AcceleratingBiz®

Growth CEO’s AI Briefing · Edition #102

Paid On Results

Tuesday, September 1, 2026 · 4 Developments · 2 Perspectives · 2 Case Studies · 2 Resources

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

Two items today move in the same direction, and neither is a model release. The largest AI vendor has started billing some customers only when the work completes, and one of the largest professional services firms has started paying bonuses for the judgment a model cannot supply.

  • Outcome pricing. OpenAI is billing select large accounts only when the AI finishes the job.
  • New ad surface. ChatGPT Ads opened self-serve buying globally at a $1 billion annualized run rate.
  • EU designation. ChatGPT is now regulated as a search engine, with obligations due by January.
  • Generated interfaces. Runway's Solaris draws the screen frame by frame with no code underneath.
  • Rationed access. Tomasz Tunguz argues labs are sorting buyers into camps with whitelists and defaults.
  • Portfolio churn. Every moved drafting back to Claude and kept its automated pipeline on Opus 5.
  • Self-improving product. Amplitude's agent found, shipped, and measured its own product changes.
  • Human skills. EY put $100 million behind judgment and collaboration, nominated by peers.
  • Design method. A former Apple design lead published a critic-subagent loop worth stealing.
  • Open model tuning. Fireworks opened its training API, making fine-tuning a procurement decision.

The pricing story is the one to read twice, because your enterprise buyers will read it too.

Developments

Pricing

OpenAI has started letting some large customers pay only when the AI finishes the job

The Information reported that OpenAI has quietly begun offering a small set of major accounts outcome-based pricing, billing when a task such as a resolved customer support request actually completes rather than per token consumed. Terms, customers, and prices are undisclosed, and OpenAI has not announced the arrangement. Reported per-resolution rates cluster near a dollar rather than a cent, because the vendor now absorbs the failure risk. Token billing has been difficult for buyers to forecast and harder for their finance teams to defend.

So what

When the largest model vendor prices on completed work, your enterprise buyers will ask you for the same terms within two quarters. Run the arithmetic now on which of your workflows have a clean, auditable definition of completion and a success rate comfortably above your gross margin floor. Price on outcomes only where both conditions hold, and keep seats or consumption for everything else. A defensible published success rate is a stronger differentiator in that negotiation than a lower seat price.

Confidence: Watch this. Reported by The Information from unnamed sources and not confirmed by OpenAI.

Go deeper The Information · The Next Web

Go-to-Market

September 1: ChatGPT Ads opened self-serve buying at a $1 billion run rate

OpenAI said on August 31 that ChatGPT Ads crossed a $1 billion annualized revenue run rate less than 200 days after launch, and opened self-serve buying across global markets today. A redesigned onboarding flow builds a campaign in under 60 seconds from a website URL, analyzing the business, proposing relevant ChatGPT user journeys, and generating creative and copy. New advertisers get $500 in matching credits. This puts a paid surface inside the assistant your buyers increasingly use to shortlist vendors.

So what

You have spent a year worrying about losing organic citations inside ChatGPT, and you can now buy placement instead. Run one contained test this month against a defined CAC ceiling, because auction pricing on new inventory is never cheaper than at the open. Measure it against your answer-engine referral cohort rather than your paid search cohort, since the intent and the path to close are different. Treat the 60-second campaign builder as a first draft that your marketer edits, not a strategy.

Hype check

The $1 billion figure is OpenAI's own annualization of under 200 days of selling. No advertiser retention, renewal, or return-on-spend data has been disclosed, and launch-period spending on a novel surface is not a durable run rate.

Confidence: Established fact that the announcement and self-serve launch occurred; the revenue figure is company-reported.

Go deeper OpenAI · Marketing Brew

Legal & Regulatory

August 31: The EU designated ChatGPT a Very Large Online Search Engine

The European Commission designated ChatGPT under the Digital Services Act on August 31, the first time an AI assistant has been regulated as a search engine. OpenAI declared 159.1 million monthly EU users against a 45 million threshold, and the Commission chose the search category because the service retrieves from the open web. OpenAI has four months, until January 2027, to meet systemic risk assessment, minor protection, and algorithmic transparency obligations. Reddit and Roblox were designated as Very Large Online Platforms the same day.

