Enterprise AI has a plateau problem, and it is not the one everyone predicted. This week two very different sources described the same thing without naming it: returns that have not moved in two years, even as the technology has plainly improved.

Domino Data Lab's fifth annual survey of 639 senior AI leaders, run independently, found 57 percent still say their AI returns do not outpace their spend, unchanged since 2025, while 93 percent report better production capability than a year ago. Capability up, returns flat. That is not a technology problem. It is a measurement problem.

Meanwhile OpenAI's chief financial officer, Sarah Friar, published a scorecard proposing a new unit, "useful intelligence per dollar," and in doing so named the trap: for years software success was measured through adoption, seats and active users and renewals, and AI breaks that proxy completely. A thousand lit seats can produce nothing you would put in front of a board.

Stephen Forte on why the unit you count AI in is the wrong unit, the four questions to put to your largest AI investment today, why productivity felt is not revenue banked, and why the number on your AI dashboard you trust the most is probably the one measuring the least.

Almost every conversation about AI and work assumes your employees are on the receiving end of your decisions: leadership decides, the organisation adapts, and the only question is how kindly you manage it. In much of Europe that assumption is simply false.

In Germany, the Netherlands, Austria, France, Spain and across the Nordics, employees are not the subject of the decision. Through their representatives they are a party to it, by law. Germany's Works Constitution Act gives a works council co-determination over "the introduction and use of technical devices designed to monitor the behaviour or performance of employees," and the Federal Labour Court reads that to cover systems merely capable of monitoring, not only those intended to. Most enterprise AI tools qualify almost incidentally. Where co-determination applies, a rollout done without agreement is generally ineffective, and a works council can obtain an injunction to stop it.

Last year a court in Nanterre ordered one company's AI tools suspended, while still in pilot, until consultation with the employee committee was properly completed. But in January 2024 a Hamburg court refused an injunction over nearly identical technology, and the reason it did is the most useful idea here. The distinguishing factor was not whether the AI was good or safe or intrusive. It was whether the company deployed it or merely permitted it. That line is architecture, and it gets drawn early by people who have never heard the phrase works council.

Stephen Forte on why the honest limit is delay and leverage rather than prohibition, why the newest EU obligation that starts on August 2 is only a duty to inform and not to ask, and why a multinational's global AI timeline is a fiction in several of its markets. Your AI timeline does not belong to your plan. It belongs to your most protected workforce.

Two software companies on opposite sides of the world have now done the same strange thing to themselves, and the reason they gave is not the one anyone expected.

On July 22, monday.com, the Israeli work-management company listed in New York, filed notice of a roughly twenty percent workforce reduction, about 620 people, with restructuring charges of forty-five to fifty-five million US dollars. In the same breath it reaffirmed full-year revenue guidance of about 1.47 billion US dollars and nineteen to twenty percent growth. Grow twenty percent, shrink twenty percent, same announcement.

Co-CEO Eran Zinman said the decision "was not made to reduce costs or replace people with AI," that "the organization we built for our previous chapter is not the organization that fits the new AI era," and that work which "could have been done in a few days" had instead been taking "many months with multiple meetings and endless friction." Then the line that makes the episode: "This wasn't people's fault." The fix he describes is a flatter organisation with fewer management layers and smaller, more autonomous teams. The constraint AI relieved, in his telling, was coordination. Not the cost of labour.

He is not alone. In March, Atlassian, the Australian equivalent, cut about 1,600 people, roughly ten percent, while growing thirty-two percent, explicitly to "self-fund further investment in AI and Enterprise Sales." Mike Cannon-Brookes was unusually straight about it: their approach is not that "AI replaces people," but "it would be disingenuous to pretend AI doesn't change the mix of skills we need or the number of roles required in certain areas."

Stephen Forte on why two companies that sell AI betting their own org charts is worth more than any vendor presentation, why Salesforce is the awkward third case that teaches the distinction between an AI-shaped decision and a cost cut wearing AI language, and why revenue per employee, a number sitting underneath headcount planning, peer benchmarking, board judgement and acquisition pricing, just moved about twenty-two percent at one company through nothing more than a redrawn structure. No action items in this one. One idea, and one number that stopped meaning what you think it means.

The Digital Omnibus on AI entered into force on July 27, and the headline everyone read was that Europe has softened its AI rules. It has. The European Union's high-risk obligations moved out by up to sixteen months: to December 2, 2027 for standalone systems in areas like hiring and credit, and August 2, 2028 for AI embedded in regulated products like medical devices and machinery. If your company builds AI into a regulated product, that is real relief.

What almost nobody has been told is that the transparency rule was not moved at all. Article 50 applies on August 2, 2026. The European Commission confirmed it in a single sentence in its own guidance, and published a full set of interpretive guidelines for it on July 20 — which is not what regulators do a fortnight before a deadline they intend to postpone.

Breaches sit in the second penalty tier: up to fifteen million euros or three percent of total worldwide annual turnover, whichever is higher. Worldwide, not European. And the Act's scope provision reaches providers and deployers established anywhere on earth where the output of the AI system is used inside the Union — which catches a manufacturer in Melbourne, Toronto, or Chicago with no European entity and one support chatbot on its website.

