The Org Chart Is Being Rewritten by Software That Doesn't Sleep

02What the 12% did differently

We spoke to operators at nine companies — three banks, two insurers, a logistics group, two software firms and an Israeli defense-tech company — who got agents past the pilot wall. None of them started with the model. All of them started with a workflow they already measured. "We didn't buy an agent to do customer service," one insurance COO told us. "We bought it to close the 14 minutes between a claim landing and a human reading it. We knew that number for years. That's why we could prove it moved."

Banking and insurance lead the production charts at 47%; government trails at 14%. The reason is not budget. It is that regulated industries already had measured processes, governed data and human approval checkpoints — exactly the four things every successful deployment we saw had in common. The BCG/Forrester payback data says the same thing from the other side: sales-development agents pay back in 3.4 months because the funnel was already instrumented; legal and compliance take 11.2 months because nobody had measured them before.

"We didn't buy an agent to do customer service. We bought it to close the 14 minutes between a claim landing and a human reading it."

03The new job title nobody had in 2024

Fifty-six percent of enterprises now name a dedicated agent owner, up from 11% in 2024. That is the fastest-growing role in corporate org charts since the chief digital officer — and unlike that title, it comes with a P&L. The best ones we met are not data scientists. They are former operations managers who can read a cost-per-task chart and argue with a vendor about it.

The role exists because the bottleneck moved. Sixty-four percent of companies cite evaluation gaps — no automated way to know whether an agent got worse after a prompt change — as their single biggest blocker. Only 38% run full evaluation coverage. Forty-one percent have already rolled an agent back from production at least once. Twenty-two percent report negative twelve-month ROI, and the post-mortems almost never blame the model. They blame scoping and the absence of tests.

04The cost line that grew 7.2×

Enterprise LLM bills grew 7.2× year over year. For the first time, the model line item is showing up in board decks next to cloud spend, and CFOs are asking the question they asked about cloud in 2019: what is our unit cost, and who owns it? The companies that can answer — cost per resolved ticket, per closed deal, per reviewed contract — are the ones still expanding. The ones that can only report "tokens" are the ones on Gartner's cancellation list.

This is also why the Chinese open-weight wave matters to a CFO in Frankfurt. DeepSeek V4, Kimi K3, GLM-5.3 and Qwen3.8 all shipped this month with open weights, and every one of them pushes the price of a capable API toward zero. The premium Western labs will keep on reliability, safety and tooling — not on the raw model.

05The Israeli angle

Israel's first half closed at $8.4 billion across 129 rounds, and the composition tells the transformation story better than any survey. The largest rounds were not in models. They were in the plumbing: Vast Data's $1 billion at a $30 billion valuation, ScaleOps' $130 million for compute efficiency, ZutaCore's $100 million for data-center cooling, and PointFive's $60 million for cost control. Israeli founders read the same 7.2× number and built for it. Cyera's two $1 billion tranches say the same thing about data security: you cannot let an agent touch data you have not classified.

06The playbook, in one screen

Pick one workflow you already measure. Instrument cost per task before day one. Buy or build evaluation before you buy agents. Put a named owner with budget authority on it. Keep a human approval checkpoint until the rollback count is zero for a quarter. Then — and only then — add the second agent. The 22% coordinating three or more agents in production did not get there by starting with three.