So what

The precedent that reaches you is categorical rather than about OpenAI. A conversational product that retrieves third-party web content is a search engine in EU law regardless of what your marketing calls it. If you serve EU users at any meaningful scale, have counsel test your monthly active user methodology against the 45 million threshold and inventory what a systemic risk assessment would require. That is a two-quarter program, so the time to scope it is before you cross the line, not after the designation letter arrives.

Confidence: Established fact. The designation is published by the European Commission.

Go deeper European Commission · Ars Technica

Product & Technology

August 31: Runway's Solaris generates the interface itself, with no code underneath

Runway introduced Solaris, which it calls an Interface World Model, rendering websites and apps as real-time video and drawing each frame in response to clicks and drags rather than emitting HTML and components. In Runway's own study, 250 evaluators preferred Solaris over pages coded by Claude Opus 5 in 71% of in-scene behavior comparisons and 61% on instruction following. Runway acknowledged text legibility problems, drift over long sessions, and convincing but wrong screens. Access is request-only early research with no API, pricing, or launch date.

So what

This is a research preview and not a build decision, but it prices a direction. If a model can generate the next screen state, prebuilding every state becomes optional, which changes what a design system is for and where your frontend cost sits. Do not restructure a roadmap around a request-only demo. Do put one question to your product leaders this quarter: what would our product look like if the interface were generated per user rather than shipped to everyone, and which parts must stay deterministic regardless.

Hype check

The 71% and 61% preference figures come from a study Runway ran on its own model, with no independent replication and no published methodology for evaluator selection or task mix.

Confidence: Watch this. An early access research preview evaluated by its own vendor.

Go deeper Runway · The New Stack

Perspectives

AI Strategy & Transformation

Tomasz Tunguz: frontier access is being rationed, and you are being sorted into a camp

Tunguz argues the frontier AI market is closing from both ends of its supply chain. Labs are building whitelists and blacklists, naming exclusive partners, and rationing their strongest models, while governments add export restrictions, nationality screening, and tiered access. Downstream, enterprises are standardizing on one or two named vendors and products increasingly ship with a default model inside. His lead example is Salesforce naming Anthropic, which makes Claude the default inside the largest CRM and Slack.

So what

If your product depends on a single frontier model, your access is a commercial relationship someone else can reprice or revoke, and OpenAI cutting off Cursor last week was the live demonstration. Keep a tested fallback behind an abstraction layer and know your switching cost in engineer-weeks before you need the number. Then ask a second question most teams skip: which camp have your largest customers already joined, because their vendor standard increasingly dictates which model you are permitted to run on their data.

Confidence: Informed perspective. A venture investor's synthesis of a pattern, not a measured market study.

Go deeper Tomasz Tunguz · TLDR AI

Product & Technology

Every moved drafting back to Claude, and the useful part is what it kept elsewhere

Every published its hands-on review of Anthropic's Fable 5.1 on September 1, calling it faster, clearer, and less argumentative than the original Fable, and said the team is moving work back to Claude after a summer that pushed much of it to ChatGPT and Codex. The reservations are specific: the model still overruns briefs, can invent quotations, and sometimes keeps working after being told to stop. Every's automated editing pipeline is staying on Opus 5 regardless of the upgrade.

So what

The signal is not which model won. It is that a serious team re-evaluates its model portfolio every few weeks and deliberately runs different models for human drafting and for automated pipelines. Budget that churn as a standing operating cost with a named owner and a review cadence, rather than treating model selection as a one-time architecture decision you revisit at renewal. Note also that the failure modes Every names, invented quotations and ignored stop instructions, are precisely the ones a benchmark score will never surface.

Confidence: Informed perspective. Hands-on testing by one team on its own workloads, not a controlled evaluation.

Go deeper Every Vibe Check · Every

Case Studies

Product & Technology

Amplitude let an agent choose and ship its own product changes, then measure them

Amplitude ran its Wave agent on Amplitude's own product for several months, letting it find problems in behavioral data, ship fixes, and A/B test whether they worked. In one case Wave found that users ignored suggestion cards, personalized them to each user's recent work, and had a working pull request open in about 15 minutes, producing a 160% lift in suggestions accepted. The agent drafted the spec and routed it to a Cursor cloud agent that opened a PR across four files, which a lead engineer reviewed, approved, and merged. Amplitude reports key metrics moving 50% to 3x.

