The market just repriced AI: it pays for proof and control, not capability.

The shift: from Deploying AI capability you can buy to Controlling and proving AI you can stop

Enterprise AI stopped paying for capability you can buy and started paying for AI you can govern, count, and prove.

In March the board asked the deployment questions: are we using it, are we spending enough, do we have the skills. By April adoption hit 88 percent of organizations and got called the floor, and the binding constraint slid downstream from the model to the ownership, measurement, and control layer wrapped around it. From May into June that flip hardened into a control-and-trust market. The new board questions are who owns it, can we prove it net of the bill, and can we stop it. The capital settled the argument on a single Tuesday in June: investors put 200 million dollars into a company that watches agents the same week a marquee capability vendor lost 53 percent of its revenue, and the only new thing the biggest vendors packaged to sell was control.

March artifacts

• Are we using it? March board question one, answered by 88 percent adoption that no longer separates anyone. This Week's AI Signals: What a CEO Should Carry Into the Boardroom
• Are we spending enough? March board question two, now buying 2.59 trillion dollars of global AI that mostly cannot be tied to a result. Three AI signals this week, and the board question they raise
• Do we have the skills? March board question three, reframed by Brookings as a management distribution decision, not a skills gap. The AI adoption gap is a management decision in disguise

June artifacts

• Who owns it? June board question one: one enterprise expected 500 agents a week and found 2,000, with no control plane over them. How many AI agents has your Series B actually shipped?
• Can we prove it net of the bill? June board question two: only 21 percent of S&P 500 firms can cite any measurable AI benefit at all. How to build an AI ROI model your Series C board will trust
• Can we stop it? June board question three: Gartner expects 40 percent of enterprises to demote or decommission an autonomous agent by 2027. How to set up agent operations before your first 2 a.m. incident

Data signals

• 11% - Only 11 of 100 scored production agents were both capable and well-defended, while 40 percent paired broad power with thin defense. Shipping an agent is no longer evidence of anything. Rank the high-reach, low-defense agents and harden that short list before writing another blanket policy.
• 3 to 5% - Across three independent sources the proven-value rate holds near 3 to 5 percent, not survey noise but a structural ceiling. Only 5 percent capture substantial value at scale, 3 percent have measurable ROI, and just 4 percent of 951 firms cleared 30 percent savings. Read your own AI line against this band before the H2 budget call.
• 12 months in 4 - ServiceNow and Uber each burned a full year of AI token budget in roughly four months as agents ate subscription economics. Consumption pricing removed the headcount ceiling finance relied on for a decade. Name an owner of the agent cost ledger and cap spend per workflow before metered billing lands.

Unresolved tensions

• Everyone adopts, almost nobody captures value: AI is universal: 88 percent of organizations use it for at least one function and adoption is openly called the floor. vs Proven value holds near 3 to 5 percent at scale, with only 5 percent of companies capturing substantial AI value.. Adoption no longer differentiates anyone, so the whole competitive question moves to whether one named owner runs the workflow, the measurement, and the cost ledger. A board still reporting adoption is reporting the floor. Evidence: this-weeks-ai-signals-ceo-boardroom, how-to-write-ai-roi-report-cfo-board, ai-adoption-vs-productivity-myth.
• Confidence runs inverse to control: 93 percent of security executives say they are fully confident in their AI governance. vs Only 16 percent of organizations have high confidence detecting AI agent-specific threats, the capability that confidence claims to cover.. The most exposed organization is the one that mistakes a stated policy for a working control system, because confidence shows green on the board slide right up until an agent surprises it. More than half of generative AI use has already moved to mobile that those tools cannot see, widening the shadow surface beneath the confidence. Evidence: shadow-ai-board-question, ai-agent-scope-violations-board-series-c.
• Consolidation cures sprawl but concentrates lock-in: Standardizing on one AI platform is the disciplined fix for tool sprawl, overlap, and ungoverned burn. vs It does not remove risk, it concentrates it: only 6 percent of executives believe they could stop using their main AI vendor without interruption.. The safe-looking move and the safe move diverge, so the real decision is which dependencies are cheap to escape, the model, versus expensive, the platform and the agent memory. Getting that wrong trades a sprawl problem for a hostage problem. Evidence: single-ai-platform-safer-myth, ai-budget-capability-or-oversight.
• AI gets blamed for cuts the CFO made: Layoffs at Meta, Cloudflare, and Snap are reported as AI workforce replacement with productivity claims of 2x to 100x attached. vs Only about 8 percent of 2026 job-cut announcements actually cite AI, and the lift comes from org redesign plus AI, not deployment.. If a CEO believes deployment produces the productivity, they cut before the redesign and the gains never arrive. Most headline cuts are capex reallocation wearing an AI label, and copying them imports the cost without the redesign. Evidence: ai-layoffs-cfo-decision-myth, ai-productivity-multiplier-myth-ceo, ai-entry-level-pipeline-myth-ceo.

