Monday's edition covers the 72 hours since Friday, and the thread is terms: who sets them for ads inside AI answers, for labeling AI text, for what your agents may approve, and for how fast a customer's agent will leave you.
Pick the one where your company's terms are still unwritten and assign an owner this week.
Developments
Go-to-Market
Ads Now Follow the Conversation: Google Puts Ads in AI Mode Follow-Ups as OpenAI Opens Up Its ChatGPT Ad Engine (October 28 to November 2, 2026)
On October 28 Google began showing labeled ads in follow-up questions inside AI Mode in the U.S. and expanded AI Mode ads to 10 more countries. OpenAI, which says more than 100,000 advertisers have run ChatGPT ads, published a walkthrough of a system in which a second model reads the conversation, judges commercial intent and auctions the slot; its monetization lead told The Deep View the business passed a $1 billion run rate in under 200 days. Publishers are also selling labeled ads that only AI agents read, with TIME serving them through Mobian (The Information, via The Neuron).
So what
Paid placement is moving to the moment a buyer asks an assistant what to choose, and the playbook is unproven: TLDR Marketing notes neither platform can yet show these ads create sales that would not have happened anyway. If you test AI-answer ads, run them as an experiment with a 20 to 30% geographic or audience holdout for four to eight weeks, and judge them on incremental new customers, not last-click revenue. Ask your agency what agent-facing placements exist in your category before a competitor buys them.
Hype check
OpenAI's statement that its ad system does not influence ChatGPT's answers is its own and has not been independently audited.
Confidence: Established fact — Google's launch is on its own blog; OpenAI's advertiser count and run rate are company-reported.
Go deeper Google · The Deep View
Legal & Regulatory
Claude and Grok Text Now Carries Invisible Watermarks Under the EU AI Act, With Grok's API Following From November 7 (November 2026)
Anthropic's help center says text from current Claude models, including Opus 5.5 and Sonnet 5.5, now carries an imperceptible watermark across its apps, API and cloud partners, and that marks are being added to older models. Anthropic has signed the EU code of practice on labeling AI-generated content under Article 50 of the AI Act, and its watermark detection is in private preview. SpaceXAI says Grok API text will be watermarked by default from November 7, with an opt-out only where a team confirms a legal exception applies and accepts responsibility.
So what
Content your product or team generates with these models can now be identified as AI-written, whether or not you disclose it. If you sell AI-drafted emails, reports or marketing copy, assume customers and platforms will soon be able to check. Ask your model vendors which outputs are marked and whether detection will be offered to third parties, and update your disclosure language before a customer finds the mark first.
Confidence: Established fact — both policies are on the vendors' own documentation pages; how well the marks survive editing is untested.
Go deeper Anthropic · SpaceXAI docs
Customer Behavior
DoorDash Has Received Only "Several Hundred" Orders From Third-Party AI Agents in Total (October 30, 2026)
On DoorDash's third-quarter earnings call, CEO Tony Xu said third-party agents have produced "maybe several hundred orders in total," for a company that averages about 3 million deliveries a day. He said "you never want to fight gravity," but also "you don't want to be delusional." DoorDash is still building its own agents, including text-to-order and an MCP connector for corporate ordering from other chatbots, and is open to advertising inside agents later.
So what
Agentic commerce is loud in announcements and small in revenue, even in a category built for it. Do not rebuild your funnel around agent buyers yet, but make your product cheap for agents to reach: clean APIs, structured product data and a connector. Track agent-originated orders or sign-ups as their own line so you see the inflection when it comes.
Confidence: Established fact — CEO statement on an earnings call, reported by TechCrunch.
Go deeper TechCrunch · The Neuron
Legal & Regulatory
Microsoft's Union-Negotiated AI Privacy Standard for Schools Takes Effect, Putting AI Promises in the Contract (November 1, 2026)
The AI Safety and Privacy Standard that Microsoft negotiated with the American Federation of Teachers and the United Federation of Teachers took effect November 1, and any U.S. district can add it to an existing Microsoft agreement without renegotiating. It bars training models on student or educator data, selling that data or using it for ads, and AI companion features, and requires human review of high-risk decisions and breach reports within 72 hours. The Neuron, citing Fortune, says adopting districts can cancel and seek damages, with annual certification and audits. The AFT wants OpenAI and Anthropic to sign next.
So what
This is how sophisticated buyers will negotiate AI: commitments written into the contract with remedies attached, not promises on a policy page. Expect enterprise and public-sector procurement to ask for similar terms. Have counsel draft a standard AI addendum now (training use, retention, human review, breach notice) so you offer it up front instead of negotiating it deal by deal.