So what

The transferable asset is the loop, not the tool. An agent that proposes a change is a demo; an agent wired into the measurement system that tells it whether the change worked is a compounding capability, and most companies already own all three pieces without connecting them. Ask where your behavioral data, your spec process, and your deploy pipeline are close enough to close that loop on a low-risk surface such as onboarding copy or empty states. Keep the human merge gate that Amplitude kept.

Hype check

A vendor reporting its own agent's results on its own product, selecting the examples, with no disclosure of how many proposed changes failed their tests or were rejected at review.

Confidence: Informed perspective. Vendor-reported and self-referential, though the mechanism is documented in detail.

Go deeper Amplitude · The Neuron

People & Organization

August 31: EY put $100 million behind judgment, collaboration, and adaptability

EY's US business committed $100 million this fiscal year to bonuses for employees who demonstrate human skills such as leadership, judgment, business acumen, collaboration, and adaptability, alongside technology adoption. Spot awards top out at $500, while individuals and teams whose contributions move the firm are eligible for cash prizes of $10,000 to $25,000, a ceiling five times higher than the prior program. Employees can nominate colleagues at any level, and there is no cap on what one person can receive. It reverses the firm's messaging of a year ago.

So what

A firm whose product is billable human hours is paying to keep the human contribution legible while it automates everything around it, and it is paying peers rather than managers to identify that contribution. If your compensation plan still rewards output volume, AI has already broken your measurement, because volume is now nearly free. Add a peer-nominated award for judgment calls and cross-team work, sized large enough that people notice, before your strongest performers conclude the only rewarded skill left is prompting.

Confidence: Established fact. Announced by EY and reported independently by CBS News and Fortune.

Go deeper EY · CBS News

Resources

Product & Technology

Seven techniques for getting design out of a model built to be average

Anshu Chimala, who led design and engineering research at Apple for 12 years, published a working process for AI design in Lenny's Newsletter on September 1. His diagnosis is structural rather than about prompting: a next-token predictor makes the most agreeable choice at every step, which is design by committee by construction. The techniques are concrete, including seeding a random string as creative input, running a large critic model against screenshots in a fresh context while a cheap model does the building, and spending most of the polish removing what the model added.

So what

The critic-subagent pattern generalizes well past design and is the cheapest quality upgrade available to most teams. An expensive model judges only the finished output in a clean context, a cheap model does the production work, and an objective stopping rule keeps the loop from burning tokens forever. Chimala reports the critic accounted for under 10% of output tokens while roughly halving the cost of the alternative. Try it this week on one artifact your team produces repeatedly and rates inconsistently, such as proposals or release notes.

Confidence: Informed perspective. A practitioner method with worked examples rather than measured outcomes.

Go deeper Lenny's Newsletter · Sakana AI on seed strings

AI Strategy & Transformation

Fireworks opened its training API so teams can tune an open model on their own work

Fireworks made its Training API and Lab generally available on August 31, letting teams train open models for a specific job and deploy them from the same platform without standing up their own infrastructure. It lands in the middle of a run of evidence that buyers are moving routine work to cheaper open models. AT&T now routes roughly 40% of employee AI usage to open models it hosts itself, and open models took 62% of Vercel's tokens.

So what

Fine-tuning an open model is becoming a procurement decision rather than a machine learning project, which changes the build-versus-buy math for any workflow you run at volume. The candidate is narrow, repetitive work where a frontier model is overqualified and you already hold labeled examples, such as ticket classification or document extraction. Pick one, measure quality against your current model on your own tasks rather than a public benchmark, and keep the frontier model wired in as the fallback path.

Confidence: Established fact. General availability was announced by Fireworks on its engineering blog.

Go deeper Fireworks · The Neuron

Sources in this edition: The Neuron / Mindstream, Lenny's Newsletter, Every (AI & I / How I AI), The Information, The Rundown AI, The Deep View, Superhuman AI, TLDR AI, TLDR Marketing

← Latest Editions and Search