The AI decisions that matter this week

☕ The 60-second read

1. Frontier models now ship every ~11 days. Your architecture can't be married to one lab. Decision → design for model portability now.
2. The agent ROI gap is a discipline gap, not a model gap. 80% of well-run agent deployments show measurable ROI — but only ~23% of organizations see significant ROI. The delta is evals, ownership, and success criteria. Decision → fund the eval layer before the next model.
3. 1M-context is now table stakes at last year's 200K price. OpenAI, Anthropic, Google and DeepSeek all shipped it. Decision → revisit RAG you built to dodge context limits.
4. Capital is a tidal wave: ~$300B in Q1 venture, 80% of it AI. Google, OpenAI and Microsoft are buying everything. Decision → write portability into vendor contracts.
5. The talent/skills gap is the #1 adoption barrier — above budget. Decision → your constraint is people, not models.

If you do one thing this week: pull your top agent initiative and ask the three questions in Signal of the Week. If you can't answer them, you're in the 22% headed for negative ROI.

🔭 Signal of the Week

The agent ROI gap is a management problem in disguise

The headlines this quarter are euphoric: by mid-2026, ~54% of enterprises are running AI agents in production, and 80% of those deployments report measurable ROI. But read the same data one layer down and a very different picture appears: only about 23–29% of organizations report significant ROI at the company level, and 22% of agent deployments are at negative ROI by the 12-month mark.

Both are true. The reconciliation is the whole story:

What to do — the three questions to ask of every agent initiative:

1. Success criterion: Can you state, in one sentence, the number that defines success and who is accountable for it? (If not, you're in the 41%.)
2. Eval harness: Do you have an automated way to know when the agent is wrong before a customer does? (No evals = no production.)
3. Workflow fit: Is this a high-volume, rule-bound, costly-error process — or did you pick a flashy demo? (Pick boring and expensive.)

The contrarian implication: the right next investment for most enterprises is not a better model or more pilots — it's an eval and observability layer plus a single accountable owner per agent. The org chart is now a bigger lever on ROI than the model leaderboard.

📡 Intelligence brief

🔢 The number

Global enterprise AI spend in 2026 — with 86% of organizations increasing budgets — yet only ~29% report significant ROI. The money is committed; the returns are not. This is a management problem, not a budget problem. The leaders who close their ROI gap in 2026 won't be the ones who spent the most — they'll be the ones who scoped, measured, and owned the best.

🧭 Contrarian take

Everyone is hiring a "Head of AI." Most shouldn't — yet.

With AI strategy under board scrutiny, the reflex is to appoint a senior AI leader. But 75% of execs admit their AI strategy is "more for show," and 54% of the C-suite say adoption is "tearing the company apart." Dropping a new Head of AI into that — with no eval infrastructure, no success criteria, and no owned workflows — manufactures a scapegoat, not a solution.

The non-consensus move: earn the role before you fill it. Ship two boring, measurable agent wins first. Then hire a Head of AI to scale what's already working. Capability-first, title-second.

👀 What we're watching next

• Anthropic's internal "Mythos" model — reportedly a "step change," possibly landing Q2.
• xAI's Grok 5 — a rumored 6T-parameter MoE, which would be the largest model publicly announced.
• The first wave of agent-ROI postmortems — as 12-month marks arrive on 2025 deployments.