Good morning. Here's what's worth acting on today.
BREAKING STORY
ChatGPT Wallet: OpenAI's agents can now pay at checkout
On 13 November OpenAI launched ChatGPT Wallet, which lets its dots (OpenAI's AI agents that carry out tasks for you) finish a purchase rather than only recommend one. You add a card once; the dot finds the product, shows you the items, total and card, and you approve each purchase with Face ID or a fingerprint. Payment runs on a token, so the card number is not shared with the agent or the merchant. For now it is open to a limited number of US Pro subscribers with US-issued Visa, Mastercard or American Express cards. TechCrunch notes some retailers block AI agents; Amazon is the largest.
My take: If you sell online, more of your visits may come from an agent that has to read your page and complete checkout without a person clicking around. Check this week whether your store or platform blocks AI agents, and decide on purpose whether you want that. Then make your product pages easy for an agent to read: one clear price, stock status, shipping cost and returns policy in plain text, not inside images. OpenAI's commerce documentation for merchants is the place to start if you want to be included.
MARKET TRENDS
AI chatbot links drove about 5% of Douyin purchases in October
China's Singles Day shopping festival showed this year how much buying now starts in an AI chat. Data from WPIC Marketing + Technologies, reported by Reuters, says roughly 5% of Douyin purchases in October came through AI chatbot links, and more than 20% for considered purchases such as supplements. JD.com's shopping assistant drew nearly 80 million users, more than triple last year, while overall sales grew only 2.4% between 7 and 31 October, according to Syntun. Finimize reported the same Douyin and Syntun figures.
How to make money from it: The share is highest where buyers research before they pay. If you sell supplements, gear, courses or software, publish the page an assistant would want to quote: who the product is for, who it is not for, a comparison with the two alternatives people ask about, and answers to the five questions customers email you most. Keep it factual and current, because an assistant that recommends a product is now one tap from buying it.
GROWTH HACK / DIY
Unlimited seats, pay per outcome
The idea: Stop charging per login. Let the whole team in free, and charge for each finished, checked job against a monthly budget the customer sets. Free invites let your product spread inside an account, and outcome billing means customers pay only when the work gets done. On 10 November, Zendesk launched plans with no per-seat charge and an Outcome Budget, with conversational outcomes starting at $0.99 (as of 14 Nov 2026). Its CEO told TNW Zendesk will end the year with about $400m of AI revenue, up from about zero two years ago, mostly from agents paid per resolution; about 100 customers tested seatless pricing. These are company figures, not audited results.
Do it today:
Define one outcome, such as "an invoice matched to a payment and not reversed within 7 days", and publish what never counts (merged tickets, test data, failed runs), as Zendesk does in its outcome definitions.
Add a check before you charge, such as customer confirmation or no reopen within a set window; Intercom's Fin, at $0.99 per outcome as of 14 Nov 2026, charges at most once per conversation.
Let customers set the cap, with alerts at 80% and no charge above it without approval, plus a small platform fee or minimum spend to cover your costs.
What you'll need: an event log of completed jobs, metered billing in your payment tool, a customer-facing charge list, and about a week to set up before testing with a few customers for two billing cycles.
AGENT LAB
Judge the outcome: a second model watches your agent
Shopify's Cyber Chaos Monkey runs AI agents that attack its own live systems. To keep them safe, a second "judge" agent on a different model reads the worker's full session in real time and can kill the run if it breaks the rules. Shopify reports tens of thousands of engagements without causing an outage or incident. OpenAI's Auto-review uses a similar second agent to approve actions so humans avoid review fatigue.
How it works: Plain code, the orchestrator, sends each proposed action to the judge and waits for allow or block. It also checks the judge is alive with a regular heartbeat. If the judge times out, crashes or returns nonsense, the run ends. A missing verdict counts as a no.
Try it: Take the agent that sends email or touches your store. Write five to ten "never" rules. Before every send, pay or delete step, call a small model from a different provider with those rules, with a ten-second timeout. Route every error to "stop and alert", then kill the judge mid-run to prove nothing goes out. Test it on 20 actions you have labelled, 10 of them bad.
TRENDING AI RESEARCH
People change their opinion on model efficacy even as rankings hold
What it says: A working paper from Stevens Institute of Technology and UMass Boston studied 1,260,967 blind Chatbot Arena votes (May 2023 to August 2024) on AI models whose versions never changed. For the same pair of models, the share of votes rejecting both answers rose 1.71 points a month from an average of 17.1%. In fourteen pairs judged often in both years, it rose in every one, from 15.5% to 23.3%. In 2024, voters still picked the stronger model as often as before; they just accepted less.
