Last week, reports put OpenAI’s yearly sales pace $18 billion below the figure investors had heard in late September. Shares of Nvidia, Oracle and CoreWeave fell the day the reports came out. All three sell chips or computing power to AI labs like OpenAI. Their growth depends on whether those labs can keep paying for what they’ve agreed to buy.

The filings below show which companies bear the risk if OpenAI’s sales grow more slowly than expected. They also show the size of each company’s commitments to its AI customers.

1. OpenAI: $50 billion or $68 billion

OpenAI told its investors that its annualized revenue was roughly $50 billion at the end of September. Annualized revenue takes what a company earned over a recent stretch, such as a month, and multiplies it out to a full year. It tells you how fast sales are coming in today.

In late September, news reports had put that figure at $68 billion. According to a person familiar with OpenAI’s finances, that larger figure included revenue from OpenAI’s partners in addition to OpenAI’s own sales.

OpenAI is a private company. It doesn’t publish audited financial reports the way public companies must. Investors who want verifiable figures have to look at the reports of public companies that do business with OpenAI.

Subscribe now

2. Nvidia: up to $105 billion

SB Energy is developing a large data center campus in Ohio. OpenAI has agreed to rent space on the campus and run Nvidia’s AI systems there.

Nvidia has signed guarantees that back OpenAI’s leases with SB Energy. If OpenAI goes bankrupt or stops paying rent, Nvidia pays SB Energy the difference between a minimum lease value set in the agreements and whatever SB Energy recovers by renting the space to someone else or selling it. Nvidia’s payments are capped at $105 billion. OpenAI has agreed to pay Nvidia back for anything Nvidia pays out. The guarantee ends once OpenAI earns a strong enough credit rating, the grade lenders give a company based on how likely it is to pay its debts. Until OpenAI earns that grade, Nvidia’s guarantee protects SB Energy from losing income on the leases. The leases start when the first buildings open, expected in 2028.

Subscribe now

3. Oracle: $664 billion in backlog

Oracle has $664 billion in remaining performance obligations. That’s the total value of contracts customers have already signed for services Oracle hasn’t delivered yet. Think of a landlord who signs a tenant to a five-year lease. From the day the lease is signed, the landlord knows five years of rent is coming. Each month’s rent still comes in one month at a time. Oracle expects to earn about 13% of its backlog, roughly $86 billion, over the next 12 months. Oracle says the backlog grew from $455 billion a year earlier mainly because of a few very large cloud contracts. It doesn’t say who signed them. You can’t tell how much of the $664 billion depends on OpenAI.

Last quarter, Oracle spent $28.5 billion on data centers and equipment. Its day-to-day business brought in $23.1 billion in cash. Customers paying in advance made up $11.4 billion of that cash. In the same quarter, Oracle raised $19.9 billion by selling new shares. It pays to build its data centers today. Its customers pay for the computing power inside them over the life of their contracts.

Subscribe now

4. CoreWeave: one customer, 36% of revenue

CoreWeave rents out computing power built on Nvidia chips to AI companies. Three customers paid 72% of its revenue last quarter. The largest one accounted for 36% on its own. If that one customer cut back, CoreWeave would lose about a third of its sales. CoreWeave hasn’t identified the customer, but it does disclose two of its largest contracts. OpenAI agreed in September 2025 to pay CoreWeave up to about $6.5 billion for computing power through May 2031. Meta agreed in March 2026 to pay up to about $21 billion through December 2032. CoreWeave expects OpenAI and Meta to become major customers in the years ahead.

Subscribe now

5. Applied Digital: CoreWeave’s landlord

Applied Digital builds and owns data centers, then leases them to companies that run AI computing inside them. Its first campus, Polaris Forge 1, is leased to CoreWeave. CoreWeave pays that rent out of the revenue it earns from its AI customers.

Last quarter, Applied Digital collected $65.8 million in base rent and recorded $77.4 million in interest expense. It has signed leases for 1.41 gigawatts of capacity across five campuses, worth about $36 billion in rent over their initial terms. Only 250 megawatts of that is built and running today. That’s less than a fifth, since a gigawatt is 1,000 megawatts. The company carried $6.4 billion in debt at the end of last quarter. It also earned $35.8 million in interest on $3.7 billion in cash on hand. After that income, its net interest cost was $41.6 million, below its base rent of $65.8.

