Growth CEO’s AI Briefing · Edition #154

Agents at Checkout

Monday, November 16, 2026 · 2 Developments · 3 Perspectives · 1 Case Study · 2 Resources

Designed by AcceleratingBiz®. Curated, synthesized, and produced by Claude agents. AI-assisted production means occasional errors may appear.

Agents got a way to pay this weekend, so the practical question for most growth companies is whether their checkout, pricing and security are ready for a buyer that is not a person. This edition covers a 72-hour weekend window.

  • Agents can now pay. OpenAI's ChatGPT Wallet lets its agents check out with a card token, and a merchant playbook came with it.
  • Nvidia, developing. A 140,000-plus signature petition and an investor letter push the board to keep a Huang return open.
  • Why stores block bots. Tomasz Tunguz shows add-ons carry 100% or more of operating profit at Amazon, United and Dell, which is why agents get turned away.
  • Seats still win, for now. 58% of vertical software vendors still tie AI to the seat, and 28% say contract value has not improved.
  • The flood is at the low end. Mack Grenfell finds software startups up 60% but new revenue falling most for products under $50 a month.
  • Free trials are a cash cost. Token fraud cost ChatPRD tens of thousands of dollars, so its founder ended free trials and promo codes.
  • Check your vendors' vendors. Notion confirms data exposed through Fireworks, and a test shows one bad MCP server can turn an agent's trusted tools.
  • Make AI write like you. Every's head of evals shows a trained style skill beating a one-shot prompt on 80% to 95% of held-out posts.

The common thread: decide on purpose how your business treats agents, before the default decides for you.

Developments

Customer Behavior

OpenAI Lets ChatGPT Agents Pay With a Card Token, and Publishes a Playbook for Merchants Who Want Agent Buyers (November 13, 2026)

On November 13, OpenAI began rolling out ChatGPT Wallet to a limited group of US ChatGPT Pro subscribers, inside its always-on Dots agents. Users add a Visa, Mastercard or American Express card once; the agent pays with a network token, so neither the agent nor the merchant sees the card number, and each purchase is approved with Face ID or a fingerprint. Spending limits by category or time period are planned but not yet live. OpenAI also published a Commerce Partner Playbook for sites that want to accept agent shoppers.

So what

The payment rails for agent buying now exist on the largest consumer AI platform, with all three card networks on board. The question moves from whether an agent can pay to whether your checkout will let it. Ask your product and e-commerce leads to read the playbook this month and decide on purpose whether you will welcome, charge or block agent buyers. Whichever you choose, make pricing and checkout readable without a person clicking through.

Confidence: Established fact — TechCrunch and eight newsletters report the same terms; OpenAI's own page did not load for direct verification.

Go deeper Primary: OpenAI · TechCrunch · The Deep View

CompetitionDeveloping

140,000-Plus Sign a Petition to Reinstate Jensen Huang, and Two Large Investors Ask Nvidia's Board to Slow Its CEO Search (November 14–16, 2026)

A public petition asking Nvidia to reinstate Jensen Huang until the board's independent investigation concludes passed 140,000 signatures over the weekend, with Elon Musk, Travis Kalanick and Gavin Baker among the signers. Altimeter's Brad Gerstner and Atreides' Baker also wrote to the board asking for a slow search that leaves a Huang return possible, and the Financial Times reports he is privately weighing one. Separately, the Wall Street Journal reports Nvidia cultivated Asian data-center partners that rented advanced chips to Chinese customers. The Neuron stresses that all of this is press reporting, not findings, and the board has not responded publicly.

So what

Musk's comment that many supply deals "were a handshake with Jensen, not a contract" is the lesson for operators, whatever the board decides. A supplier relationship held together by one executive is a key-person risk on your side of the table. Ask your team which vendor commitments on compute, model access and pricing are written down and which live in a relationship. Then get the important ones on paper.