Why it matters: As AI improves, users' bar rises with it, so a product that stands still gets judged as declining. Falling satisfaction scores are not proof that quality fell. Caveats: not yet peer reviewed, Arena voters are AI enthusiasts, and the data ends in August 2024.
How you can use it: Before deciding a tool or model "got worse", rerun a saved set of your own prompts and compare old and new answers blind. If you build AI products, freeze 30 to 50 prompts with today's outputs, re-score them blind monthly, and report satisfaction against that fixed baseline beside the raw score. A falling raw score with a flat baseline calls for a planned upgrade.
TOOL REVIEW
Latch: teach an AI bookkeeping
What it does: You do a routine bookkeeping task on screen and talk it through. Latch records your voice, screen and clicks, then builds a "sidekick" that repeats the task in tools like Xero, Google Sheets and Stripe, asking when something doesn't fit. Its privacy notice allows training AI on recordings, so ask about opting out first.
Price (as of 14 Nov 2026): Free for 50 credits a month; Solo is $50 a month ($40 billed annually) for 500 credits; extra credits cost $100 per 1,000. The pricing page doesn't say how many credits a run uses; check your balance before and after a test run.
Who it's for: Bookkeepers and small-business owners doing the monthly close.
Verdict: Worth a free-plan test if you or your bookkeeper repeat the same click-heavy close task every month in US tools like Xero, but hold off on paying until Latch says how many credits a run costs.
WHAT GOT FUNDED YESTERDAY
Morsa (morsa.ai) · manufacturing AI · India. Founded by Satvik Kapoor (CEO) and Vedant Singh (CTO), the Delhi company builds an "AI COO" for factories that assigns tasks, chases suppliers and re-plans schedules. It works inside the WhatsApp groups plant teams already use, and says on-time task completion at one glass maker rose from 30% to 75% in four weeks. It raised a $1.5M pre-seed round led by Jackson Square Ventures, with the a16z Scout Fund.
Pro Climbing League (proclimbing.com) · sports media · UK. Co-founded by Danaan Markey (CEO), Charlie Boscoe, Neil Wyatt and Eelco Bulten, it runs a head-to-head professional bouldering league with a 2027 season in Berlin, Paris and London. It says it sold 4,000 tickets on the first day of sales and is not owned by a gym, brand or agency. It raised a £3.2M seed round led by Redrice Ventures, with Active Partners, Accelerate Ventures and the International Bouldering Alliance.
Refilled (refilled.com.au) · office hardware · Australia. Founded by Ryan Nelson (CEO), the Sydney company makes the Refiller, an office machine that chills and filters tap water, serves it still or sparkling with flavours, and reorders its own supplies. It says it has kept 1.6 million single-use bottles out of landfill. It raised a $5.6M Series A led by OIF Ventures, with Melt Ventures and David Paradice.
Sagittarius Logistics (sagittariuslogistics.com) · space launch · US. Co-founded by Jonathan Slavik (CEO), Dave Masten (CTO), Jenna Edwards (COO) and Daniel Sinkel, the Colorado company is building fully reusable rockets that launch from inland sites. It says both stages can throttle back and land downrange or back at the pad, which supports overland flight away from the crowded coastal launch sites. It raised a $1.2M pre-seed round with Nova Threshold, Felicis Scout Fund, Ingenuity VC and Space Venture Studios; no lead was named.
Liflex Science (liflex.bio) · biotech · South Korea. Founded by Seoul National University professor Kim Jin-hong (co-CEO), with co-founder Kim Yi-jun of Ewha Womans University, the company develops anti-ageing RNA drugs, led by an osteoarthritis therapy it aims to take into clinical trials in the second half of 2027. It says it plans follow-on RNA drugs for obesity, kidney and heart disease, and is looking at genetic circuits as a second platform. It raised a KRW 7B (about $4.8M) Series A from seven investors including Korea Development Bank, Medytox Venture Investment and Murex Partners; no lead was named.
QUICK HITS
Nvidia's 8-K filing says Jensen Huang has resigned as CEO and from the board, effective immediately; CFO Colette Kress is interim CEO, and results still come on 17 November.
llama.cpp reached v1.0, so the open-source library for running AI models on your own computer now promises a stable API.
Adobe released an MCP server that lets ChatGPT, Claude and Gemini use features in Photoshop, Illustrator, Premiere and After Effects.
That's today. Hit reply and tell me what you're building. I read every email.
Dev