Last week, paid subscribers got a calculator for finding the break-even point of an AI system that still depends on human oversight, built from the Presto Automation SEC case.

GET THE BREAK EVEN POINT CALCULATOR

Presto sold stock to investors during those same years and told them the AI completed more than 95% of orders without help from restaurant staff.

What Presto didn’t tell investors was that it had hired people in the Philippines and India to listen to the drive-thru orders and type them in. Those people handled what a US regulator later described as ‘the vast majority of drive-thru orders.’

That regulator was the Securities and Exchange Commission (SEC), the agency that polices what public companies tell their investors. The SEC investigated Presto and concluded in January 2025 that the company’s statements to investors had been false and misleading. It explained that Presto counted an order as completed by the AI as long as no restaurant employee stepped in, including orders Presto’s own overseas staff had typed. Presto accepted the SEC’s order without admitting or denying its findings.

By the time the SEC published those findings, Presto’s stock had been delisted from Nasdaq, and the business had been sold to new owners. The SEC didn’t fine Presto because of the company’s financial condition and its cooperation with the investigation.

Using people to support an AI system is legal. You may run one that way yourself if someone on your staff finishes the tasks your AI can’t do or reviews what it produces. Presto had two problems that you can check for in your own AI system.

1. Presto’s AI cost more to run than it earned. Each restaurant paid Presto a fee, and Presto paid the off-site staff who typed in that restaurant’s orders. The product could only make money when the AI entered enough orders on its own. Presto’s AI was far below that level.

2. Presto described its AI with a figure that left out its own staff. Investors saw 95% and had no way to know how many orders people typed.

Read more

How many large US companies report a measurable AI result

Every quarter, a public company’s executives review its financial results with investors on an earnings call. FactSet, a financial data company, searched last quarter’s S&P 500 earnings calls for the term ‘AI’ and found it on 67% of those calls.

Morgan Stanley went a little deeper with its analysis, counting the companies that cited at least one measurable benefit from AI on their earnings calls and at investor conferences. They found that only 24% of S&P 500 companies did. That means roughly three-quarters of these companies gave investors no measurable benefit from adopting AI.

Morgan Stanley’s definition of a measurable benefit is generous. It includes a bank that told investors it had 117 AI solutions running. The bank attached no revenue or savings to that count. Even with such a broad definition, only 24% of S&P 500 companies qualified.

That 24% reflects what executives chose to tell investors. A company in the other 76% could still be benefiting from AI without having reported a figure.

Microsoft is one of the companies that reports AI figures to investors. I compared what its executives said on earnings calls with the annual report it filed.

Subscribe now

What Microsoft told investors about Copilot

On July 29, Microsoft CEO Satya Nadella told investors on the company’s earnings call, ‘we now have over 30 million paid Microsoft 365 Copilot seats.’ Copilot is the AI assistant Microsoft sells inside Word, Excel, and Outlook. A seat is one paid license for one person.

Microsoft filed its annual report with the Securities and Exchange Commission (SEC) that same day. The SEC is the US agency that regulates public companies. The annual report is the yearly document in which a company reports its finances, risks, and each division’s results. I searched Microsoft’s report for ‘30 million’ and for any count of Copilot seats. Neither search returned a Copilot figure.

The phrase ‘Copilot seat’ appears once in the report. Microsoft wrote, ‘Cost of revenue increased $2.6 billion or 12% driven by investments in AI infrastructure to support Microsoft 365 Copilot seat and usage growth.’ Cost of revenue is what a company spends to deliver the products it sells.

On the call, Microsoft told investors how many Copilot seats it has. In its annual report, Microsoft reported what those seats cost to grow.

That matters because ‘30 million seats’ is the kind of figure people repeat as proof that AI is paying off at large companies. A company answers for a figure differently depending on the document it’s in.

If you’re new here, these are more pieces from The Data Letter on AI economics:

• Calculate AI ROI Your CFO Will Approve. Put a dollar figure on one AI workflow in about 90 minutes. You get a spreadsheet with the break-even formula built in, nine questions for the person who runs the workflow, and a one-page summary for your finance team.

• NVIDIA and AI Inference Economics in 2026. NVIDIA and Google say the cost of serving AI is falling by up to 10x. I trace who keeps those savings and why your workload hasn’t gotten lighter.

