AcceleratingBiz®

Growth CEO’s AI Briefing · Edition #149

Own the Judgment

Monday, November 9, 2026 · 3 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.

This Monday edition covers the 72 hours since Friday’s issue. Today’s thread: AI keeps getting more capable and cheaper, and what separates good results from expensive ones is human judgment about what “done” means.

  • Frontier gets cheaper. Anthropic released Claude Fable 5.2 and cut its price by 20%.
  • Agents overbuy. In Visa’s new benchmark, most failed AI shopping runs still ended in a purchase.
  • Money gets choosier. OpenAI raised less than it sought as Microsoft pulled back on some capacity deals.
  • Almost, not quite. Ethan Mollick finds Fable 5.2 expert-level but blind to its own weak work.
  • Keep the map. Notion’s Ivan Zhao warns teams not to lose track of what their agents build.
  • The model sets headcount. Growth without hiring works where one more customer does not need one more person.
  • Agents with job titles. A 25-person company runs on named agents with clear roles and limits.
  • Price the unit. Seven ways to handle token costs, and why finance should sign off first.
  • Define done first. Coding agents met most requirements but finished only about a quarter of tasks.

This week, put a named person in charge of judging important AI output, and write down what “done” looks like before the work starts.

Developments

Technology

Anthropic Releases Claude Fable 5.2 and Cuts Its Price 20%, While Still Recommending Opus 5.5 for Most Work (November 9, 2026)

Anthropic released Claude Fable 5.2, its most capable generally available model, along with Mythos 5.2, a version with looser cybersecurity and biology safeguards that only vetted organizations can use. Fable’s price falls 20% to $8 per million input tokens and $40 per million output tokens, with an optional fast mode at double that, and it is available on Anthropic’s platform, AWS, Google Cloud and Azure. Anthropic says answers are up to 30% faster than version 5.1 and that refusals of harmless requests fell about 55%. It still recommends Opus 5.5 for most coding and knowledge tasks.

So what

Price cuts at the top of the range change which jobs are worth sending to the strongest model, so ask your engineering lead to rerun cost estimates for your heaviest AI workflows this month. Keep cheaper models as the default and reserve Fable for long, hard, multi-step work where the extra capability pays for itself. If you resell model access inside your product, revisit your margin assumptions before customers ask why your price did not move.

Hype check

The speed, benchmark and refusal figures are Anthropic’s own, the launch lands weeks before an expected IPO, and Ethan Mollick’s early review (below) was based on pre-release access.

Confidence: Established fact — launch, pricing and availability are confirmed by Anthropic and independent coverage; performance claims are self-reported.

Go deeper Anthropic · The Next Web · One Useful Thing

Customer BehaviorDeveloping

Visa’s First AI Shopping Benchmark: Agents Failed 42.8% of Runs, and Most Failures Still Ended in a Purchase (November 5, 2026)

Visa tested 22 AI models from 12 labs on 15 realistic shopping tasks, such as buying a TV or a laptop, and found an overall pass rate of 57.2%. The best models passed 82.2% of attempts and the worst 28.9%, and the five most expensive models beat the five cheapest by 67.1% to 50.2%. About 54% of failed runs still ended with the agent buying something, and all three attempts at the same scenario passed only 34.2% of the time. Visa’s September survey found 72% of U.S. consumers have used an AI assistant, but only 23% trust one to pay on their behalf.

So what

The costly failure is not a missed sale but a confident wrong purchase, which turns into a return, a chargeback or a support ticket with your name on it. If you sell online, test how agents describe and buy your products, and make sure return and cancellation flows can handle orders the customer did not really intend. If you build agent features, add a clear “don’t buy yet” path and require human confirmation above a set amount. Expect the payment networks to write the rules here.

Confidence: Established fact — figures are from Visa’s release; Visa did not name the models it tested and has a commercial interest in setting standards for agent payments.

Go deeper Visa · The Deep View

BusinessDeveloping

The AI Buildout Gets Pricier: OpenAI Raises $11 Billion at $1.25 Trillion, Short of Its $30 Billion Goal, as Microsoft Pulls Back on Neocloud Deals (November 5–8, 2026)

OpenAI has raised about $11 billion of a round that could reach $20 billion at a $1.25 trillion valuation, below the $30 billion at $1.4 trillion it reportedly sought in September, according to Bloomberg. The same day, Bloomberg reported that prospective buyers of Anthropic’s IPO shares may value it at as little as $1.1 trillion, against roughly $2 trillion expected a month earlier. Bloomberg also reported that Microsoft is pulling back from parts of up to $70.9 billion in capacity deals with Nscale, Nebius and IREN over delays, cost overruns and permitting. The FT reported, citing one unnamed person, that Anthropic agreed to pay about $3.4 billion to take over 1.7 gigawatts of gas generator orders and skip multi-year waits.

