Every Friday I read the week's signals from two sources, the investments of early-stage investors and the findings of research institutions, the same evidence base my annual report on AI and work is built on. The investors are never the story. Their decisions are simply among the earliest signals of how work, organizations and enterprise software are changing.
For five months, artificial intelligence was the most cited reason for job cuts in the United States. In August it fell to fourth place. In the same week that this number appeared, the market paid the highest prices yet recorded for software that takes over whole jobs, and the builders of enterprise systems moved to the assumption that a machine, not a person, will operate them. Those two facts sound like a contradiction and are in truth one finding. The story and the transaction have parted ways. The label is leaving the statistics. The handover of work is not.
Work is now bought by the finished task, and the supervision comes bundled
On September 8, Cognition (cognition.com), the company behind the software engineering agent Devin, raised more than 2 billion dollars at a valuation of 48 billion, nearly double its level of May, in a round co-led by Andreessen Horowitz and Accel and joined by Benchmark, Bessemer Venture Partners (bvp.com) and General Catalyst, among others. By the company's own account, run-rate revenue has grown from 492 million dollars in May to almost 900 million, with Nvidia, GE Aerospace, Citi and Mercedes-Benz among the customers. The more instructive document appeared the same day: Bessemer's investment memo, titled "Devin's next commit", which states more plainly than any pitch what is actually being purchased. The unit of work is no longer the suggestion a person accepts or rejects. It is the ticket, the whole task handed over the way one hands work to a colleague, while the human moves into the role of the architect who sets goals and priorities. And the memo justifies the bet not with model quality but with governance, sandboxed environments and audit controls as the condition under which a bank is allowed to delegate at all.
Harvey, the legal AI company, showed the same pattern a day later. It raised 550 million dollars at a valuation of 15.5 billion, according to reporting by TechCrunch and LawSites, with Kleiner Perkins among the returning investors, on the back of more than 400 million dollars in annual recurring revenue by its own account. Part of the announcement was an acquisition: Guardrails AI, a firm that builds safety and control tooling for agents. Supervision of machine work is becoming part of the product rather than a task left to the customer. What the market prices at these levels is not a tool. It is completed work with the oversight included.
Enterprise software is being rebuilt around the assumption that the agent is the user
A generation of enterprise software assumed a person at the keyboard, filling forms so that the system of record stays true. The week's smaller rounds show that assumption being retired. Lightfield raised 47 million dollars in a Series A led by Andreessen Horowitz, with Greylock among the investors, for what it calls the CRM for companies that run on agents. The system gathers its own context from mail, calendar, calls and chat instead of waiting for a salesperson to type it in, and by the founder's account whole Salesforce installations are migrating over. Forus (forus.com) raised 150 million dollars at a 3 billion valuation, led by Bain Capital Ventures with General Catalyst returning, for agents that run the entire administrative path between a prescription and a treatment, insurance approvals, appeals, assistance programs and pharmacy routing. The company reports that the median approval time fell from more than seven days to 1.1 days and that nursing staff spend 70 percent less time on paperwork, figures that are its own. And Bessemer led a small round with a sharp thesis: 10 million dollars for Harmoni, which automates the administrative work around factory operators, time capture, ERP transactions, quality reports, and argues its case with the labor shortage in American manufacturing rather than with headcount savings.
The buyers are moving in the same direction as the builders. Rogo (rogo.com), the analyst platform for finance, announced a strategic investment from Barclays, BNP Paribas, Citi, MUFG and Société Générale, among others, which by the company's count puts nine global banks on its investor list. The institutions whose junior analysts have always built the models, memos and decks are buying into the platform that now produces them.
The label leaves the statistics while the adjustment happens at the entrance
According to Challenger, Gray & Christmas (challengergray.com), whose monthly count of announced job cuts is a private tally and not an official statistic, artificial intelligence fell from first to fourth place among stated reasons in August: 3,462 cuts attributed to AI, the lowest monthly figure since December 2025, after five consecutive months from March through July in which AI led the list. Over the year to date, AI remains reason number one, with around 116,000 announced cuts, roughly 22 percent of the total. Restructuring now leads the monthly list. Nothing in this week's market suggests that the technology retreated in August. What retreated is the willingness to name it.
The official numbers of the same week locate the real adjustment. The Bureau of Labor Statistics counted 162,000 new jobs in August, far above the 12-month average of 31,000 and the roughly 53,000 the consensus expected according to CNBC, with unemployment steady at 4.1 percent. Inside that headline, the information sector lost another 23,000 jobs. Weekly initial claims stood at 206,000, historically low. And iCIMS, a maker of recruiting software reporting from its own platform data, counted job postings up 13 percent on the year while actual hires rose only 2 percent and fell for the second month in a row. People are rarely quitting and rarely being fired. They are also rarely being hired. The adjustment sits at the entrance, which is exactly where a transformation without a label would sit. One more figure from the same report: AI-related postings make up about 4 percent of American hiring demand, while the share of workers teaching themselves AI skills has risen from 22 to 30 percent within a year, faster than employers train them.
