AI Hiring Pulse — Week of 2026-05-18

📊 Pattern of the week — the Agent tilt has crossed a threshold

58% of every new AI role opened last week is at the Agent station. Foundation-model layers (Pretrain + Post-train) together made up 8%. The hiring market is no longer building models — it is building things on top of models.

Agent is over-represented in new flow versus the active book (57.7% vs 46.8%) — meaning the marginal hire is more application-focused than the running portfolio. Conversely Serve is being under-hired at the margin: 11.7% of new vs 16.8% of stock. The infra build-out is past peak. Companies are consuming serving infra from others (OpenAI, Anthropic, Bedrock, Azure AI) and pouring marginal headcount into agent / vertical work.

The mirror image of the headline is the Eval Gate gap — 1.4% of new hires, half its already-low 2.8% baseline. As the application surface area explodes, the eval surface area is shrinking, not growing. This is the second consecutive pulse where Eval Gate hiring is anomalously thin.

🪞 Two ways to open 30+ AI roles in one week

Salesforce and Google both opened more than 30 AI roles last week. Same headline number, opposite shapes.

Salesforce is making a pure agents-on-top-of-someone-else's-model bet. Thirty Agent hires and no investment, this week, in anything else — no data plumbing, no fine-tuning, no eval, no infra. That is consistent with the public Agentforce posture, but it is also a one-sided strategic position: every quality and cost lever lives at whichever foundation lab Salesforce is sourcing from. Google's distribution is the polar opposite — every layer of the stack is staffed in the same week.

The applicant-relevant read: if you want pure agent / orchestration work, Salesforce's hiring shape says they will keep opening these roles. If you want post-training or eval work, Salesforce has shown you in one week that you should look elsewhere.

🚨 4 inflections this week

1. JPMorgan opened 29 AI roles; the rest of the US bulge-bracket banks combined opened 0. Goldman Sachs, Bank of America, Citi, Wells Fargo, Morgan Stanley, American Express — all zero new AI postings in our tracking window. JPMorgan's shape was full-stack (18 Agent, 6 Ship, 1 each at Data / Post-train / Serve / Eval). This is the first week we have seen one US bank account for the entire sector's AI hiring flow. Capital One — the prior multi-quarter leader — opened 11, suggesting the post-Brex absorption is complete and JPMorgan now carries the banking-AI hiring book.

2. Box (BOX) disclosed $7M of Q1 FY26 revenue tied to AI-driven early renewals, on its 2026-05-27 earnings call (transcript already ingested in our trajectory pipeline). Aaron Levie's framing: AI should be "margin neutral over the medium term," enabled by falling inference cost. This is the first time Box has put a dollar figure on AI-attributed revenue. It is small ($7M against ~$280M quarterly) but it is directly attributable, which most enterprise AI revenue claims are not. Adds a grade-able number to the "Box agents thesis" we have been tracking.

3. Adobe / Salesforce / J&J all in the top 7 for new AI hires. Adobe (35), Salesforce (32), Johnson & Johnson (28). The presence of J&J at this rank — a pharma + medical devices conglomerate, not a tech company — is the more interesting half. J&J's shape was balanced: Agent-heavy but with Data, Post-train, Serve, Ship representation. Pharma AI hiring is no longer niche.

4. Serve hiring is decelerating at the margin. 99 new this week against an active book of 1,306 — implied turnover of 7.6% per week is below most other stages. NVIDIA alone accounted for 12 of those 99. Strip NVIDIA and Serve is 87/847 = 10.3% of new hires versus 16.8% of stock. The serving-infra build-out is past its peak; companies are increasingly buying it instead of building it.

🔮 Prediction watch

Predictions to grade

• Box's "AI margin neutral over the medium term" target (Aaron Levie, 2026-05-27 earnings call) is now grade-able. The $7M Q1 FY26 disclosure is a baseline. Watch the AI-attributed revenue figure against AI inference COGS in Q2-Q4 FY26 calls; if AI gross margin contracts, the thesis fails.
• The Eval Gate gap. Last pulse flagged Capital One's Eval=0 anomaly. This pulse flags it sector-wide: Eval hiring at 1.4% of new flow vs 2.8% of stock means the eval workforce is shrinking in relative terms while application surface area explodes. Either eval is being absorbed silently into Post-train / Agent roles (the benign read), or the industry is shipping agents faster than it can grade them (the alarming read). We will pull a sample of 30 Agent JDs next pulse and check whether they contain meaningful eval responsibilities.
• Salesforce upstream hires = 0 will stay 0. Prediction: across the next four weekly pulses, Salesforce opens fewer than 5 Pretrain + Post-train + Serve roles combined. Cashable at 2026-06-22. If this holds, Salesforce is on record as a pure agents-on-foundation-models bet — which has implications for who they will keep paying inference bills to.

Predictions cashed

• The "Capital One post-Brex surge is one-off" call (implicit in last pulse's framing) cashed positive. Capital One opened 11 this week, well below its post-Brex high. Brex absorption is complete; the headline-grabbing 317-active number is now the running baseline, not a growth rate.

📰 Reading list

1. Box Q1 FY26 earnings transcript — for the $7M AI-attributed renewal disclosure and the "margin neutral" framing. The first explicit AI revenue number in this earnings cycle that is not vibes.
2. JPMorgan AI roles in our index — 29 roles in one week is a real shift; the shape (18 Agent, full-stack secondary) is worth eyeballing directly.
3. Salesforce in our index vs Google in our index — same week, same volume, opposite strategies. The clearest single-page case for "AI hiring shape > AI hiring count" we have produced.