So what

Capital is still flowing, but on tougher terms, and new computing capacity is getting slower and costlier to bring online. For growth companies, expect AI vendors to push harder on price, commitments and usage limits over the next year, as the recent caps at Apple and on Grok Bot already showed. Keep AI contracts short where you can, build so you can switch models, and ask your investors how they are marking AI-exposed positions after last week’s drop.

Confidence: Informed perspective — funding and Microsoft figures come from Bloomberg as reported by TechCrunch and AI Weekly; the Anthropic generator deal rests on a single unnamed source and is unconfirmed.

Go deeper TechCrunch · AI Weekly (Bloomberg on Microsoft) · The Deep View

Perspectives

AI Strategy & Transformation

Ethan Mollick on Fable 5.2: Expert-Level Work Across Fields, but It Never Noticed When Its Own Work Wasn’t Good Enough (November 9, 2026)

Mollick, who tested Fable 5.2 before launch, calls it “almost” AGI: as good as most experts on many computer-based tasks, but still uneven. It screened 22 business ideas and wrote a plan he says would impress a team of his MBA students, and produced an entrepreneurship paper he rates ahead of second-year PhD work. It also built a course that was accurate but joyless, a 182-page paper that never cut anything, and, on a 3D world-building project, an 87,500-word internal rule book that slowed output from hundreds of worlds a day to about a dozen. Its critic and fact-checking agents caught errors but never asked whether the work was too long, repetitive or dull.

So what

The scarce skill is no longer producing the work but judging it, and that judgment stays with your people. When you put agents on long projects, name an owner who reviews the output for fit and quality, not just accuracy, at set checkpoints. Watch for agent systems that pile up rules and approvals, and periodically have an outside reviewer simplify them. Mollick stopped short of letting Fable deal with real customers, a sensible line to hold for now.

Confidence: Informed perspective — a careful practitioner’s hands-on tests, not a controlled study.

Go deeper One Useful Thing · Anthropic

People & OrganizationNew Angle

Notion’s Ivan Zhao: “I Don’t Think We Should Surrender Our Understanding” as Agents Write the Code (November 5, 2026)

At Every’s Thesis: 2027 conference in Brooklyn, Notion cofounder and CEO Ivan Zhao warned that when every engineer works through their own coding agent, a medium-size codebase can stop making sense to the people responsible for it within a week. Notion now routes its agents’ work through a shared kanban board so people can see what the agents are doing. Every CEO Dan Shipper closed the day arguing that once everyone has the same model, its default answer becomes cheap and the perspective a person brings on top of it is what matters. Zhao said last week that agents now create about half of Notion’s pages.

So what

Speed from agents is easy to measure; lost understanding is not, until something breaks and no one can explain why. Make agent work visible in one place your team already uses, such as a board or ticket queue, rather than leaving it in individual sessions. Ask your engineering and operations leads whether someone can still explain how each critical system works, and hold regular reviews where people walk through what agents changed.

Confidence: Informed perspective — conference remarks as reported by Every; Notion’s practice is self-described.

Go deeper Every: Thesis 2027 recap · Every: Thesis 2027

People & Organization

Growth Without Hiring Depends on Your Business Model: Revenue and Headcount Split at Big Software Firms but Not at Banks or Consultancies (November 2026)

An FT analysis by John Burn-Murdoch, summarized by The Signal, finds that revenue at Alphabet, Meta, Microsoft and Spotify kept rising after 2022 while staff stayed flat or fell, while at Bank of America, Visa, Accenture and Deloitte revenue and headcount rose together. He attributes most recent U.S. job weakness to rate rises from 2022, though AI may now be keeping hiring soft. CNBC profiled owners with no employees who use AI for bookkeeping, marketing and support, including one expecting $2.5 million to $3 million in revenue across 12 brands. Meanwhile, U.S. employers unexpectedly added 70,000 professional and business services jobs in October.

So what

Whether AI lets you grow without adding people depends on whether one more customer needs one more person. Split your cost base into work that scales with customers and work that does not, and set 2027 hiring plans by function rather than one company-wide ratio. Be skeptical of plans that promise flat headcount across a services-heavy business, and track revenue per employee by team, not just overall.

