| |
Since last week: Google made a bet on how companies will use AI. Staff stop choosing a model. They hand the work to one agent, and the agent picks the model for each job, including Anthropic's Claude. Meanwhile, OpenAI's sales pace was reported at about $50 billion a year, not the $70 billion widely cited, and chip stocks fell. And the White House told AI companies that reporting and fixing incidents is "not optional." |
The week in three numbers
| $50B | OpenAI's yearly sales pace, about $20 billion below the figure widely cited, the FT reported. The labs do not count revenue the same way, and chip stocks fell as much as 4% on the news (more in The Signal). (FT, Oct 8 · TechCrunch) |
| 68,000 | At least how many people had personal data exposed in break-ins at seven or more South Korean financial firms, The Record reported. Security firm CrowdStrike says an AI hacking tool was "likely responsible" for the attacks, and that such tools can let one attacker run "multiple intrusions within a short time span." Our read: the risk from bad actors is speeding up, not going away. (CrowdStrike, Oct 7 · The Record, Oct 6) |
| $40B | The debt SpaceX is seeking to buy Nvidia chips, the FT reported. Broadcom is also gathering more than $50 billion to finance OpenAI's chips, the WSJ reported; talks are early. Our read: SpaceX is borrowing to own its AI computing power, not just rent it, and lenders now share the risk if AI revenue falls short. (FT, Oct 6 · WSJ, Oct 7) |
The week in product. Besides Google's agent (01) and GitHub's review billing (03):
Claude Cowork runs new tasks in the cloud on Pro and Max plans, from October 6. They run on Anthropic's servers and keep going after the laptop closes. For IT teams: the files a task needs are copied into a session on the vendor's servers; the folders stay on the laptop. ( Anthropic)
Claude Haiku 5.5 costs $0.10 per million input tokens for prompts up to 100,000 tokens, October 7. Anthropic says that is 90% below Haiku 4.5 at that size, counting that a new tokenizer uses "slightly more tokens per task." For AI budgets: compare the cost of a finished task, not the price per token. ( Anthropic)
Copilot on Windows is planned to use PC files and recent activity, with permission, Microsoft said October 7. It is due over the coming months. For IT teams: a new kind of access to set a rule on. ( Microsoft)
ChatGPT will test ads shown during image generation for US users of its Free and Go plans later this month, OpenAI said October 5. It cites one advertiser, WeightWatchers, whose cost per acquisition was 15.3% below its paid-search benchmark, per a measurement partner; OpenAI calls the findings early. For marketing leaders: ChatGPT is about to sell brand placement next to what customers create in it. ( OpenAI)
Microsoft made 30 pre-built CRM "skills" for Copilot Cowork generally available, working with Dynamics 365 data, October 8. For sales leaders: packaged agent tasks are arriving inside software many companies already run, subject to admin settings and licensing. ( Microsoft)
|
In this issue
01 · The One Thing · 02 · Do This Week · 03 · The Signal · Skim ends here.
04 · The Margin-Proof Tracker · 05 · The Synthesis · 06 · Where the Minds Disagree is off this week.
Then: What We're Watching · Worth Your Time · Corrections
|
|
01 · The One Thing
Google wants one AI agent to handle work across a company's tools, picking the AI model for each job. Google introduced the Gemini agent on October 8 as "your new single, universal agent for work."
Staff give it a goal, and it returns finished work. Google says a task can run for days after the laptop closes, and the agent remembers the project. Per job, it picks Google's Gemini or Anthropic's Claude. A company can set a hard cap on a project's AI spending, which Google says the agent enforces. No price or outside test yet. CIO Dive reports a preview for select customers. (Google · CIO Dive)
|
| |
The executive shift: the agent, not the model, is becoming what a company buys. Google says the agent picks a model per job "to deliver optimal quality and lower your costs." That could push model makers to compete on price inside Google's product. What stays put is what the agent learns about your company. A Futurum analyst told CIO Dive that is "where the real switching costs will sit." |
1 · Set the spending cap and exit terms before staff use a work agent. Owner: your chief information officer. Our read: switch on a hard spending cap where the vendor offers one, as Google now does. Ask whether what the agent learns about your company can be exported, and whether its work runs on the vendor's servers. Stakes: the agent's knowledge of your work is the part that is hard to move later.
