Function-Selective AI-Era Headcount Cuts

Sources#

• 2026 State of Scaling: The Great Sorting
• Cloudflare Cut 20%. Atlassian Cut 10%, Mostly Engineers. Bending Spoons Cuts Up to 75%. All Three Are Growing Faster or Earning More. So How Lean Can B2B Teams Really Get?

Summary#

The vault's headcount evidence splits by population. Private AI-forward growers add people (Firm AI-Spend Intensity and Headcount Growth: ~10% at intensive AI spenders, 146% median at ICONIQ's 100%+ growers), while mature public incumbents announce AI-attributed cuts of roughly 10–20% (ICONIQ's 2026 timeline, carried on the same page). That timeline gives a percentage per company and nothing else. It does not say which functions were cut or what happened afterwards.

Jason Lemkin's How Lean Can B2B Teams Really Get? (SaaStr, 2026-10-17, practitioner-opinion) supplies both for two names on that list, Cloudflare and Atlassian, plus a third company that cuts as its business model, Bending Spoons. His argument is that which function you cut, and at what growth stage, determines how lean you can go. The three cases cut different functions for different reasons, and each was followed by a better quarter.

Evidence note. practitioner-opinion, with three kinds of number that need keeping apart. (1) Filing-derived: revenue, growth, guidance, margins, headcount at year end and at IPO, cut sizes and charges, sourced in the post to SEC filings, the Bending Spoons S-1 and quarterly results. These are secondhand here and were not re-checked against the filings. (2) The author's own arithmetic: both Cloudflare revenue-per-employee figures, including the ~4,050 post-cut headcount, which is his estimate. (3) SaaStr self-report: revenue, team size and agent counts for his own company, unaudited and internally inconsistent (below). The sample is three companies chosen for their good outcomes. ICONIQ's timeline lists fourteen AI-citing cutters, and the post discusses only the two with strong results. Stock moves are one-day reactions. The post also sells the agent-run model its author runs.

Three cuts, three functions#

Cloudflare cut the function that scales with revenue but does not produce it. The stated aim was that G&A stop growing in step with revenue. Management credited the quarter's growth to Workers and agentic workloads, so on the company's own account the acceleration came from demand, and the cut lowered the cost of serving it. Two details matter beyond the headline. Agent usage came before the cut: internal AI tool usage rose more than 600% in three months, with thousands of agent sessions a day across engineering, HR, finance and marketing. This is the order Organizational Complements to AI predicts works, with the workflow rebuilt first and headcount following. And Prince expects more employees in 2027 than at any point in 2026, hired for different roles. That makes the cut a change of composition rather than a lasting reduction, closer to the backfill-suppression and role-mix channels recorded on Firm AI-Spend Intensity and Headcount Growth and Role Averaging, Not Role Elimination than to substitution.

Atlassian cut engineering at a company whose growth was in trouble, and moved the money into AI and enterprise sales. Cloud growth accelerated and the stock rose 35% the next day. Lemkin flags the limits himself: FY2027 guidance steps total growth down to ~13% and ARR growth to 18%, partly because Data Center revenue is expected to fall ~17% as customers migrate to cloud, and about one point of FY26 growth came from acquisitions. Some of the cloud acceleration is therefore migration out of the company's own on-premises product rather than new demand. The company also split the CTO role into two AI-focused positions, so engineering's working method changed along with its size. The MCP-server and Teamwork Graph CLI usage figure (past one million monthly active users) is the company's own claim.

Bending Spoons runs acquired products for margin, not growth. NRR of 95% means the acquired base shrinks each year, and the company grows by buying more of it. Lemkin draws the boundary: the case shows that a mature, slow-growth product can run on a small fraction of its team at much higher profit, and it does not show that a product can grow 30% at that headcount. Vimeo is the cost. An ambitious pre-close roadmap was followed by layoffs. More than 90% of the quarter's pull requests were written or co-written by AI (company-reported).