• How Uber Kept AI Spend Flat Through a 9.4x Increase in Requests. Uber’s engineers grew their weekly AI agent requests 9.4x and kept total AI spend flat. I break down the changes behind that result and give you a worksheet that finds where your own AI spend goes.

• LLM Inference Costs. Running a model accounts for 60% to 80% of total LLM spend over a year, far more than training it. I cover five ways to cut that cost and a client case where annual inference spend fell from $50K to $15K.

Three documents where a company reports AI results

To compare Microsoft’s statements with its annual report, I sorted each AI figure by the document it appeared in. I used three groups. Here’s what each means:

• Said. An executive stated the figure on an earnings call.

• Furnished. The figure is in the press release a company publishes on the day of its earnings call to announce that quarter’s financial results. The company sends that press release to the SEC, so it becomes part of the public record. Under SEC rules, it doesn’t count as a filed document, so the company faces less legal liability than with its annual report.

• Filed. The figure is in the company’s annual report, the 10-K. Before filing, the company has an outside accounting firm audit its financial statements. The firm’s audit report is published inside the 10-K.

Microsoft’s AI claims compared with its annual report

On its April and July earnings calls this year, Microsoft’s CEO gave investors three AI figures. I looked for each one in Microsoft’s press releases and in its 10-K.

• 30 million paid Copilot seats (Furnished). Nadella stated it on the July call. Microsoft’s July press release includes it too. The 10-K has no Copilot seat count. Microsoft reports there that Microsoft 365 Commercial seats grew 6%, a figure for the product family that includes Copilot. Microsoft credits Copilot for revenue growth in one sentence of the 10-K. That sentence shares the credit with a second product and gives no dollar amount.

• $37 billion AI revenue run rate (Furnished). On the April call and in the April press release, Nadella said, ‘Our AI business surpassed an annual revenue run rate of $37 billion.’ A run rate estimates a full year of revenue from a recent shorter period. If a business earned $1 billion last month, its annual run rate is $12 billion. The phrase ‘run rate’ doesn’t appear in the 10-K.

For Intelligent Cloud, the division that includes its Azure cloud business, Microsoft reports that revenue rose $31.5 billion, or 30%, and cost of revenue rose $17.7 billion, or 44%. Microsoft attributes the cost increase to ‘investments in AI infrastructure’. It attributes the Azure revenue growth to ‘demand for services across the platform’ and gives no AI share.

• 43 minutes saved per day (Said). On the July call, Nadella said a trial at NHS England, the public health service, showed Copilot ‘saved employees an average of 43 minutes per day.’ The July press release doesn’t include it. I wouldn’t expect a customer’s trial result in an annual report. This figure is the only one of the three that measures what Copilot does for the people who use it. Microsoft gave it on the call alone.

Microsoft’s 10-K attributes two cost increases to AI, $2.6 billion in the division that sells Copilot and $17.7 billion in Intelligent Cloud. The Copilot seat count and the $37 billion AI revenue run rate appear in Microsoft’s press releases. Neither one appears in the 10-K.

Both divisions earned more than those cost increases. Operating income rose $14.1 billion in the division that sells Copilot and $12.4 billion in Intelligent Cloud. Microsoft’s 10-K doesn’t say how much of either increase came from AI.

Subscribe now

Why Microsoft’s annual report includes OpenAI revenue and no Copilot seat count

Microsoft’s 10-K does include one AI revenue figure. Microsoft recorded $24.1 billion in revenue from OpenAI in fiscal 2026, which covers only its business with OpenAI.

Microsoft owns roughly a 25% stake in OpenAI. US accounting standards treat a business with that level of ownership as a related party and require a company to disclose transactions with related parties.

That figure is in the 10-K because a rule requires it. No rule requires Microsoft to file its Copilot seat count or its AI revenue run rate. Microsoft didn’t file them.

The SEC doesn’t require companies to report the revenue or savings they attribute to AI. Its general standard requires a company to disclose a trend that management knows about and that’s ‘reasonably likely to have material effects’ on its finances. Material means large enough to affect an investor’s decision. Company management decides whether a trend meets that standard.

When the SEC has addressed AI directly, it has focused on honesty. Its former chair, Gary Gensler, commented on AI washing and said companies that discuss AI ‘should say what they’re doing, and do what they’re saying.’

Subscribe now

What Microsoft’s filing tells you about corporate AI claims

Microsoft broke no rules by leaving the Copilot seat count out of its 10-K. It filed what its AI growth cost. It kept the size of that growth in a press release and on a call.

About 3/4 of S&P 500 companies gave investors no measurable AI results on their calls last quarter. Microsoft is one of the S&P 500 companies that gives investors AI results. Microsoft put none of the three results I checked in its 10-K.

The case that AI is paying off at large companies currently rests on what executives choose to say on earnings calls. Microsoft’s filed annual report includes the spending it attributes to AI and no figure for what that spending returned.

Which company’s AI claims should I compare with its annual report next?

Paid subscribers of The Data Letter get every audit I publish, including company claim audits, AI infrastructure spending breakdowns, and explainers on AI costs.

Upgrade

So I stopped comparing feature lists and started breaking things.

I built the same small agent three times, once in Python, once in LangGraph, and once in n8n. The agent is deliberately plain, so the only thing changing between runs is the tool underneath.

Then I put each one through two failures every agent meets eventually. I crashed it at the worst possible moment, and I restarted the whole tool while a run was still going. One tool brought a paused run back from the dead. Another did the same thing twice, the kind of mistake that charges a customer twice. The protection that would have stopped it was missing from all three.

My paid subscribers make it possible for me to keep running builds and tests like these.

Below, I break down exactly how I ran each failure, what every tool did when it hit them, and the one piece you have to build yourself no matter which tool you pick.

Read more

By the end of this, you’ll be able to see every step your agent takes, with a rule that stops it before it does something you didn’t approve.

You want to follow an agent through every step, so you catch a mistake the first time it happens instead of after it has cost you. In yesterday’s live, I gave you the three parts that make it possible: log what the agent read and did, trace a single run from the request to the final action, and gate the actions you’d want to sign off on, so the agent pauses for you before it runs one.

Today we’ll build the agent and add all three parts.

We’re building it on a support agent that issues refunds through Stripe. This is the same observability layer I built for my own agents at Asaura AI. The setup works the same on any agent you run.

We’ll build the agent, add the three parts, and add a rule that pauses the agent before a step you’d want to check, so a person signs off before it goes through. On this agent, that step is a refund. On yours, it might be sending an email or changing a record.

Read more

He never got the chance, so nobody in the room knew what they were missing by cutting him out.

Friday, I called this the Compounding bucket: the work that’s worth more the more your teams build. When every department can build its own tools this fast, someone has to see how those tools fit together and what each one costs the company over time. This skill gets more valuable the more your company builds. This article gives you a scorecard to measure this. You give an internal tool five scores, and it tells you what that tool will cost your company and whether you should back it, fix it, or flag it to leadership.

A tool that works today has cleared the lowest bar. It ran once, for one team, on the machines it was built on. What that tool costs your company shows up over the year that follows. Software maintenance and evolution runs to roughly 60%-80% of a tool’s total lifetime cost, and yearly upkeep alone often runs 15%-25% of what the tool cost to build. A tool a contractor builds in an afternoon costs far more than that afternoon. It brings a year of upkeep, security fixes, and support that nobody counted when they approved it.

A contractor building a tool for one department answers one question. Can I get this working for them now? That’s the right question for them to answer. Your value is answering the one their job doesn’t ask them to. What will this tool cost the company once it’s running, and does it work well with everything else your teams have built?

Read more

I’m really excited for this one today.

When we heard the story of how Uber blew through their budget by April this year, it was a cautionary tale about how fast agentic AI can rack up costs when a large engineering team adopts it at scale.

Getting all those headlines, however, didn’t mean Uber’s engineering team gave up on finding more efficient ways to use AI in their work. A couple of weeks ago, they published a follow-up blog post on their success since then. Weekly agent requests had grown 9.4x since earlier this year; AI agents were now producing more than 70% of the company’s code-change submissions; engineers were running more than 30,000 agent tasks a day; and total AI spend had stayed flat since April. Cost per 1,000 requests was down 34% from the April peak. Cost per session was down 52% from the June high.

The new story at Uber is: usage kept climbing while spend stopped climbing. Between April and August, Uber’s engineering team completely rebuilt how they use Claude Code.

Uber’s situation earlier this year isn’t unique. 93% of enterprises are still exceeding their AI budgets. The reason is that 60% of total agentic AI spend goes to response refinement, which is the iterative back-and-forth of the agent checking its own work, correcting mistakes, and improving outputs before it returns anything to the person who asked. That’s separate from the initial call to the model. On top of that, only 22% of per-task AI cost is the token cost itself. The other 78% is everything wrapped around the model call: tool calls, database queries, human review, and compliance infrastructure that never appears as a line item on the token invoice.