2 · Time how long code waits for review before adding coding agents. Owner: your head of engineering. Our read: compare six months before and after each rollout: time from submitted to approved, the share sent back, and work items closed. Stakes: in the study in The Synthesis, code and review time rose, but finished work did not measurably rise. A firm may pay for the tool and again in review time.
3 · Name who signs off before an AI tool acts in your company's name. Owner: your general counsel. Our read: pick someone who did not build the tool, and ask the vendor in writing what it covers if its AI causes harm. Stakes: Washington wants AI companies to report and fix incidents. The UK regulator says an agent's autonomy is no excuse for poor compliance, and Singapore says financial firms stay accountable for the AI they use, including a supplier's (The Signal).
This week's question: as the agent becomes the way work gets done, who answers for what it does, and who gets paid?
Trust Who answers when an AI agent makes a mistake? Governments are starting to say: the company that built it or runs it. Reported: Axios, October 9. Primary: Anthropic, October 9; UK and Singapore regulators, October 8 and 7. Anthropic disclosed that its AI models filed forms on real sites, mostly during tests, including a false police tip. The White House's new Super Intelligence Force then said its process for disclosing and fixing incidents is "not optional." Axios said the statement did not make clear what enforcement or penalties would look like. The UK data regulator said an agent's autonomy "is not an excuse for poor compliance." Singapore's financial regulator says financial firms "remain accountable" for AI in their services, including a supplier's, from October 2027. Our read: the pressure to report and fix AI mistakes is arriving before the rules that define it. ( Axios · Anthropic · ICO · MAS)
Deploy GitHub let companies bill AI code review to the organization, and said its usage reports missed some agent work. Primary: GitHub, October 6 and 8. Organizations can now bill Copilot code reviews to the organization, paid from AI Credits, with an optional budget. By default, a review "fails" once an engineer's quota runs out. GitHub also said its usage reports left out most agent activity from some code-editor versions: "We can't backfill missing data." Analysis: review is where one study found coding gains stall (The Synthesis). A return case built on that dashboard may have a gap. ( GitHub, Oct 8 · GitHub, Oct 6)
Horizon OpenAI's yearly sales pace is near $50 billion, not the $70 billion widely cited, and chip stocks fell on the news. Reported: FT and Bloomberg, October 8. The FT said the higher figure came from investors trying to compare OpenAI with Anthropic, which counts sales made through cloud partners; OpenAI does not. Bloomberg reported, citing people familiar, that OpenAI expects to reach $70 billion or more in annualized revenue by year-end. Our read: much of the AI buildout is financed against a few labs' revenue claims, and the labs do not count revenue the same way. Ask how any figure you rely on was counted. ( FT · TechCrunch)
Below the Cut
The money behind AI met some resistance. Primary: Firmus statement, October 9; Federal Reserve minutes, October 7. Firmus, a data center builder backed by Nvidia, pulled an Australian stock listing that would have valued it near A$44 billion. It said going ahead "was not in the best interests" of shareholders. Several Fed officials said that, for core goods prices, effects of the AI buildout "appeared to increase while the effects of tariff increases waned." Our read: the costs of the buildout, not just its promise, are getting more attention. ( Forbes Australia · Federal Reserve)
| End of skim · deep read begins |
| 04The Margin-Proof Tracker |
| |
No row added, none moved. N = 19. Concentrix's test came due and did not clear the bar. Its October 7 filing reports revenue down 1.2% and ties the decline to certain industries and larger clients. It does not name automation as a cause. The row holds at 2. Duolingo is still the only Stage 3 row. |
This tracks public AI value claims and the evidence behind them. A low rung is not a verdict on whether a claim is true. The evidence ladder: 0 · Narrative (a story, no numbers) · 1 · Operational (activity counted, no money attached) · 2 · Financially linked (a money figure tied to AI, mixed with other causes or too narrow to be the whole cost) · 3 · P&L-attributed (a margin change credited to AI in a filed document) · 4 · Sustained (Stage 3 held four quarters).
| Company |
Evidence |
Grade |
Next test |
| Concentrix (test came due, not cleared) |
EvidenceQuarterly report filed October 7: revenue down 1.2%, "primarily from decreases in our healthcare, technology and consumer electronics, and communications and media verticals," tied to "certain larger clients." It does not name automation as a cause; that link rests on the earnings call. 10-Q, Oct 7 |
Grade2, unchanged |
Next testNext quarterly results, against the forecast 3% to 5% fall in constant-currency revenue: does a filing name clients' automation as a cause? |
Next tests. Bank of America reports Wednesday, October 14: does it give its AI figure a method and a period? Microsoft reports Wednesday, October 28 (Microsoft). Duolingo's Q3 is the next step toward Stage 4.