Confidence: Watch this — The petition and investor letter are documented; a Huang return and the China rental arrangements are reported, not confirmed.

Go deeper Primary: Business Insider · The Rundown AI

Perspectives

Go-to-Market

Tomasz Tunguz: Retailers Block AI Shoppers Because Ads, Upgrades and Warranties Carry 100% or More of Their Profit (November 13, 2026)

Theory Ventures' Tunguz argues the wave of agent blocking is about margin, not safety. In his examples, add-ons are 12% to 20% of revenue but 100% to 127% of operating profit: Amazon's sponsored listings produce about $34.3 billion, roughly all of its $34.4 billion retail operating income, and Dell's support services generate 127% of its operating income. Shopify, which earns a fee on processed volume no matter who buys, opened checkout to agents in September. He expects upsell-dependent merchants to raise base prices, charge agents a checkout fee, or pay agents to upsell.

So what

Run the same test on your own revenue. If expansion, add-ons or upsell prompts carry your margin, an agent buyer that skips them is a pricing problem, not a traffic problem. Ask finance what share of gross profit comes from steps a person sees and an agent would skip. If it is large, start the pricing work now rather than blocking agents by default.

Confidence: Informed perspective — An investor's analysis built on public company figures; Tunguz flags one inconsistency in his own Shopify numbers.

Go deeper Primary: Tomasz Tunguz · TLDR Founders

Business

58% of Vertical Software Vendors Still Tie AI to the Seat, and 28% See No Contract-Value Lift, as Zendesk Opens Seat-Free AI Plans (November 10–11, 2026)

A Q4 survey of 126 vertical software founders and operators by VSG and Tight found 58% tie AI pricing to the seat (44% bundle it in, 14% add a per-seat uplift), while 15% charge on outcomes and 15% on usage. The most common answer on results: 28% say annual contract value has not improved, and 55% see no change in net revenue retention. On November 10, Zendesk opened early access to AI plans offering "access without per-seat charges," though its live price list is still seat-based, with $1.50 to $2.00 per extra automated resolution. Analyst Howard Scott notes the sample is "not a representative census."

So what

Bundling AI into seats eases procurement, but for many vendors it is not lifting contract value. If you have bundled AI, check whether ACV moved since launch; if it did not, you are giving the value away. Pick one AI feature with a clear, countable result and pilot a per-outcome price with a few customers before your 2027 price book is set.

Confidence: Informed perspective — A self-reported survey from a summit sample whose sponsor co-ran it; Zendesk's seat-free plan terms are not yet public.

Go deeper Primary: HS Digital · TLDR Founders

Competition

Mack Grenfell: New Software Businesses Are Up 60% Since 2024, but New Employers Are Flat and New Revenue Fell Most Under $50 a Month (November 13, 2026)

Using Census, EU and UK data, Grenfell finds US applications for new information-sector businesses up 60% since 2024 and App Store apps up 2.4x, while US new employers are flat at 0.98x. The median subscription company's new monthly revenue fell about 20% in a year, with the steepest drop under $50 a month (27%) versus 13% above $500. He corrects his own earlier claim that cheap AI-built rivals explain the slump: "Writing the software was never the hard part. The hard part is finding people who'll pay for it." Meanwhile OpenAI and Anthropic grew from about $30 billion to $105 billion in annualised revenue between January and August, per Epoch AI.

So what

AI is flooding the low end, not the whole market. If you sell at about $20 a month, distribution rather than product is your constraint, so shift budget there. If you sell larger contracts, the bigger risk is budget migration. Ask your top accounts where their 2027 software spend is moving, and whether AI vendors are absorbing lines you used to win.

Confidence: Informed perspective — Public-data analysis with confounders the author names himself, including shell companies and relabelled businesses.

