What Would You Say… You Do Here?

THE NUMBER: 3.5x — what 95th-percentile firms now consume in AI per worker compared to typical firms, per OpenAI’s first enterprise B2B Signals report released yesterday. That ratio was 2x a year ago. Twelve months from now it’s 5x. “What… would you say… you DO here?” That’s Bob Slydell in Office Space, sitting across a folding conference table from Peter Gibbons, asking the question every consulting firm gets paid to ask and most CEOs are too polite to. It’s also the question OpenAI just answered with a chart. In 2026 the answer “I use Claude” is not the right answer. The right answer is “I’m one of the engineers growing the gap.” The companies running with that answer are pulling 3.5x ahead of the companies still measuring AI by license utilization. Bob and Bob are walking the floor again. They have data this time.

Yesterday OpenAI released its first B2B Signals report. The lead number is the most important data point a CEO has been handed in 2026 and almost nobody is naming it. Firms in the 95th percentile of AI use now consume 3.5 times as much intelligence per worker as the typical firm. A year ago that gap was 2x. Message volume explains only 36 percent of it. The rest is depth. Longer prompts. Richer context. Multi-step delegated workflows. Frontier firms are sending 16x as many Codex messages per worker. They are not buying more seats. They are using each seat to do more.

This is the structural bifurcation we have been writing toward for six weeks. Heat (Apr 30) said the smart money is hedged and the retail money is paying full sticker. Porsche In The Driveway (May 3) said you bought the Porsche and you are driving it like a bicycle. I Drink Your Milkshake (May 4) said the labs are walking into your operations through your PE sponsor’s checkbook. Name Of The Game (May 5) said the trillion-dollar aftermarket is three layers of buyers running a brokerage script. This morning’s No One Sets Off My Evil Detector said the harness offensive at Code with Claude SF was the lab’s answer to an 80x Q1 growth admission Dario Amodei put on stage Wednesday. All of it points to the same place. The AI-native company is not a company that uses AI. It is a company that has reorganized around AI, with different role definitions, different headcount math, different review process, and the productivity gap that re-org produces is now measurable at 3.5x and widening.

What’s new today is the bookend. Same news cycle: DeepL fires 25% of its workforce citing AI. Cloudflare (NYSE: NET) cut 1,100. AI is now responsible for 26 percent of April’s job cuts per Challenger, Gray & Christmas. Daniela Amodei stands on the Code with Claude stage and declares Claude’s transition from “chatbot” to “colleague” complete. MIT just published the Acemoglu paper in the May print issue of the Quarterly Journal of Economics, arguing that automation 1980-2016 was used to control wage premium, not boost productivity, and 60 to 90 percent of productivity gains were burned offsetting that targeting. The bifurcation is not a contradiction. It is the pattern. The productive companies are getting 3.5x. The non-productive companies are firing 25%. The wage-premium worker is exactly who AI replaces first. Acemoglu’s paper is not a forecast. It is the chapter that comes before the one we are now living in.

Bob and Bob are sitting at the conference table. They have new questions.

🧠 The 3.5x Gap Is The Org Chart, Not The License Bill

OpenAI’s B2B Signals report yesterday cracked open the operator question every CFO has been ducking. The 95th-percentile firm uses 3.5x as much AI per worker as the typical firm. That gap was 2x a year ago. The research underneath the headline is what matters. Message volume, the seat-count metric every IT department reports up the chain, accounts for only 36 percent of the difference. The other 64 percent is depth. Longer prompts, richer context, multi-step delegated workflows, more substantive outputs per interaction.

The depth metric is the one that does not show up on a license-utilization dashboard. It shows up in what your engineers ship.

Frontier firms send 16x as many Codex messages per worker. OpenAI named Cisco (NASDAQ: CSCO) as the case in the report. Codex helped Cisco cut build times by roughly 20 percent, save 1,500 engineering hours per month, and increase defect-resolution throughput 10 to 15 times. The Cisco team described the shift as treating Codex “as part of the team” rather than as a tool. Read those last six words slowly. Part of the team. Not a productivity hack. Not a vendor relationship. A team member. That is the linguistic register of an org chart that has been redrawn, and the productivity numbers that follow are the consequence.