The week also brought the most serious attempt so far to put boundaries on the debate, from the vendor side of all places. The Anthropic Institute published economic scenarios for transformative AI along with an interactive explorer. In its moderate and substantial scenarios, the American unemployment rate in early 2030 stands at 3.9 and 4.6 percent, inside the historical normal band. The job-killer picture appears only in an extreme scenario. The labor share of income, however, falls from 60 to 56 percent already in the substantial scenario. These are model outputs under stated assumptions, not forecasts, and precisely as such they move the question: away from whether the jobs disappear, toward who is paid for the work that remains.
Signals at the margin
Compute keeps turning into a procurement item. NEA co-led an 875 million dollar round for Positron at a 5 billion valuation, an inference chip built on the bet that memory and bandwidth, the running costs of a machine workforce, decide the economics. Figure (figure.ai), the humanoid robot maker, committed 3.5 billion dollars for compute at Nscale according to Forbes, nearly twice everything it has ever raised. Oracle reported a cloud order backlog of 664 billion dollars, and Qualcomm entered the data center chip business with Amazon as anchor customer, warrants worth around 4 billion dollars and a purchase frame of up to 60 billion, per CNBC. In defense, Mach Industries raised 600 million dollars at a 3.7 billion valuation, doubling in three months, with Ribbit Capital, long a pure fintech specialist, among the investors. The physical substructure of machine work keeps absorbing capital, whatever the layoff statistics say.
The week at a glance
Cognition, over 2b at 48b valuation (Sep 8). The finished task becomes the unit of purchase, governance included. (Andreessen Horowitz, Accel (co-leads), Benchmark, Bessemer, General Catalyst returning)
Harvey, 550m at 15.5b, per press reports (Sep 9). Legal AI buys its own supervision layer with Guardrails AI. (Kleiner Perkins among returning investors)
Lightfield, 47m Series A (Sep 9). The CRM assumes agents, not salespeople, keep the record. (Andreessen Horowitz (lead), Greylock, Coatue, Lightspeed)
Forus, 150m at 3b valuation (Sep 8). Agents run the whole administrative path between prescription and treatment. (Bain Capital Ventures (lead), General Catalyst returning)
Harmoni, 10m Series A (Sep 9). Factory admin work is automated against the labor shortage, not against headcount. (Bessemer (lead))
Rogo, strategic round, amount undisclosed (Sep 10). The banks buy into the platform doing their analysts' groundwork. (Barclays, BNP Paribas, Citi, MUFG, Société Générale a.o.)
Positron, 875m at 5b valuation (Sep 10). Inference cost becomes the payroll of the machine workforce. (NEA, Atreides, Valor a.o. (co-leads))
Mach Industries, 600m at 3.7b valuation (Sep 10). A longtime fintech specialist now funds defense manufacturing. (Ribbit Capital, Sequoia, Infinite, Bedrock)
Figure, 3.5b compute commitment, per Forbes (Sep 4). The robot maker commits double its lifetime funding to compute. (Nscale (supplier, not a round))
What this says about the AI label
My annual report, Work After AI, appears next Tuesday, September 15. Its founding edition carries a German year title that translates as "The great error about AI the job killer", and the first of the three forces it names is the AI label: statements about artificial intelligence and jobs follow the position of the speaker, not a shared body of data. A company cutting costs reaches for the label as long as it helps, and sets it down when it stops helping. This week the label demonstrated the thesis by leaving. After five months as America's most convenient explanation for job cuts, AI dropped to fourth place in a month in which the market paid 48 billion dollars for a maker of software colleagues and five global banks bought into the automation of their own analysts' groundwork. The transformation did not pause in August. The explanation just stopped being useful.
That is why the report argues for watching the transactions and the entrance rather than the announcements. What organizations buy tells you which work is being handed over. Whom they stop hiring tells you where. Both signals ran clearly this week, and neither of them carried the label.
Reading the week
None of this is a scoreboard of who invested how much. The deals matter as evidence, and the evidence of this week points one way: the story about artificial intelligence and jobs is decoupling from the transaction. The announcements soften, the label rotates out of the press releases, and the measuring gets harder at exactly the moment the handover gets faster. Whoever keeps steering by the announcements will find that the label disappeared long before the change did, and that they did not see it go.
Gerhard Kürner is CEO of 506.ai, the European platform for Service-as-a-Software and agentic engineering, and author of Work After AI. Around 1,000 conversations with boards, owners, and PE funds across DACH and Europe.