Reading the AI Headlines

Read enough WSJ coverage of corporate AI adoption and you start to notice the genre has its own grammar. The vocabulary is consistent across outlets: AI transformation, agentic workflow, productivity uplift, AI center of excellence, and at least one phrase per quarter that nobody can actually define, like semantic discoverability membrane. The narrative arc is just as predictable. There is the panic appointment, the bulk license buy, the press release with the seven-figure hours-saved number, the photo op at the vendor summit, the CEO equity event, and the layoff announcement that lands a quarter later and gets rebranded in the earnings call as AI-enabled efficiency. Every beat is recognizable because every beat is happening somewhere right now.

The funny part is how legible the whole sequence becomes from inside corporate AI work.

The headline economy

Open the Wall Street Journal or the Microsoft customer-stories page on any given Tuesday and the genre is recognizable. Impact, a hundred-person firm, saved twenty thousand hours a year on Copilot with a posted ROI of $1.72 million omg_wow.gif. Newman’s Own tripled their marketing campaigns and saved seventy hours a month. Vodafone’s legal team got back four hours per person per week. A UK government pilot of twenty thousand users averaged twenty-six minutes a day saved. Microsoft’s own legal department says it’s running thirty-two percent faster. Sixty-four percent of the Fortune 500 now have an active Copilot deployment, up from forty percent at the end of 2024.

The numbers are real reports from real firms; what they actually measure is rarely the thing the headline implies. Self-reported time savings depend on the honor system. Vendor case studies are written by the vendor’s marketing team and presented as third-party validation. Adoption percentages count licenses provisioned, not weekly active users on a real workflow. Salesforce’s pricing rollout for Agentforce is a useful tell: launched at $2 per conversation in late 2024, walked back after customer revolt, replaced in 2026 with a flex-credit scheme at $500 per 100,000 credits, twenty cents per standard action, fifteen cents per voice action. The vendors are no more certain about what their products are worth than the buyers are.

What’s underneath

The bulk license purchase precedes any plan for the licenses, because the purchase is the deliverable. Organic adoption settles in the low double digits, which everyone knows and nobody says. Mandatory training rolls out next, aimed at making the adoption number presentable for the vendor’s next case study. The metric ships, the vendor publishes, the CEO does the photo op. Equity vests against the announcement, not against any later measure of usage. The workforce reduction lands the next quarter, framed in earnings as AI-enabled efficiency even though the cuts were on the roadmap before any seat was bought.

The CEO is paid for the announcement, the VP who ran the rollout is paid for visibility (and on a long enough timeline, for an exit into their own fund), the vendor is paid for seats, the board is paid for the stock pop. The only group whose incentives point toward actually using the tool is the employees, and they are the ones being reduced. The gap between licenses sold and value created is where the arbitrage lives.

What the receipts look like

The actual practitioner view is not on the customer-stories page. A few-hundred-person professional services firm switches from a flat-rate enterprise AI plan to usage-based billing because the vendor moved the goalposts. Projected annual spend, if left alone, lands somewhere between a quarter and a third of a million dollars. Most users sit between zero and a few dozen dollars a month. A handful of power users rack up a few hundred dollars in forty-eight hours because they are running real work, and the firm has to decide in real time whether that’s a budget problem or proof the tool is doing what it was bought to do.

Behavioral change is what makes a rollout work, and almost no one owns it. Most users default to the most expensive model on the menu without knowing they’re doing it. Vendor admin consoles often lack basic features like sort-by-spend or CSV export. A separate API exists for full per-user observability, which is the worst possible product to deploy first because it turns the whole effort into surveillance theater. The firms that handle this well watch month-to-date spend per user, accept that their dead seats are dead seats, and own their training internally instead of outsourcing it to the vendor’s case-study team.

I’m not above any of this. The instinct to celebrate seat count instead of usage is a strong one, and the headline numbers feel like victories even when the underlying engagement stays flat.

Questions worth asking

If you are a buyer, a board member, or anyone watching a rollout get announced, the short list:

What is weekly active usage on real workflows, and how is “real” defined? Who owns adoption after the procurement check clears? If adoption is below twenty percent at ninety days, does the budget shrink or get rationalized into a bigger one? Is the rollout being announced externally before it’s been validated internally? Are headcount actions being timed to the same news cycle as the AI announcement?