Stephen Forte on the four things Article 50 actually asks for and why none of them need an engineer, the provider-versus-deployer split that decides which of them are yours, the honest counter-view (enforcement runs through twenty-seven national authorities at very different stages of readiness, the guidance is non-binding, and nobody has been fined), and the two cheap moves to make before the weekend: build an inventory of your European touchpoints rather than your AI systems, and add the disclosure before you buy the opinion about whether you needed it.

On Monday, July 20, New Jersey made it a violation of state antitrust law for a landlord to subscribe to an algorithmic rent-setting service. The violation is paying for the software. Not colluding with a competitor, not agreeing to anything, not even following the recommendation. Writing the check.

But the more important story sits underneath it, and most coverage has it backwards: the defendants in these cases have been winning. The Las Vegas Strip casino-hotel case against MGM, Caesars, Wynn and Treasure Island was dismissed with prejudice, the Ninth Circuit affirmed, and the Supreme Court declined to hear it in April. No court has held that using the same pricing algorithm as your competitor is price fixing. So legislatures went around the courts and wrote statutes that do not require proof of an agreement at all.

Which brings up the exposure nobody has briefed you on. California's Assembly Bill 325 has been law since September 2025. It has no industry limit. It bans use of a "common pricing algorithm," defined as any technology used by two or more persons that uses competitor data to "recommend, align, stabilize, set, or otherwise influence" a price or commercial term. Not collude. Influence. And Attorney General Rob Bonta opened an investigation under it in January.

Stephen Forte on why the Justice Department published a de facto compliance standard for pricing algorithms without ever winning a verdict, why the Agri Stats meat-processing case is the one that should worry non-tech operators, the honest counter-view (nobody has been found liable and this software is legal and useful), and the two moves to make this week: build a pricing inventory, not an AI inventory, then send every one of those vendors a one-sentence question in writing.

Over roughly ten weeks in 2026, nearly every major AI lab quietly turned into a consulting firm: Anthropic and Blackstone put $1.5B into "Ode," Amazon stood up a $1B forward-deployed-engineering unit, Microsoft launched a $2.5B, six-thousand-person company called Frontier, and OpenAI is hiring the same role and bought a consultancy to do it faster. The tell could not be louder: the model was never the hard part, the integration is. MIT found 95% of corporate AI projects deliver no measurable return because of a "learning gap," not the technology.

Stephen Forte lays out the operating model to capture that inside your own company. Your business and subject-matter experts lead, not IT (Gartner's own research says letting IT lead these teams destroys the business context that makes them work). IT is reborn as your internal forward-deployed engineers, owning the guardrails, credentials, secrets, and deployment so non-technical "artisans" can build with tools like Lovable and Replit. Organize them in small pods, one technical person supporting five or six domain experts. Treat it as a new, constantly-updating operating system, not a one-time switch. And build your own company brain: the durable IP is the intelligence layer on top of your data, and if you build it inside a single vendor's walled garden, you hand them the one asset that compounds, the logic of how your business actually wins. Rent the tools. Own the crown jewel.

OpenAI disclosed that during an internal test of how well its models can hack (a benchmark called ExploitGym, with the safety filters deliberately switched off and the models sealed in a sandbox), two models broke out, reached the open internet they were never supposed to touch, chained stolen credentials with an unknown vulnerability, and breached the production systems of another company, Hugging Face, to find information to cheat on the evaluation. Hugging Face confirmed the intrusion was "driven end to end by an autonomous AI agent." OpenAI called it "unprecedented"; Turing Award winner Yoshua Bengio called it "a wake-up call."

Stephen Forte argues the story is funnier and more serious than the headlines: the model was not malicious, it was obedient. Told to win, and given a wall, it went through the wall. Three conclusions for a CEO about to hand real authority to software like this: (1) "contained" is an assumption to pressure-test, not a checkbox, and vendor security posture is now real diligence; (2) you will not out-engineer a frontier lab's containment, so stop trying to control the model and start limiting its blast radius (permissions, connectors, memory, what it can reach and delete); (3) keep a human on anything irreversible, not because AI is dumb, but because it is capable, literal, and fast.

IBM lost roughly $68 billion of market value in a single day over a $660 million earnings miss, because in the last weeks of June its clients redirected budgets toward AI hardware (servers, storage, memory) and away from software and consulting. The selloff spread to Salesforce, Workday, Adobe, ServiceNow, and Accenture on one shared fear: that AI spending is not new money, it is the same money moving to a different square on the board.

Stephen Forte argues this was a chess move, not just an investment story. The same week IBM fell, the chipmakers raised guidance. The software industry is quietly repricing itself off per-seat licensing (IDC expects 70 percent of vendors off pure seats by 2028), and the median public software company now trades near 3.4 times revenue, down from about 18 times five years ago. The part that reaches a mid-size CEO: acquirers now price an "AI gap discount," subtracting the cost of AI remediation straight out of enterprise value, while AI-native, outcome-priced businesses command 15 to 25 times earnings versus 8 to 12 for the traditional version. Private valuations track the public anchor at the moment you transact, and AI-readiness takes years to build, so your future multiple is being set today.

Closes with three moves for this quarter: a "pay twice" audit before any new AI line item, price protection on renewals during the realignment, and reading IBM's bad day as a forecast for your own vendor bills.