Confidence: Informed perspective — company-level patterns, not causal proof; the FT article was not reviewed directly.

Go deeper John Burn-Murdoch thread · The Signal · CNBC via TLDR Founders

Case Studies

Operations & Finance

A 25-Person Company Run With Named Agents: Hackmamba’s Founder Built Native Apps Without Knowing Swift and Retired Most of His Skill Library (November 9, 2026)

William Imoh, founder of the 25-person firm Hackmamba, used OpenAI’s Codex to build native iOS and Android versions of Mural, an AI language tutor, despite never having written Swift or Kotlin. He runs the business with Grok Bot agents given Lord of the Rings identities, from a chief of staff to marketing, sales and engineering, connected to Google Workspace, Linear, Notion, Pipedrive and QuickBooks, with banking and customer data kept off-limits. He also built his own HR tool that combines email, calendar and meeting notes to rank candidates. As models improved, most of his custom skills became unnecessary; the few that survived encode lasting expertise such as growth strategy.

So what

Small teams can now operate with the reach of much larger ones, but the leverage comes from access and clear roles, not clever prompts. If you are piloting agents, give each one a defined job and a named human owner, connect it to the systems it needs, and draw explicit lines around money and customer data. Review your prompt and skill libraries each quarter and retire what newer models no longer need.

Confidence: Informed perspective — one founder’s self-reported setup from a podcast interview.

Go deeper How I AI episode · ChatPRD write-up

Resources

Reference

Before Your Next AI Pricing Change: Seven Ways to Handle Token Costs, and Why the Same Deal Can Be $10 Million or $3.3 Million of Revenue (November 6–7, 2026)

In a guest post on OnlyCFO, investor Gokul Rajaram lists seven ways to handle token costs: absorb them, cut them, charge by usage, cap usage, let customers bring their own model key, move work to the device, or wait for prices to fall. Vanta’s CEO says the company is absorbing token costs and accepting lower margins because only about 25% of customers use its agent, and reaching 75% would improve retention. Rajaram’s advice is to price on a unit your product controls, such as a personalized email from an AI sales agent, and to know what each unit costs you. A Deep View weekend edition adds the accounting side: whether resold model access counts as gross or net revenue depends on who controls the price and the customer, and pay-per-outcome revenue cannot be booked at signing.

So what

Pricing and finance are now the same decision. Before launching credits, overages or outcome pricing, pull your last thousand agent runs, calculate the cost of each sellable unit, and have your controller confirm how the revenue will be recognized. A model that shows three times the revenue on a board slide may be net revenue to your auditor.

Hype check

The Deep View edition was sponsored by Hyphenate, which sells revenue-recognition software; the accounting principles it describes are standard, but the pitch is not independent.

Confidence: Informed perspective — practitioner frameworks; the accounting examples are illustrative, not advice on any specific contract.

Go deeper OnlyCFO · The Deep View

Tools

“Mostly Done” Is Not Done: Coding Agents Met 69.9% of Hidden Requirements but Finished Only 26.4% of Tasks, and a Two-Chat Fix (October 26, 2026)

The Co-Design Bench study tested ten coding-agent models on 273 real app-editing tasks drawn from more than 26,000 user sessions, graded against 2,331 expert-written acceptance criteria the models never saw. The best model met 69.9% of the criteria on an average task but fully completed only 26.4% of tasks, because one missed requirement fails the whole job. Graders checked each criterion against evidence, then a separate verifier challenged every pass. The Neuron turns that into a workflow: have the AI write a pass-or-fail checklist first, then open a fresh chat whose only job is to prove the work wrong.

So what

Most AI output that disappoints executives is not wrong; it is incomplete in ways no one wrote down. Require a short checklist of acceptance criteria before any significant AI-produced deliverable, and have a second reviewer, a person or a separate AI session, try to break it before it ships. It takes minutes and catches the quiet misses that erode trust in AI work.

Confidence: Established fact — benchmark figures are from the arXiv paper; the two-chat workflow is The Neuron’s adaptation and was not tested in the study.

Go deeper arXiv paper · The Neuron

Sources in this edition: One Useful Thing, The Neuron / Mindstream, The Deep View, The Rundown AI, TLDR AI, TLDR Founders, TheSequence, The Signal, Every (AI & I / How I AI), How I AI (Lenny's Podcast Network)

← Latest Editions and Search