Palantir just put a number on enterprise AI spend: 1.94 billion dollars in a single quarter, up 93%. A Munich court told Suno that training on GEMA's catalogue was not fair game, and ordered it to open the books. A survey of 1,000 managers found 59% now use AI to help decide who gets laid off. Money, law and headcount all moved in the same week.

Palantir reported 1.935 billion dollars of revenue for Q2 on 3 August, up 93% year on year. US commercial revenue hit 764 million dollars, up 149%, and US government revenue reached 809 million dollars, up 90%. The company closed 220 deals worth at least 1 million dollars, 73 of them over 10 million, and raised full-year guidance to roughly 8.15 billion dollars. Adjusted operating income came in at 1.194 billion dollars, a 62% margin.

Why it matters: The standing objection to enterprise AI spend was that it was pilot money dressed up as production. A 62% adjusted operating margin on 93% growth is difficult to explain as pilot budget. Read the shape of it though: 809 million of that revenue is US government, and 220 deals across a company this size means concentration. The money is real, and it is landing with a small number of vendors who sell deployments and charge for outcomes.

What to do:

Source: CNBC

On 31 July, the Munich Regional Court's 42nd Civil Chamber ruled against Suno in GEMA's copyright case (42 O 763/25), covering six songs including “Rasputin,” “Atemlos” and “Mambo No. 5.” The court rejected the text-and-data-mining defence, finding the works “were not only analysed during training but retained in the models in a reproducible form,” and dismissed a US fair use argument because simple prompts produced output substantially similar to the originals. Suno must stop training on the works, disclose the revenue it earned from them and pay damages, with the amount still to be set. The judgment is not yet enforceable pending appeal. Days earlier, Universal, Sony, Warner, Believe, BMG, Concord and HYBE proposed chart eligibility rules that would require a track to be substantially human made.

Why it matters: Memorisation just became a legal test rather than a research footnote. “We only analysed the data” stops working as a defence in Germany once a model can reproduce recognisable originals from a plain prompt. Any business shipping a generative feature trained on third-party content now has to prove what its model retained, which is an evidence problem your policy page cannot answer. The labels are squeezing distribution at the same time.

What to do:

Source: JUVE Patent

A ResumeTemplates.com survey of 1,000 US managers with direct reports, published 31 July, found 59% use AI when deciding who to lay off and 58% when deciding who to fire. One in four use it often or all the time. More striking: 17% let AI make the call often or all the time with no supervision, while 57% said they would never do that. A separate poll released the same week found 87% of Americans want a human to sign off before AI ends someone's job.

Why it matters: A layoff is a documented, litigable event. If the input to that decision is a model output nobody can reconstruct six months later, the employer carries the evidentiary risk, not the vendor. 17% of managers running that unsupervised is live exposure today, and the gap between what managers say they do and what 87% of the public expects is the shape of the next employment claim.

What to do:

Source: HR Dive

UNAM invalidated roughly 3,000 of more than 150,000 entrance exam results after AI surveillance from contractor Territorium Life flagged suspected cheating, with about 21,962 undergraduate places at stake. Around 58,000 applicants have now been called back for in-person control exams, and the 10 August semester start is in doubt. Students report acquaintances buying programs for about 150 dollars that solved the exam automatically, while others who scored well, including one applicant with 94 out of 120, are stuck in limbo.

Why it's interesting: The detection worked and the institution still lost, because it had no proportionate way to act on what it found.

Key takeaway: Detection without a graded response process turns one cheating problem into 58,000 admin problems.

Source: CNN

The science YouTuber, who has 3.2 million subscribers, slowed uploads across his channels after viewers spotted the phrase “I appreciate the pushback” in a Complexly video and questioned whether he had written it. Green admitted using ChatGPT for research and said “the level of dopamine I've been getting from interacting with LLMs is not healthy for me or good for the world.” He said he plans to make fewer, more deliberate videos.

Why it's interesting: The audience caught the tell before anyone disclosed anything, and the cost landed on the creator's output schedule.