Same Wednesday, Anthropic‘s Daniela Amodei stood on the Code with Claude stage in San Francisco and declared Claude’s transition from “chatbot” to “colleague” complete. Same story, different lab. Six months ago you typed into a prompt window. Today the model is in your Slack. In your Excel. In your inbox triage. In the agent fleet that runs your customer support overnight. The platform companies (Microsoft Agent 365, Google Agent Mode leaked Wednesday, Apple iOS 27 multi-model selector, Meta Hatch, OpenAI ChatGPT Apps, ServiceNow + Accenture’s same-day FDE program) are all sprinting to wrap that colleague in a directory service. Six platforms competing for the harness slot. Two labs competing for the model slot. One operator stuck deciding whether to let the colleague into the building.

The 3.5x is what happens when you let it in and reorganize around it. The license bill is downstream of the org chart, not the other way around.

The action item: Audit your engineering org by role, not by headcount. How many engineers in your stack are operating in the depth percentile? Long prompts, multi-step workflows, agent supervision, code reviews of compound-engineering output. How many are still using AI as autocomplete on the same flow they ran in 2023? The first group is the only group whose comp can scale with output. The second group is the wage-premium group from the Acemoglu paper, ten months from now.

🦞 Bob And Bob Are Asking The Wage Premium Question

The same day OpenAI publishes the depth-gap data, DeepL announces it is cutting 25 percent of its workforce, roughly 250 jobs, citing AI. Read that twice. An AI company is firing a quarter of its people because of AI. Cloudflare cuts 1,100 due to AI changes. Challenger Gray’s April report names AI the cause of 26 percent of the month’s job cuts, the highest share since the firm began tracking the variable.

The press release version of the DeepL story is going to be about operational efficiency. The unwritten part is that DeepL was a Claude-and-GPT-wrapped translation business that just watched its core product get commoditized by the labs underneath it. Anthropic and OpenAI ship translation for free now, as a side-effect of frontier capability. DeepL’s product disappeared into someone else’s API. The 25 percent cut is what that looks like at the headcount line. The “operations” framing is the part of the press release you are supposed to read. The part you are not supposed to read is the product line. Every “operations” press release in the next twelve months is going to have a product-line obituary buried underneath it. This is the Clayton Christensen frame we keep coming back to. The bridges across the moat are cheap, and there are a hundred of them now.

Then the MIT economics department published the Acemoglu paper. Five hours after the DeepL announcement. May print issue, Quarterly Journal of Economics. Title: Automation and Rent Dissipation: Implications for Wages, Inequality, and Productivity. The findings are the part of the paper that should be on every CFO’s desk by Friday morning.

Automation 1980-2016 explains 52 percent of the growth in U.S. income inequality. About 10 percentage points of that growth come from a single mechanism, firms specifically replacing workers earning a “wage premium,” meaning workers paid more than peers with similar qualifications. Inefficient wage-premium targeting offset 60 to 90 percent of the productivity gains automation could have produced.

Daron Acemoglu puts it cleanly: “The higher the wage of the worker in a particular industry or occupation or task, the more attractive automation becomes to firms.” The CFO’s incentive is not productivity. The CFO’s incentive is to absorb inflation without absorbing wage growth. In a naturally inflationary world, productivity is the wage-control mechanism, not the output mechanism. That is the Solow paradox restated as CFO behavior. Same line we used in Porsche In The Driveway on May 3: “You can see the computer age everywhere except in the productivity statistics.” Acemoglu’s paper is the proof underneath that quote. The productivity gains were burned controlling labor cost.

In 2026 the new tractor is Claude. The new lever is Codex. The wage-premium worker is the senior associate, the mid-level analyst, the junior engineer with three years of experience and a $180K base. The translation analyst at DeepL is the Acemoglu paper in human form. AI does not kill jobs in aggregate. It kills wage premium. The aggregate count rebounds in role types nobody had a name for two years ago: agent supervisors, FDEs, PMs who code, compound-knowledge reviewers. The wage line stays flat while the inflation line keeps climbing. The CFO smiles. The board smiles. The wage premium walked.

Bob and Bob have the consulting playbook from 1985. Same questions. Different shoes.

Why this matters: The CFO conversation is no longer “what’s our AI strategy.” It is “what’s our wage-premium strategy.” Pull a histogram of your $150K-and-up knowledge-worker headcount. Cross-reference it with the depth-gap profile from the OpenAI report. The intersection of wage-premium plus low-AI-depth is where the cuts happen first. The reverse intersection of wage-premium plus high-AI-depth is the headcount the labs are hiring out from under you. Plan for both before someone else plans for you.

📉 The a16z Chart Is Not The Story. The Footnote Is.

Yesterday a16z published The AI Job Apocalypse Is A Complete Fiction: the bull case, with twenty-six charts. Software engineer demand is rising. Open PM jobs are at a 2022 high. Hiring growth is stronger in “AI-augmented” industries than in “AI-substitute” industries. Job postings for devs will exceed pre-pandemic levels in less than two years. Aggregate employment is roughly unchanged. Doomers are wrong. AI is more job-maker than job-taker.

This is true at the level a16z is writing. It is not the story.

Read the fine print. AI-exposure is driving above-trend wage growth, especially in systems design. Hiring growth is stronger in AI-augmented roles. The roles in decline are customer-service representatives and medical transcriptionists. a16z classifies the entire knowledge-work bifurcation under “neutral aggregate with reallocation between jobs and tasks.” The aggregate number is up. The composition has changed completely.

The 2023 software engineer was a six-figure professional who shipped pull requests. The 2026 software engineer in a16z’s chart is a different human entirely. PMs who code. Designers who prompt. Forward-deployed engineers from Anthropic’s JV. Agent supervisors running fleets of subagents under one keyboard. Andrej Karpathy nailed the role definition when he said on Dwarkesh: stop trying to hire agents. Start supervising them. The supervisor is the new senior engineer. The agent fleet is the new bench. The 1x coder is the customer-service rep of the OpenAI report.

The aggregate count rises because the Bureau of Labor Statistics classifies the supervisor and the 1x coder under the same job title. The same line item in the BLS quarterly release means a different thing than it did three years ago. This is exactly the substitution Acemoglu describes from the 1980s. Bookkeepers did not disappear when Excel arrived. They rebranded as FP&A analysts at higher comp. What disappeared was the typist. The same compression is rolling through software in 2026. Aggregate engineer count rises. Aggregate engineer wage rises. Output per engineer rises 3.5x. The headcount you are about to lose is the headcount in the middle of your salary band that did not yet learn to drive the agent fleet.

a16z is right about the macro. The macro does not help you on Friday’s headcount review.

The bottom line for executives: The narrative the press is reading from a16z is “engineer demand is up, AI is fine.” The narrative your CFO will read from the same data eight months from now is “we can hire 1.5 AI-pilled engineers at $300K each instead of 4 1x coders at $180K each, and our output will triple.” Same payroll budget. Fewer humans. Three times the output. That is the trade. Get there before your competitor does, or get there second after they hire the people who would have closed your gap.

💲 The Q3 Earnings-Call Question Nobody Has Asked Yet

Anthropic’s Dario Amodei stood on stage Wednesday at Code with Claude SF and admitted Q1 grew 80x against a 10x plan. SemiAnalysis had estimated $44B annualized run rate the day before. The CEO confirmed it from the keynote stage. The Information reported Anthropic now accounts for more than 40 percent of Google Cloud‘s $462B revenue backlog. One private company is 40 percent of the largest cloud-services backlog ever assembled. The supply side of the depth gap is real. The lab is grading itself with the Q1 print. The question for the operator is what the Q3 print does to the buyer.

Public-company analysts are about to discover the question they should have been asking the whole time. It is not “what percentage of your engineers use AI tools?” That is the seat-count metric. The seat-count metric is the AI version of a cost-per-click report: relevant, measurable, and roughly meaningless to the productivity number that determines next year’s revenue.

The question is: how many of your engineers are pulling away from the average, and how fast is that count growing?

The CFO who answers that question with a real number (“42 percent of our engineers are operating in the top depth quintile, up from 11 percent six months ago”) gets the multiple. The CFO who answers with seat-count utilization (“96 percent of our engineers have a Claude license”) gets the downgrade. The 3.5x gap is the new gross margin. The CFO who can show the gap inside their company is closing the gap with their competitors. The CFO who cannot is admitting their wage line is about to lose to someone else’s payroll.

Three quarters from now, the earnings-call analyst question is going to be a standing item. It will sound like a softball. It will not be one. Every public AI-exposed company is going to be ranked on the same axis investors used to rank cloud-spend efficiency in 2017 and gross-margin trajectory in 2009. The companies that have prepared for the question will have the language. The ones that have not will sound like they don’t know what they’re doing. They probably don’t.

Here’s what to do: Build the metric internally before someone else builds it externally. Pull your engineering org by depth quintile. Track the count of agent-supervisor roles, PMs who ship code, FDEs deployed in production, anyone whose token-spend-to-output ratio looks like Cisco’s 1,500-engineering-hours-saved chart. Make it a standing slide in your board pack. In Q3 the CEO who has it will be the CEO who has the answer. The CEO who doesn’t will have to explain why.