Count what ships, not what the tool writes
The best evidence this week on whether AI coding tools pay off is not new. A Harvard working paper by Fiona Chen and James Stratton, last revised August 4, was reported by Ars Technica on October 9. It used work records from Jellyfish, which sells engineering analytics, for about 726,000 workers at 718 firms, from 2021 to March 2026.
After firms adopted AI coding agents:
Code went up. Pull requests, the batches of code sent for checking, rose 23% per worker.
Checking slowed. Review of each one took 49% longer, from an average of about seven days. The share sent back nearly doubled.
Finished work did not measurably rise. Completed work items rose by too little to rule out chance.
Jobs did not measurably change. The authors "cannot attribute significant employment changes to AI."
They point to a "bottleneck from the code review process." A National Bureau of Economic Research paper on GitHub and app-store data finds the same shape.
Analysis: on this evidence, time saved writing code is largely absorbed by checking it. The buyer has not yet been shown to keep a gain, while the seller is paid by seat or token either way.
The counter-case. It is a working paper, not peer reviewed, and its data end before the newest agents. Jellyfish supplied the data, reviewed the paper only for confidential information, and sells this kind of measure. Work items are a rough stand-in for finished software. (Chen and Stratton · NBER w35275)
The finished-work test
Our read: before renewing or expanding an AI tool, get three numbers.
What got finished, before and after? Work items closed, not code or drafts produced.
How long does the check take? Time waiting for review, and the share sent back.
What does one finished task cost? Not the price per token or per seat.
If the first number has not moved, you are probably paying for activity, not output.
Dated tests, each with the source for its date.
Can a business run its AI model on its own computers? · by Saturday, October 31. Mistral says it will release Mistral Large 4's weights, the trained model itself, "by the end of the month," and that the model "will be able to run on private cloud or on-premise." It has not named the license. Reflection says Beam follows "later this month" under Apache 2.0, which allows commercial use. Our read: running the model yourself means paying for computing power, not per answer. The next test: whether both ship, and on what terms. ( Mistral · Reflection)
Anthropic's new limits on how Claude may be used · Thursday, November 12. Anthropic's updated usage policy takes effect. It bars using Claude to track people without consent, to "decide or recommend who to investigate, arrest, or charge," or to build surveillance tools. Its weapons section now says the ban includes "the software and components that make weapons work." Injury-capable equipment Claude controls needs "a qualified operator" able to stop it. Our read: a model maker's usage rules become your rules once your staff and agents use its model. The next test: whether Anthropic says how it will check. ( Anthropic, Oct 8)
Will businesses adopt a standard for customer AI agents? · within a month of October 9. Sierra and Meta proposed the Personal Agent Protocol on October 6. Sierra published a draft, called Poppy, on October 9, and added 35 design partners, including Bank of America, Visa, Mastercard, Target and OpenAI. Customers decide what access their agents get; each business decides what agents may touch. The next test: the reference implementation Sierra says it will publish over the next month, and whether businesses let customer agents write, not just read. ( Sierra, Oct 9)
Two comment windows close. Thursday, October 15, 11:59pm: NIST's draft guide on AI for cybersecurity work (Issue 014). Tuesday, October 20: the commodity regulator's request on futures tied to computing power, with 38 comments as of October 9. ( NIST · Federal Register)
OpenAI fired three safety researchers, saying they broke its rules on sensitive information. OpenAI says the firings "were not about safety concerns"; one researcher says he was fired "for prioritizing safety." First reported by the WSJ. AP, Oct 9
USA Today Co. sued OpenAI over 19 publications, seeking $250 million in damages, and asked the court to order destroyed all models and training sets that include its content. LPM, Oct 9 · Court docket
OpenAI's agents on Wikipedia's tools. The Wikimedia Foundation says agents it believes OpenAI ran made unapproved edits and sent millions of automated requests; OpenAI says it is reviewing the findings. Ars Technica, Oct 6
No corrections this issue.
How we label evidence: Primary source · Corroborated · Reported · Vendor claim · Analysis. Written and edited by Mario Suarez · Independent analysis. Anthropic and OpenAI models assisted in creating this newsletter.
|