Monday moves

• Move AI value accountability to the CFO: Only 21 percent of S&P 500 firms can cite any measurable AI benefit, yet 76 percent of programs where the CFO owns AI value report significant returns. The Q3 budget call decides next year's AI line, and the owner you name now sets whether the board sees proof net of the bill or another promise. This Monday: Move AI value reporting from the CIO to the CFO in next month's board pack, with one P&L line per workflow stated net of the cost to run and oversee it. The Two Percent Fix: What Happens When Finance Owns AI Value
• Name one owner per agent with a kill switch: One enterprise expected 500 agents a week and found 2,000, and only 25 percent of CIOs can monitor every agent their teams created. No vendor console discovers the cron-job script an engineer wrote in March, so a one-page agent register and a tested kill protocol have to come before any platform purchase. This Monday: Commission a one-page agent register naming every production agent, its human owner, and a kill protocol tested to stop any agent in under five minutes, due before the next platform contract. The agent register question Series B boards should ask before Q3 close
• Cap agent spend per workflow before the meter: ServiceNow and Uber each burned a full year of AI token budget in roughly four months, GitHub flipped Copilot to metered credits June 1 and Anthropic on June 15, and one allotment lost 8 percent in the first two hours of metered billing. The ceiling finance relied on disappears with consumption pricing, so it has to come back through the contract before the meter runs. This Monday: Write a hard per-workflow spend ceiling and a named cost-ledger owner into every agent vendor contract before it signs, plus the renewal guardrail the GitHub and Anthropic meters now make load-bearing. Does your Series C need an agent FinOps function before Q3 close?

What to watch

• The EU Commission classification consultation on its 167-page draft high-risk guidelines closes June 23, and the draft reads intended purpose from product documentation and marketing. (within 90 days): It tells whether the December 2027 deferral actually bought time or whether classification, the decision that determines if obligations apply at all, is being settled now based on what companies ship and say this quarter. Trigger: The EU Commission finalizes guidance that ties high-risk status to current marketing and docs, forcing 2026 product and positioning changes at named vendors despite the postponed compliance date.
• Where new AI capital lands: money flowing into agent-watching and agent-constraining companies versus pure capability vendors, after Coralogix raised 200 million dollars the week C3.ai fell 53 percent. (Q3 2026): It confirms whether the market has durably repriced AI on provable control rather than capability, which would make the trust and oversight layer, not the model, the place value accrues. Trigger: Another capability-only vendor re-rates sharply down while an oversight vendor posts a marquee raise, or a Big Four firm or a hyperscaler ships agent governance as a named product line.
• Agent FinOps moving from anecdote to a standing named function as more enterprises follow Uber and ServiceNow in capping per-tool spend after burning a full-year token budget in roughly four months, and Anthropic in moving agent billing to non-rollover credits. (H2 2026): It will tell us whether boards force a named owner of the agent cost ledger, the operating role that did not exist 18 months ago, before consumption pricing quietly breaks unit economics. Trigger: A Fortune 500 or named late-stage company discloses a hard per-employee or per-agent spend cap and ties the agent bill to features shipped in a board pack.
• Gartner's call that 40 percent of enterprises will demote or decommission autonomous agents by 2027, tracked against an agent-rollback rate already at 74 percent. (H2 2026): It marks whether the deployed-to-dependable gap is forcing a real retrenchment, separating firms that built a control plane of register, kill switch, and owner from those that shipped agents they cannot stop. Trigger: A high-profile public agent incident at an S&P 500 or named enterprise triggers a visible rollback and a board-level postmortem within the quarter.