Key takeaway: Your audience is now pattern-matching your writing for AI, so decide your disclosure line before they decide it for you.

Source: TechCrunch

Intelligence, the company behind Design Arena, raised a 7.9 million dollar seed led by Index Ventures with Conviction, A* and Valkyrie participating, announced 3 August. The platform shows people A-versus-B comparisons of AI-generated websites, images and other visual work, and sells the resulting preference data to labs training media models. Co-founder Grace Li said design was “the missing bottleneck for a lot of these models to make improvements,” and 5.3 million people have used it.

Why it's interesting: Automated benchmarks can check whether code runs. Nobody had a clean way to score whether a layout looks good, so somebody built a market for the opinion.

Key takeaway: Subjective quality is now a data category with a price, which is worth knowing if your product competes on taste.

Source: TechCrunch

Context: With 59% of managers now using AI on layoff decisions and 17% doing it unsupervised, the useful move is not banning the tool. It is knowing exactly where AI already sits inside decisions you would have to defend in front of a tribunal. Use this to build that register in one sitting.

Prompt: You are an operations and employment risk advisor. I run [company type] with [number] employees in [country]. Here is how we currently use AI: [list tools and where they are used]. Build me a decision register with four columns: the decision, whether AI touches it today, whether that decision is reversible, and who signs off by name. Then flag every row where an irreversible decision about a person has no named human signature, and give me the specific control that closes each gap.

Do not: Do not recommend that we stop using AI. Assume the tools stay and the controls have to work around them.

If/Then: If I have not told you which decisions are legally protected in [country], then ask me before you rank the risks rather than guessing.

Example: A 40-person agency, Northgate Studio, ran this and found three rows with no signature: shortlisting for redundancy, contractor renewals and performance improvement plans. They added a named partner sign-off on the first two and kept the prompt and output on file for both. When a contractor challenged a non-renewal six weeks later, the file answered it in a day.

As of 30 July, Amazon Bedrock Agents, launched in November 2023, is now Bedrock Agents Classic and sits in maintenance mode. Accounts that have not called CreateAgent or InvokeInlineAgent in the past 12 months get a 403 error when they try. Existing agents keep running, pricing is unchanged and no end-of-life date has been set, but the model catalogue is frozen and AWS points new builds at AgentCore.

Why it's important: This is the first major cloud agent framework to be effectively retired, less than three years after launch. Anyone who built on it is not broken today, but they are on a platform that will never get another feature, and the move to AgentCore is a CLI job for simple agents and real engineering work for anything with custom orchestration.

Business takeaway: Budget agent framework migration as a recurring cost, and keep your business logic outside the vendor's orchestration layer.

Source: AWS documentation

Simile, founded by Stanford PhD Joon Sung Park, raised a 200 million dollar Series B at a 2 billion dollar valuation led by Greenoaks, five months after a 100 million dollar Series A. The company builds simulated users for product and marketing research, and CVS Health is both a marquee customer and an investor in the round.

Why it's important: A 2 billion dollar valuation on synthetic respondents is a bet that a simulated panel is good enough to replace a real one for a large share of decisions. It is cheap and fast, and it is trained on how people behaved before, which makes it least reliable at the exact moment behaviour changes.

Business takeaway: Synthetic panels are useful for screening many ideas quickly, but keep real customers in the loop for the decision you cannot reverse.

Source: TechCrunch

Alibaba launched QwenWork on 2 August, folding QoderWork, MuleRun and Wukong into one enterprise agent platform that runs desktop, cloud and collaborative agents and plugs into DingTalk. It runs on Qwen3.8-Max, a 2.4 trillion parameter model with about 95 billion active parameters and a one million token context window, and is in public beta as a web app and a PC client.

Why it's important: The capability gap between Chinese and US enterprise agent platforms is now small enough that price decides it for a lot of buyers. So the buying question moves to jurisdiction: where your operational data sits, and whose law reaches it once it is there.

Business takeaway: If you are pricing agent platforms, put data residency and legal jurisdiction in the comparison table alongside cost per task.

Source: GuruFocus

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas