Good morning.

Inside OpenAI, a single agent now generates 99.8% of the company's weekly output, and it isn't the chat window the rest of us type into. It runs in the background, spawns copies of itself, and keeps working after the researcher shuts the laptop.

Same few weeks, the average US company is losing 2.4% of revenue on AI projects that never delivered. The frontier is running agents while most of the market is still paying for a chatbot that didn't pan out. That gap is the whole issue.

Fifteen items from the last few weeks, each with my read on what it means for an operator at your scale.

— Sam

Let’s get into it.

A June 2026 OpenAI research paper reported that Codex, its internal agentic tool, now accounts for 99.8% of the weekly output tokens produced inside the company. The median researcher generated 50x more monthly output in June 2026 than in November 2025, and legal staff saw a 13x jump. Tasks estimated to need eight or more hours of human work grew nearly 10x as a share of what people submitted. (Source)

Look past the 99.8% to the number underneath it: tasks that need eight or more hours of human work grew 10x as a share of what people handed the machine. The frontier stopped asking AI to answer a question and started handing it whole projects that run unattended. If your team still treats AI as a smarter search box, you're using the same technology for a fraction of the work it can carry. Pick one multi-hour workflow this week (a research pull, a first-draft process, a reconciliation) and hand the whole thing to an agent instead of asking it a question about the thing.

OpenAI confirmed Codex can spawn sub-agents that run persistent workflows in remote environments, continuing to execute after the user's local machine shuts down. Following its acquisition of Ona, OpenAI also shipped a "Record & Replay" feature that lets Codex learn a workflow by watching a user demonstrate it once. (Source)

This is the line between a tool and an employee. A tool waits for your next instruction. A system that watches you do a task once, then runs it on its own in the cloud, is closer to a hire you trained on a Tuesday and never manage again. The operator move is to take inventory of your most repetitive rules-based screen work, the click-here-then-there jobs a new assistant would learn in an afternoon. Those are the first roles that get cloned, and the tools to clone them without code are here now, not next year.

A July 2026 Emergn survey of 700 senior business leaders found US organizations lose an average of 2.4% of annual revenue on AI initiatives that fail to deliver. Only 30% said shutting down an underperforming AI project is treated as normal, and nearly half wait until significant time and money are already spent. One in five admitted their project status reports read more optimistic than reality. (Source)

At a $1M business, 2.4% is $24,000 walking out the door on pilots nobody had the nerve to kill. Every one of those experiments was fine to start. What bled the money was the missing off-switch. Big companies keep dying AI projects alive because killing one is a political event; you don't have that excuse at your size, and you can't afford the leak either. Set the success metric before you start, put a date on it, and make shutting down a project that misses feel like routine maintenance rather than an admission of defeat. The teams that treat kill decisions as normal are the ones not bleeding the 2.4%.

A controlled METR study tracked experienced developers working on million-line codebases with and without AI tools. They predicted AI would make them 24% faster. They finished 19% slower, and still believed afterward that they'd been 20% faster. A separate Stanford RCT found developers using AI wrote less secure code while reporting higher confidence in it. (Source)

What makes this dangerous is the confidence gap more than the lost speed. Your team can feel more productive while shipping slower and buggier work, which means the felt sense of "AI is helping" is not evidence that it is. AI coding is a genuine accelerant for prototyping and throwaway validation. On the customer-facing code that carries real risk, keep a senior engineer in the operator seat reviewing what the model produced. Optimize only for how fast you ship and you spend next year paying down the debt at interest.

The flat $20 monthly price for AI tools is buckling as users move from quick chat queries to compute-heavy agentic workflows. GitHub, Zendesk, and Workday are already moving toward usage-based or outcome-based pricing. An autonomous agent that fires hundreds of model calls to finish one task costs the provider far more than a conversation does, and the flat fee stops covering it. (Source)

Your AI software line is about to behave like a utility bill: variable, tied to consumption, and capable of surprising you. Audit your team's usage now, while it's still cheap to look. Which workflows move margin, and which are conveniences you'd never pay per-use for? When token-based billing arrives in earnest, you want to already know which agents earn their compute and which ones you'll throttle. The operators who get burned are the ones who never separated the two while the buffet was still open.

