 2026, and half of firms expect no role to be replaced.
- A saved hour becomes profit only if the capacity is removed or moved to the constraint. Otherwise nothing reaches the P&L.
- In the worked example below, a 2% gain in output at the constraint is worth about three and a half times a 2% payroll cut in net EBITDA, provided the market buys the extra output.
What the evidence says
AI is paying off through growth and individual time savings, not through headcount cuts. The findings come from three studies: Baslandze et al. (CFO survey, about 750 firms, 2026), Brynjolfsson, Li and Raymond (customer support, 5,179 agents) and Dillon et al. (Copilot field experiment, 7,137 workers, 66 firms).
- Replacement is small and concentrated. CFOs expect AI to cut aggregate employment by about 0.4% in 2026. Large AI-investing firms expect about −0.7%, and the routine clerical share is expected to fall about 2 points by 2028.
- Cost-cutting does not show up as the payoff. Labor-cost reduction ranks among the weakest motives for AI investment (2.0 out of 4) and shows no robust link to productivity gains. The gains correlate with product development and reaching customers.
- Time savings are real but modest, and they stay with the worker. Copilot users saved about 2 hours a week on email, roughly 5% of a working week. Output and tasks did not change, and part of the time went back to shorter workdays.
- AI helps weaker workers most. In customer support, productivity rose 14% on average, 34% for the lowest-skill agents and not at all for the top performers.
- The spend is mostly recurring. 55–64% of AI spending is subscriptions and services, not capital.
Where staff reduction can work, and what it is worth
Staff substitution is plausible for routine clerical and standardized support work at large firms, and unproven elsewhere.
| Likelier to pay | Weak or unproven |
|---|---|
| Routine clerical and administrative work (data entry, bookkeeping, transaction processing) | Management, engineering, legal, sales (described as “enhanced”, not replaced) |
| Tier-1 customer support and high-volume standardized queues | Judgment-heavy work: treasury, tax, M&A |
| Large firms with enough volume to cover subscriptions and integration | Small teams, where savings are too thin to cover recurring costs |
None of the three papers measures ROI on staff substitution. Our conclusion: with today’s tools, a staff-reduction project still struggles to deliver even a small positive economic effect, and it carries high organizational risk.
Why headcount projects disappoint, and what they do to the next AI project
Four problems recur across the studies.
- Freed time is not saved cash. Without removing or redeploying capacity, nothing reaches the P&L. Exception handling, error checking and governance also absorb part of the gain.
- Savings sit at non-constraints. Faster email does not raise what the company can sell.
- The people who make AI work are the ones threatened. The tool learns from top performers’ tacit knowledge, yet they gain little from it. Under a cut-staff narrative, they have every reason to withhold it.
- Adoption depends on management and training, not on the tool. Firm identity explained far more of the variation in usage than worker traits. The top barriers reported by non-adopters were immature technology (42%), an untrained workforce (36%) and privacy concerns (36%).
A visible “AI means layoffs” program also damages every project that follows it:
- Employees under-report usage or turn to unsanctioned tools.
- Experienced staff stop documenting how they work, which starves the AI of its best training material.
- Strong performers are the ones who can leave.
- Later projects that need cooperation, such as process redesign or data sharing, start with low trust. Skeptics can be won over: adherence to AI suggestions rose over time in the support study once people saw value.
A better target: the business constraint
The constraint is the one step that limits how much the whole company can deliver or sell. One hour gained there is one hour gained for the whole system, while an hour saved elsewhere is usually idle time. The studies point the same way: gains appear on the revenue side, and individual time savings do not become value without process redesign.
As a former CFO in manufacturing, banking and telecom, my first question about any saving is where it shows up in the P&L. Savings at a non-constraint rarely do.