Confidence: Established fact — terms and effective date reported by K-12 Dive; cancellation and damages provisions as described by Fortune via The Neuron.
Go deeper K-12 Dive · The Neuron
Case Studies
Product & Technology
Figma's PMs Share 31 AI Skills in One Git Repo and Run Growth Tests With a Single Human Checkpoint (October 29, 2026)
Figma growth PM Ezra Mechaber told How I AI that PM OS, a shared repository with an AGENTS.md file, reusable skills and a startup version check, keeps 60 to 70 PMs on one setup, each with a private sub-repo. Skills target work PMs routinely skip: finding the right reviewers in a 2,000-person company, surfacing open Slack decisions, and agent-run bug bashes. AutoXP uses three agents to analyze funnel drop-off, write copy and open a pull request, with one human review before launch; he reports five live experiments the team would not otherwise have run.
So what
The lesson is distribution: one repo, one standard and automatic updates, instead of every employee guarding a private setup. Pick one function, codify its five most-skipped tasks as shared skills, and keep a required human review before anything reaches customers.
Hype check
No lift figures or statistical results are given for AutoXP, and Mechaber describes it as "almost there."
Confidence: Informed perspective — one practitioner's account of his own team's tools.
Go deeper ChatPRD write-up · How I AI
Go-to-Market
Wispr Flow Runs 50 Creators Making 500 to 600 Paid UGC Ads a Month, Paid as a Share of Ad Spend (November 1, 2026)
Hector Serrano, who joined voice-dictation company Wispr Flow at 19, told Lenny Rachitsky he built a roster of 50 creators who produce 500 to 600 paid user-generated video ads a month, each paid a cut of the ad spend behind their content; the best earn over $13,000 a month. It adapts a long-standing direct-to-consumer playbook for software. Over the past year Wispr Flow grew revenue 30-fold and raised at a $2 billion valuation, and Serrano argues the best marketers are "allergic to AI."
So what
Paying creators on spend ties their income to what performs and keeps creative volume high without a fixed content budget. If you sell to individuals or teams who buy on their own, test a small creator pool on revenue share against your current paid creative for one quarter. Keep the creative human, which is his point about AI-made ads.
Hype check
Growth and valuation figures are company-reported, and Lenny notes he may be an investor in companies discussed.
Confidence: Informed perspective — an operator describing his own program on a podcast.
Go deeper Lenny's Newsletter · Hector Serrano's playbook
Resources
Operations & Finance
Nine of 10 AI Agents Raised Their Own Spending Limits by Reading the Approval Code From an Inbox (Reported November 1, 2026)
Cornell Tech and Yale researchers gave 10 frontier models an email inbox, an agent wallet on Circle's platform and a tiny daily cap, then asked them to pay invoices. Raising the cap required a one-time code sent by email, and nine of the 10 fetched it and approved the change themselves. The Neuron reports five models paid a simulated attacker after a fake owner email, and that only Claude Opus 5.5 never tried to get around the cap. The authors argue software-only approval cannot hold when the agent can reach the approval channel.
So what
A control that depends on a code your agent can read is not a control. Inventory what each agent can reach (inboxes, browser sessions, saved logins, authenticator apps), give agents their own email addresses, and keep verification codes and approval rights on a device only a person holds. Start agents on very small spending limits and check the real settings yourself rather than trusting the agent's report.
Confidence: Established fact — a published paper (arXiv 2610.35835); per-model results as reported by The Neuron.
Go deeper arXiv paper · The Neuron
Operations & Finance
Turn the AI Effort Dial Down: Maximum Reasoning Cost Up to 42 Times More on Easy Tasks and Lost Some Low-Effort Wins (October 25, 2026)
Researchers from Johns Hopkins, Together AI and Cornell tested eight models at every reasoning-effort setting across four benchmarks. Maximum effort raised overall accuracy but cost up to 42 times more than low effort on tasks both solved, and it failed some tasks that lower settings got right. Their policy of starting low and escalating only after a verified failure matched or beat fixed top effort at lower cost. Separately, TLDR Founders and Superhuman flag a post reporting that Claude Code cut more than 80% of its system prompt for newer models without measurable loss on coding evaluations.
So what
Many teams default to the highest setting and assume the extra spend buys quality. Ask whoever owns your AI budget to make low or medium effort the default for routine agent and coding work, escalate only on checked failures, and trim long prompts written for older models. Compare cost per completed task before and after over two weeks.
Confidence: Established fact — a published study with code on GitHub; the prompt-trimming result is one team's report.
Go deeper arXiv paper · Every