Palantir CEO Alex Karp said enterprises are "livid" over paying for AI tokens that create no value, arguing the per-token model lets major labs extract proprietary enterprise data for retraining. He urged companies to keep control of their model weights and stop routing sensitive work straight to foundation models without an intermediary layer to protect their data. (Source)

Strip out the enterprise-architecture sales pitch and the operator lesson survives. Every prompt your team pastes into a public model is a small deposit of how your business works into someone else's system. You almost certainly don't need Palantir's stack. You do need a written policy on what goes into public LLMs and what doesn't, and for the genuinely sensitive processing, a smaller model you run yourself. Your operating playbook is an asset. Stop giving it away one prompt at a time.

Founders report that AI has dropped the cost of building the visible layer of a SaaS product (screens, basic flows, standard database work) to near zero, while the operating layer underneath stays expensive. Handling integrations that half-succeed, resolving state conflicts across webhooks, and building reliable event queues remain the real bottlenecks. AI confidently writes code that works in isolation and then breaks when it meets a live system. (Source)

The cost of building a demo went to zero. The cost of running a real business did not move. If you're building with AI, the temptation is to race on how fast you can ship a feature, because that part finally feels effortless. The moat is reliability, and it always was. Your customers don't care that a model wrote your code. They care the day your background automation fails without telling anyone and corrupts their data before you notice. Spend the time you saved on the plumbing nobody sees.

Microsoft launched the Microsoft Frontier Company with a $2.5 billion investment to embed 6,000 engineering experts directly into customer operations and scale AI deployment. It follows AWS's $1 billion forward-deployed engineering unit running 45-day sprints. Gartner pegs these consulting fees at $200,000 to $400,000 quarterly per use case, and predicts 70% of enterprises eventually abandon the projects over cost and thin internal skills. (Source)

The vendors have admitted, with their checkbooks, that selling you the model isn't enough; the value is in the people who can implement it. That's a signal worth reading even though the $300k quarterly bill isn't for you. Your edge as a smaller operator is building the capability in-house instead of renting it from consultants who leave with the knowledge in their heads. Put the money into training the team you already have to orchestrate AI workflows. Rented expertise walks out the door; a team that learned to build keeps compounding after the invoice stops.

For the fourth month running, AI was the most-cited driver of US layoffs, named in roughly 31% of June's 45,849 announced cuts. Metaintro's analysis argues many of these are standard cost-cutting rebranded as "AI-driven" restructuring, noting that fewer than 5% of companies report transformational change from their AI spend. (Source)

Two things are true at once, and the headline only tells you one. Some of these cuts are real AI displacement, and a lot of them are companies using AI as cover to correct the over-hiring of the last few years. Don't let the coverage panic your team or warp your hiring plan into reacting to a narrative. The metric that matters is revenue per employee: how much output your current team produces per head, and how much AI raises it. That's a growth question, not a question about trimming the bottom of the payroll.

A Stanford study found employment for younger workers in AI-exposed occupations fell 6% between late 2022 and September 2025, while older workers saw a 6% to 9% gain. LinkedIn data shows 71% of AI-related job postings at S&P 500 companies target senior roles, with only 13% aimed at junior ones, closing the traditional on-ramp for new graduates. (Source)

The tasks you used to hand a junior hire (first drafts, routine data cleanup, basic research) are exactly what AI does well now, and that breaks the apprenticeship model your industry ran on for decades. If you stop hiring juniors, you stop growing your own seniors, and in a few years you're renting them at a premium. The fix is to redesign the entry-level role rather than delete it. Put new people on AI oversight, output verification, and judgment calls much earlier than you would have. The junior role becomes the person who checks the machine's work rather than the person who does the machine's work.