The comparison below uses assumed numbers for a company with €100m revenue, a 40% contribution margin (revenue minus truly variable costs) and a €15m payroll.
| Cut 2% of payroll | Raise constraint output by 2% | |
|---|---|---|
| Gross effect | €0.3m payroll saved | €2m extra revenue × 40% = €0.8m contribution |
| Assumed AI cost per year | €0.1m | €0.1m |
| Net EBITDA effect | about €0.2m | about €0.7m |
| Key condition | Capacity is actually removed | The market buys the extra output |
The second column holds only if demand exists. If the constraint sits in sales or pricing, extra production adds nothing, and the AI project should target conversion, lead response or quote turnaround instead.
Proposed AI projects by constraint type
Each project works on the constraint itself or protects it. The last column is what a pilot should measure.
| Constraint | AI project | Mechanism | Metric |
|---|---|---|---|
| Production machine or line | Predictive maintenance and AI scheduling to cut changeovers and unplanned stops at the constraint | Exploit | Constraint uptime, contribution margin per constraint hour |
| Quality control at the constraint | AI inspection upstream of the constraint | Protect from defective input | Scrap and rework at the constraint |
| Engineering, design or proposal capacity | AI drafting, specification reuse, automated configuration of quotes | Exploit | Quotes or designs per week, win rate |
| Sales capacity (senior sellers) | AI prep, lead scoring, CRM admin removal | Free scarce selling hours | Selling hours per rep, pipeline conversion |
| Expert or support capacity | An AI assistant that spreads top performers’ practices to novices | Elevate | Time to full productivity, resolutions per hour |
| Credit, underwriting or compliance review | AI pre-screening so experts only see the exceptions | Subordinate | Decision cycle time, volume cleared |
| Material and demand variability | AI forecasting for buffer sizing | Subordinate | Buffer penetration, on-time delivery |
A 6-week pilot, and how to measure success
A pilot should answer one question: does output at the constraint rise, and does the market take it?
- Week 1: find the constraint. Map the flow from order to cash and find where work queues up and capacity is never idle.
- Week 2: set the baseline. Record weekly output, uptime, rework and contribution margin per constraint hour.
- Weeks 3–5: run the pilot. Apply one AI project from the table to one constraint, with the people who run it involved from day one.
- Week 6: decide. Compare against the baseline and scale only if output at the constraint rose and the extra output sold.
Success means more output at the constraint, converted into sales. Hours saved elsewhere, licences deployed and adoption rates are inputs, not results. Move any capacity freed at non-constraints to the constraint or to growth work, which turns AI into a gain for staff instead of a threat. Re-check the constraint after each gain, since it will move.
The private equity view: EBITDA, cash and hold period
For a PE-backed company the test is what moves EBITDA and cash within the hold period. A constraint project fits that horizon better than a restructuring.
- EBITDA. In the example above, about €0.7m of net EBITDA. At an illustrative 8x exit multiple that is roughly €5.6m of enterprise value, against about €1.6m for the payroll cut (€0.2m × 8).
- Cash. Higher output ties up working capital in inventory and receivables for the first quarters, so plan for it. A payroll cut may add severance cash cost up front.
- Speed and risk. A pilot of six weeks and a sub-€0.7m budget can be reversed. A headcount program is slow to execute and costly to undo.
- Execution risk. Cooperation from the team matters. A constraint project asks them to help, and a layoff project asks them to hand over what they know.
Conclusion
Our conclusion: today it is still hard for a staff-reduction project to deliver even a small positive economic effect. Targeting headcount yields a small, uncertain saving and creates resistance that raises the cost of every later AI effort. Targeting the constraint raises what the whole company can deliver and sell, and it earns staff cooperation.
If you know which constraint limits your revenue, start there. If you do not, that is the first thing to find, and FlowKnot can help you find it.
Sources and limits
Baslandze, Edwards, Graham, McClure, Meyer, Sparks, Waddell, Weitz (2026). Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives. NBER Working Paper 34984.
Brynjolfsson, Li, Raymond (2023). Generative AI at Work. NBER Working Paper 31161.
Dillon, Jaffe, Immorlica, Stanton (2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795.
These are working papers that have not completed peer review. Most of their employment figures are CFO forecasts, not audited results, and the horizon is short.
The payroll examples, cost assumptions, exit multiple and the constraint comparison are the author’s own illustrations, not figures from these papers.