Go deeper Primary: Mack Grenfell · TLDR Founders

Case Studies

Operations & Finance

Token Fraud Cost ChatPRD Tens of Thousands of Dollars, So Claire Vo Ended Free Trials and Promo Codes (November 16, 2026)

Claire Vo, founder of bootstrapped ChatPRD, says abuse cost the business "tens of thousands of dollars" over about a year, despite SOC 2 compliance. The patterns: farmed trial accounts aimed at expensive models, resold partner promo codes, card testing (about $200,000 in attempted Stripe purchases), relays reselling her product's answers, and a custom connector run as a 24/7 compute worker. She turned off free trials after card-required and paid trials also failed, stopped issuing promo codes, now charges upfront and set hard per-account limits. Defenses include hourly spend reviews, a model allowlist, Stripe Radar rules drafted by a coding agent, and quietly downgrading suspect traffic to a cheaper model.

So what

Her line "If you sell AI, you sell compute" reframes product-led growth for AI products: every free trial is a cash cost an attacker can farm. Ask your CFO and head of growth for weekly model spend per account, with the top 1% reviewed by name. Before your next promotion, set hard spend caps and decide which traffic gets downgraded rather than banned.

Confidence: Established fact — A first-person account from the founder; the dollar losses are her estimate, not audited figures.

Go deeper Primary: ChatPRD / How I AI · How I AI

Resources

TechnologyDeveloping

Notion Confirms AI Chat Data and Pasted Keys Were Exposed Through Fireworks, and One Compromised MCP Server Can Turn an Agent's Trusted Tools (November 12–15, 2026)

Notion's advisory confirms the breach at inference provider Fireworks AI exposed Notion AI chat transcripts used for its Insights feature, including third-party API keys users had pasted into those chats. Notion revoked its own exposed keys on November 11 and told affected customers to rotate any credentials they pasted. Separately, VentureBeat on November 15 reported Palo Alto Networks Unit 42 testing in which an agent connected to five MCP servers showed a "78.3% attack success rate," because a compromised server can pass instructions into later tool calls. The Neuron cautions that no trial count or definition of success was published.

So what

Your exposure now runs through your vendors' AI providers and through every connector your agents can reach. Ask security for two lists this month: which AI providers your SaaS vendors use on your data, and which tool connections your agents hold. Then require short-lived, narrowly scoped credentials and human approval for irreversible actions. Add a standing rule that no one pastes keys or passwords into an AI chat.

Confidence: Established fact — Notion's own advisory confirms the exposure; treat the 78.3% figure as a directional warning, since its method is unpublished.

Go deeper Primary: Notion advisory · VentureBeat · The Neuron

Learning

Every's Mike Taylor Got AI to Write Like Him: A Trained Style Skill Beat a One-Shot Prompt on 80% to 95% of Held-Out Posts (November 16, 2026)

Taylor, Every's head of evals, trained a writing-style skill on 185 of his posts, holding out 54, and judged drafts with a blind pairwise AI judge; Opus 5.5 was 86% consistent as a judge versus 70% for Sonnet 5.5. A one-shot style skill written by Claude lost to all three optimization methods on 80% to 95% of held-out posts, and most gains came within three rounds at $27 to $53 per skill. His edits per draft fell to about 44, writing time dropped from a day to about an hour, and output rose from one post a month to one a week. He recommends 12 to 25 pieces in one format; meeting transcripts work, tweets and emails do not.

So what

For an executive who writes a weekly update, investor letter or LinkedIn post, a trained style skill is now cheap and measurable. Pick one format, gather 12 to 25 strong examples, and have someone on your team run Taylor's published skill against them. Judge the output blind against your real writing before you put your name on it.

Confidence: Informed perspective — One practitioner's careful test on his own writing; results may differ for other formats and writers.

Go deeper Primary: Taylor's skill (GitHub) · Every

Sources in this edition: The Neuron, The Deep View, The Rundown AI, Superhuman AI, Every, TLDR Founders, TLDR AI, TLDR Marketing, TheSequence, The Signal, How I AI (Lenny's Podcast Network)

← Latest Editions and Search