Playbook › Playbook
Headcount-vs-revenue divergence is the only hard test of AI deflation in services
Claim
The whole services cohort de-rated through H1 2026 on one thesis (Nifty IT −31%, its worst first half since 2003). Every management team responds with an identical narrative: an AI platform brand, an agentic-operations pivot, a shift toward outcome-based pricing, and reskilling rather than restructuring. The narratives are indistinguishable, so they carry no information. Bookings and TCV do not resolve it either — bookings can grow while price per unit of work falls, and both firms below reported record or near-record deal activity.
The test that does resolve it is a two-line comparison from the same quarter:
| Genpact Q2 2026 | Infosys Q1 FY27 | |
|---|---|---|
| Revenue growth | +7.1% | +2.4% cc (organic ~+0.5%) |
| Average headcount | −2.6% | flat (−52 QoQ on 328,062) |
| Guidance action | Raised the ATS target | Cut the FY27 range |
| Gross margin | +60bps | flat for four straight years |
| Forward P/E | 7.9× | 14.8× |
Same industry, same quarter, same AI, opposite readings — and the market has the cheaper multiple on the firm with the better divergence.
Why it works
Productivity gains in a seat-based business go to one of two places: the provider keeps them (revenue holds, headcount falls, margin expands) or the client takes them (revenue falls, headcount holds because the work is still being done by people, margin is defended by cutting elsewhere). There is no third destination. The sign of the divergence therefore reports the outcome of the pricing negotiation directly, before it shows up in margin — which lags by several quarters because both firms are defending margin with cost actions in the interim.
This also explains why flat gross margin is not reassuring in this cohort. See
pattern-margin-intact-while-volume-falls-defers-the-damage: an incumbent under price attack
chooses between ceding margin and ceding volume, and Infosys's four years of flat 30.2% gross
margin alongside a growth stall is the volume-ceding branch.
Refinement — "the provider kept it" splits into two states, and only one is investable
IBM Q2 2026 forced this. The original claim treats "provider keeps the productivity" as the good branch. IBM sits squarely in that branch and is still not a growth business:
| State | Gross margin | Revenue | Reading |
|---|---|---|---|
| Infosys | flat 4 years | stalled (+0.5% organic) | Client takes it and no growth — worst |
| IBM Consulting | +140bps | +0.2% | Provider keeps it, converts none of it to volume — harvesting |
| Genpact | +60bps | +7.1% | Provider keeps it and grows — best |
Harvesting is the good branch of a bad trade. No price give-back, no margin erosion, a durable profit annuity — and no growth engine, ever. IBM Consulting posted signings +6% and book-to-bill 1.05 converting to +0.2% revenue: the demand is being won and is not arriving as revenue, which is services deflation showing up one layer further down than the margin line looks for it. Value a harvester as an annuity at its current margin; do not underwrite growth from its bookings.
The forward test that separates harvesting from a genuine turn: book-to-bill >1.0 with flat revenue for two consecutive quarters confirms deflation and the segment should re-rate down. One quarter is noise.
⚠️ The test cannot be run on a segment of a multi-segment company
Firm-wide headcount is an invalid proxy for one segment, and it fails in the flattering direction. IBM-wide FY2025 revenue +7.6% against headcount −2.2% gives a +9.8pt divergence — almost exactly Genpact's +9.7pt. That number is meaningless:
- it includes Software and Infrastructure staff, who are not the seat-based business under test;
- FY2025 revenue was inflated by the z17 mainframe surge, which consumes no consultants;
- HashiCorp and Confluent added revenue and headcount on entirely different ratios.
IBM does not disclose segment headcount at all, so the test is simply unavailable — every figure in circulation (~290k, Revelio Labs) is a third-party estimate. This is the same failure class as [[pitfall-half-corrected-peer-table-invents-a-discount]]: a ratio assembled from a numerator and a denominator that do not describe the same business.
Rule: run this test only on a pure-play, or where the firm discloses headcount for the segment being tested. Otherwise fall back to the segment gross-margin line, which is disclosed and which pointed the right way on IBM (+140bps) when the headcount proxy would have flattered it.
How to apply
- Compute revenue growth minus average-headcount growth for the latest two quarters. Positive and widening is the only quantitative evidence that AI is being monetised. Genpact publishes it as a "productivity index" (1.04 in Q2 2026); most firms do not, so derive it from the revenue line and the disclosed headcount.
- Two quarters is not a trend. Weight it as evidence, not proof — Genpact's divergence is two quarters old against a prior three-year trend of declining revenue per employee ($36.3k in 2023 → $34.8k in 2025). Require four quarters before treating it as structural.
- Read the 10-K risk factors for the give-back mechanism. Genpact's states that "some of our MSAs and SOWs require certain minimum productivity benefits to be passed on to our clients" — the bear case is contractual, not merely competitive. A firm that discloses this is more trustworthy than one that declines to quantify, which is itself a signal: Infosys's CFO used the phrase "AI deflation" on the record and the CEO then refused to size it.
- Do not accept revenue-per-employee as a substitute unless the firm reports it. Derived figures (revenue ÷ period-end headcount) are noisy on acquisitions and on onsite/offshore mix shifts.
Corollary — the cluster warning
Names that look independently cheap for this reason are not independent positions. In the
2026-08-11 batch, G, INFY, DBX and AKAM were all cheap on the same AI-displacement
repricing and all resolve on the same question. Size the theme, not the tickers. See
Knowledge/Market/regime.md on Act I of the 2026 rout.
History
- 2026-08-12 — extended by the IBM full analysis. Added the harvesting third state (provider retains productivity but converts none to volume: margin +140bps on +0.2% revenue, signings +6%, book-to-bill 1.05) and the segment-proxy invalidity rule — IBM-wide headcount produced a false Genpact-shaped +9.8pt divergence for a segment whose headcount is not disclosed. EPAM added as a supporting loser case: guidance cut, blaming a "faster-than-expected shift away from task-based services."
- 2026-08-11 — written. Observed across Genpact Q2 2026 and Infosys Q1 FY27, reported within two weeks of each other, with Concentrix (revenue +1.9%, guidance cut) and Accenture (revenue +6%, bookings −2%) as supporting cases. Industry backdrop: BPO annual contract value fell 14% in 2025, the lowest since 2020.