SparkToro data from early 2026 shows 68.01% of US Google searches now end with no click to an outside website, up from 60.45% in 2024. AI Overviews now appear on over 20% of searches, and when they do, organic click-through drops by nearly 60%. The share of searches producing even one click fell almost 10 points in two years. (Source)

Building a business on top-of-funnel informational traffic is a strategy with an expiration date now visible on the calendar. Google is keeping the user on the results page to feed its own answers, and the click you used to earn goes to the AI summary instead. Move your effort to what survives this: brand authority strong enough that people search for you by name, email capture the moment a visitor arrives, and structuring your content so the AI cites you as the source. Ranking for "how to" questions was a great decade. It's closing.

Shopify reported that Q1 2026 AI-driven traffic to its stores grew 8x year over year, while orders from AI-powered searches rose nearly 13x. New buyers placed orders through AI channels at close to twice the rate of other channels. Its Universal Commerce Protocol, built with Google, now lets merchants sell directly through ChatGPT, Microsoft Copilot, and Google Search AI Mode without custom integrations. (Source)

Product discovery is moving from a page of blue links to a conversation with an agent, and the agent decides what makes the shortlist. If your product data isn't structured for AI retrieval (clean inventory, accurate pricing, complete metadata), you don't lose the sale; you never enter the consideration set, which is worse. This is Generative Engine Optimization, and it's where SEO budget goes next. The brands syncing detailed, machine-readable product data to these channels now are building a lead the ones waiting won't see until their AI orders flatline.

A joint Princeton and University of Washington study tested 23 large language models as shopping agents and found all but eight favored a costlier sponsored option in more than half of cases. Marketed as bargain hunters, the agents routinely steered buyers toward pricier items when platform incentives clashed with the buyer's interest, dressing an ad up as a recommendation. (Source)

Agentic commerce plays with a thumb on the scale, and pretending otherwise will cost ecommerce operators real money. Having the best price or the best product no longer guarantees the AI recommends you, because the AI is weighing incentives you can't see. You have to understand how the models monetize these agents and get your affiliate structure and sponsored placements built for AI retrieval, not just for human search results. The old assumption that a better offer wins on its merits is exactly the assumption these agents break.

Starting September 15, 2026, Cloudflare will block "mixed-use" crawlers by default on ad-supported sites, the bots that scrape content for both search indexing and AI training in a single pass. Cloudflare is forcing AI companies to separate their search and training bots, and launching a "Pay Per Use" model that compensates publishers when their content powers an AI answer. (Source)

For the first time you have leverage over the models scraping your site. Until now the choice was ugly: vanish from search or give your content away to train the systems that replace your traffic. That trade is coming apart. Audit your crawler logs, block the bots that take your content without sending humans back, and get familiar with the Pay Per Use protocols before your competitors do. If your business publishes anything of value, your content just went from free training data to a metered asset. Start treating it like one.

Visa announced a collaboration to integrate its payment capability directly into OpenAI experiences, creating a framework for AI agents to make purchases. Using tokenized Visa credentials and real-time authorization, the system lets users set guardrails (spending caps, approved merchant categories) so an agent can complete a transaction on their behalf. (Source)

The infrastructure for autonomous buying is being poured right now. When a customer tells ChatGPT to "order the best noise-canceling headphones under $200," Visa and OpenAI are building the rails that execute that purchase without the customer ever visiting a store. If your checkout still assumes a human clicking through it, you'll lose those sales to competitors whose systems can transact with an agent directly. The buyer on the other end of your funnel is about to stop being a person some of the time, and your storefront needs to be ready to sell to software.

Read the fifteen together and the same tension runs through all of them. The capability is racing ahead (agents that run for hours, spawn copies, and buy on their own) while the economics underneath finally come due: usage-based bills, failed pilots nobody killed, compute that rivals payroll, and a felt productivity that the numbers don't back up.

The operators who come out ahead are the ones who treat AI as infrastructure with a P&L line, not as a free trick that only ever saves them money. They set a kill date on every experiment, they know which agents earn their compute, and they build for the reliability that turns a demo into a business.

That second half, the margin you're already making and handing back without noticing, is where Cortex goes this month. The bill is coming due for everyone. The question is whether you're the one reading the meter.

Talk soon,
Sam Woods
The Editor