The rule underneath: cut depth follows growth stage#

Lemkin turns the cases into cut ranges by function. These are his judgment, not measured:

Read against the rest of the vault, this table restates the investment-then-efficiency crossing as operator advice. AI Investment Story, Not Efficiency Story reads that crossing off ICONIQ's 2026 data: growth-selected AI cohorts spend 284% of revenue on OpEx below $100M and 53% above it, cheaper than peers once mature. Lemkin's version keeps headcount where revenue is being bought (engineering at a company growing 60%, quota sales on a real pipeline) and cuts where revenue is being harvested (G&A at any stage, engineering at a mature org, everything at a no-growth acquisition). The two sources reach the same rule independently, one from operating data and one from three cases, and neither is a test of it.

No source in the vault measures cut depth by function. The nearest is McKinsey's function-level series on Firm AI-Spend Intensity and Headcount Growth, which reports the share of respondents seeing any AI-attributed decrease, not how deep a cut went. Lemkin's ranges have no measured counterpart.

The revenue-per-employee ladder, and what is inside each rung#

Lemkin's chart ("the old benchmark is dead") stacks six figures: ICONIQ typical B2B $237K (ARR per FTE, State of Software 2025), Cloudflare before the cut ~$496K, ICONIQ Pacesetters $100M+ median $655K, Cloudflare after the cut ~$690K, Bending Spoons $2.57M, SaaStr ~$6.7M. Each rung needs a caveat:

• Cloudflare's jump is arithmetic, not productivity. Q1 $639.8M × 4 / 5,156 ≈ $496K, and Q2 $696.1M × 4 / ~4,050 ≈ $687K. Revenue rose ~9% quarter on quarter, so nearly all of the ~40% rise comes from the denominator, as Lemkin says. The ~4,050 is year-end-2025 headcount minus the cut, which ignores Q1 hiring and the timing of departures; departing staff were paid base salary through the end of 2026. Treat it as an estimate of an estimate.
• The Pacesetter $655K is a selected cohort's median. AI Investment Story, Not Efficiency Story carries the selection caveat in full: top-quartile growers under ICONIQ's own AI label, no non-AI control. On ICONIQ's own State of Scaling series, the top quartile of all $100M+ companies was $478K in 2026, so pre-cut Cloudflare already sat there.
• Bending Spoons' $2.57M does not reconcile with its own IPO headcount. Q1 2026 revenue of $601.3M annualizes to ~$2.4B, and over the 2,284 employees at IPO that is ~$1.05M per head, well under half the 2025 figure. The likeliest reasons are acquired staff joining before their cuts landed, or a different FTE basis in the S-1, but the post gives neither. The "roughly 4x Cloudflare" comparison uses the 2025 number. (Wiki arithmetic on the post's figures, not a filing check.)
• SaaStr's figure contradicts itself. "$15m–$20m in revenue with 3 humans" gives $5–6.7M per human. The text says about $6M (§5), the benchmark list says about $3M (§7), and the chart says ~$6.7M. The $3M figure cannot be produced from three humans and $15–20M. None of the figures counts the agents' cost or the vendor AI SDRs that run ~90% of outbound. Cite as a range of $5–6.7M, self-reported, with the $3M left unreconciled.

The ladder mixes a typical-company figure, a selected median, quarterly-annualized estimates and one company's self-report on one axis. It shows that the ratio can be pushed up quickly by cutting the denominator. It is not evidence on AI Investment Story, Not Efficiency Story's population question.

The agent-run end: 3 humans, ~20 agents#

SaaStr is the extreme Lemkin offers, at much smaller scale: 3 humans and about 20 production agents, consolidating down from a peak near 30, with about 6 doing most of the daily work. On his account the agents run quote-to-cash end to end (Closed Won in Salesforce, contacts, bill.com invoicing, collections with a 7-day escalation; autonomous after 3 supervised deals, "one incorrect invoice since", with no denominator given), commissions, paid-campaign builds (a human still hits publish), ~90% of outbound via vendor AI SDRs, and the system of record (Notion cancelled after ~7 years). This is the AI-Native Organization mapping run at a real, if tiny, company.