What This Means For You

The AI-native company in 2026 has redrawn its org chart, redefined its role types, and rebuilt its review process around a colleague that lives in the API. The 3.5x productivity gap is the consequence of that re-org. The companies on the wrong side of the gap are the companies that bought a Codex license and called it a strategy. The Acemoglu paper that just landed says this exact pattern produced 52 percent of U.S. income inequality since 1980. It will produce the next chapter too.

Reorganize, don’t license. A Codex license without a role re-org is an expense, not an investment. The depth quintile is the measurement that matters. Build it. Track it. Pay for it.

Stop hiring engineers. Start hiring agent supervisors. The 2023 1x coder is the wage-premium worker your CFO replaces in the back half of 2026. The 2026 PM-who-codes and FDE-who-orchestrates is the headcount that compounds. Same job title. Different human.

Write the depth-gap slide before your investors ask for it. Q3 earnings calls are six weeks away. The analysts are going to start asking how many of your engineers are pulling away from the average. The CEO who has the answer wins the multiple. The CEO who does not is going to sound like the CFO of an industrial in 2009 who hadn’t run a SaaS gross-margin analysis. The chart already exists at OpenAI’s frontier customers. Build yours.

Read every “operations” press release between the lines. Twelve months from now most of them are going to have a product-line obituary buried underneath the headcount cut. The labs ship the side-effect that used to be your moat. The press release blames operations because the product blame is unspeakable. Read both. Plan for both.

The 3.5x is invisible on a seat-count report. It is the most visible thing on a competitor’s payroll line. Bob and Bob are walking the floor. The question is what would you say… you do here.

Three Questions We Think You Should Be Asking Yourself

What percentage of your engineering org is operating in the top depth quintile, and how fast is that percentage growing? If you cannot answer this in numbers, your AI strategy is a license bill. Build the metric. Pull the histogram. Tell the board what you are going to do about the bottom four quintiles.

Which roles in your company are still at headcount because the org chart hasn’t caught up with the technology? Audit by salary band. The wage-premium senior associate, the mid-level analyst, the junior engineer with three years and a $180K base: if their work is being done in production today by an agent fleet supervised by one of your AI-pilled engineers, the headcount is structural debt. Plan the migration before your CFO plans it for you.

If a Q3 earnings analyst asked you tomorrow how many of your engineers are AI-pilled, what would you say? “I don’t know” is the wrong answer. “We have 96 percent license utilization” is the wrong answer. The right answer is a percentage, a trajectory, and a set of named role types that did not exist on your org chart eighteen months ago. If you don’t have the answer, six weeks is the budget you have to build it.

The higher the wage of the worker in a particular industry or occupation or task, the more attractive automation becomes to firms.”

— Daron Acemoglu, MIT Institute Professor, Quarterly Journal of Economics, May 2026

— Harry

Signal/Noise by CO/AI is published most weeknights from New Canaan, Connecticut. The point is to make you the smartest person in the room without taking more than fifteen minutes of your morning. If we did, forward it to one person. If we didn’t, hit reply and tell us why.

Sources

• OpenAI: Introducing B2B Signals (May 6, 2026)
• AI Ready Show / Haroon: The 3.5x intelligence gap inside enterprises is widening (May 7, 2026)
• a16z: The AI Job Apocalypse Is A Complete Fiction (May 6, 2026)
• MIT News: Study — Firms often use automation to control certain workers’ wages (May 7, 2026)
• Acemoglu & Restrepo: Automation and Rent Dissipation: Implications for Wages, Inequality, and Productivity, Quarterly Journal of Economics, May 2026 print issue
• The Neuron: Anthropic doubled Claude Code limits and locked up Elon’s Memphis data center (May 7, 2026)
• TLDR-AI: Higher Limits for Claude and a Compute Deal with SpaceX (May 7, 2026)
• CNBC: Anthropic CEO Says Company Grew 80-Fold In First Quarter (May 6, 2026)
• Challenger, Gray & Christmas: AI emerges as a top cause of layoffs (April 2026 report)
• Aligned News: The Job Paradox Nobody Is Talking About (May 7, 2026)
• Andrej Karpathy on Dwarkesh Podcast (October 2025)
• Office Space (1999) — Bob Slydell scene