The workload claim is the useful part, and it matches what Human Attention as the Agent-Output Bottleneck measures at Anthropic. Humans keep "judgment calls, relationships, and checking the agents' work". A recent 13-hour day went to "checking, re-ranking, and fixing, with no new features built", and "we spend far less time doing tasks and much more time reviewing them." The operating rule he draws is the post's most transferable line: "If you cut 50% and expect agents to run unsupervised, you'll be disappointed. Someone has to own what the agents produce." At Atlassian, that is the managers and engineers who stayed. At SaaStr, it is the author. The cut is sized by how much review the remaining humans can absorb, not by how much work agents can start.

Connections#

• Firm AI-Spend Intensity and Headcount Growth — the population split this page sits on the incumbent side of: intensive AI spenders and hypergrowth companies add headcount while AI-citing public incumbents cut 10–20%. This page adds which function was cut and what followed, for two names on ICONIQ's list
• AI Investment Story, Not Efficiency Story — the investment-then-efficiency crossing measured on operating data. Lemkin's per-function rule is the same crossing stated as operator advice, and his revenue-per-employee ladder is not evidence on that page's matched-cohort question
• Human Attention as the Agent-Output Bottleneck — "someone has to own what the agents produce": a 3-person firm's review-heavy day as the small-scale case of Anthropic's measured shift from doing to checking
• Organizational Complements to AI — Cloudflare's order of operations (agent usage up 600% before the cut) is the complements-first sequence. Bending Spoons' NRR of 95% shows what the efficiency route gives up
• AI-Native Organization — SaaStr's 3 humans and ~20 agents as a running instance of the agents-as-staff mapping, with the review load attached
• Role Averaging, Not Role Elimination — Cloudflare's plan to rehire for different roles in 2027 is a change of role mix, not elimination
• Small-Business AI Adoption and Growth Expectations — the small-firm end of the size distribution: US employer firms using AI expect to add staff (net employment expectation 32.7 vs 15.2 for non-users), measured as expectation rather than as a cut announced or made

Open Questions#

• Does Cloudflare's cut stay a cut? Prince expects more employees in 2027 than at any point in 2026. Settler: Cloudflare's FY2026 and FY2027 10-K headcount, with G&A as a share of revenue, read against its growth rate. If headcount recovers while G&A share stays down, the cut was a function swap rather than a reduction.
• Did Atlassian's R&D cut cost product growth later, once cloud migration stops supplying it? Settler: Atlassian's FY2027 quarterly cloud growth against its 18% ARR guidance, with the Data Center decline separated out.

Sources#

• Cloudflare Cut 20%. Atlassian Cut 10%, Mostly Engineers. Bending Spoons Cuts Up to 75%. All Three Are Growing Faster or Earning More. So How Lean Can B2B Teams Really Get? — Jason Lemkin, Cloudflare Cut 20%. Atlassian Cut 10%, Mostly Engineers. Bending Spoons Cuts Up to 75%…, SaaStr, 2026-10-17 (practitioner-opinion). Cited for the scorecard (image, transcribed in the raw), the three company sections (filing-derived figures as quoted by the author, not re-checked), the per-function cut ranges (author's judgment), the revenue-per-employee chart (viewed at compile; matches the transcription) and SaaStr's agent setup (self-reported). The SaaStr per-human figure is internally inconsistent ($6M / $3M / $6.7M), and Bending Spoons' $2.57M does not reconcile with Q1 2026 revenue over IPO headcount (~$1.05M). Both are flagged above
• 2026 State of Scaling: The Great Sorting — ICONIQ, 2026 State of Scaling (September 2026, empirical). Cited for the public-company cut timeline that lists Cloudflare and Atlassian (p.20, secondary) and the 2026 top-quartile $100M+ ARR-per-FTE of $478K (p.42)