I went live on Friday with Why Cheaper AI Is Making Your Company’s Bill Bigger and gave you three cost categories to attribute that spend to. Context churn refers to the part of your input tokens that contains historical information from earlier interactions in the same task, which the model resends each time the agent takes another step. The unreliability tax is what your company pays for retries when a step in the agent’s loop fails. Orchestration overhead is everything you spend to make the agent run at all, including the tools it connects to, the framework it runs inside, and the monitoring watching it, regardless of whether the agent uses any of it on a given request.

Understanding those three categories is the first half of the battle. The second half is understanding what companies who’ve solved this are doing differently, so you can copy it.

Uber’s engineering post lays out roughly ten interventions they layered together to stabilize spend while agent request volume grew 9.4x. It’s a dense playbook written for teams at Uber’s scale. Below are the highest leverage interventions from that playbook, extracted and translated for teams who don’t have 5,000 engineers, with the diagnostic, math, and implementation you need to put them into practice today:

• A diagnostic tool that shows how much of your budget each of the three cost categories takes up.

• A step-by-step walkthrough for setting up your agent to send simple tasks to cheaper models and reserve the expensive ones for genuinely complex work

• The three specific changes your team can make to your prompt structure and cut your LLM spend by 59% - 70%.

• Two operational changes Uber made that don’t require code changes to your agent and can be rolled out this week.

Read more

On yesterday’s live session, I walked through why every AI workflow at your company needs a test layer, and introduced the two kinds of evals every team using AI should run. Capability evals measure what the AI does well and where it struggles. Regression evals measure whether the AI still handles the outputs your team already relies on. Together they answer whether the AI is safe to use, and whether the work you’re putting into improving your processes is paying off.

Today I’m giving you the entire testing system. A workbook with both tests (capability and regression) already set up, so you can drop your test cases straight in. A starter list of 21 conditions an AI output can be required to meet, sorted into six categories, that your team can copy from and adapt to your workflows. An AI grader prompt that scores each output against each condition, so one person can run both tests in the time it used to take to manually review a single output.

Here are some recent popular articles you may have missed:

➡ A four-artifact workflow for calculating AI ROI in a format your CFO will approve, built in 90 minutes.

➡ The intake form, screening guide, vendor checklist, and documentation workflow every enterprise AI project now runs through under U.S. state AI regulations.

Read more

Today’s build turns that definition into a working system you can copy.

A multi-agent system is an AI setup where one agent coordinates the work of two or more other agents to finish a job. The coordinating agent is called the parent. The agents it coordinates are called subagents. The parent decides which subagent handles each part of the job, reads each subagent’s output, and either combines the finished outputs into a final result or sends a subagent back with a narrower request if the output isn’t specific enough.

The reason to use a multi-agent system instead of one large agent is that each subagent gets to focus on a single task with its own instructions. That focus produces better output than asking one agent to keep every task in its memory at once. It also means the parent can send two subagents off to work at the same time, which gets the job done faster than doing every step in sequence.

The build below turns this diagram into a working system on your machine.

What it includes: the three files you can copy directly into your own setup, a step-by-step walkthrough of a live run with every part of the system labeled inside the transcript, the follow-up request that shows the parent agent adjusting its next step based on what came back, and a set of concrete swaps for adapting this pattern to other recurring jobs your team runs every week.

Here are some recent popular articles you may have missed:

➡ An n8n agent that reads your live metrics and reports them every Monday morning, grounded in a real spreadsheet so it stops inventing numbers.

➡ A four-artifact workflow for calculating AI ROI in a format your CFO will approve, built in 90 minutes.

➡ The intake form, screening guide, vendor checklist, and documentation workflow every enterprise AI project now runs through under U.S. state AI regulations.

In a few minutes, you’ll have a working multi-agent system running. Let’s build it.

Read more

That decision fits on one page. I call it an Exit Map. Down one side you list the kinds of data your company works with. Across the top, where each kind is allowed to go, whether that’s a personal AI account, your company’s approved account, or no AI at all. Public marketing can go anywhere. Internal documents go only to the company account. Send confidential client data and credentials to no AI at all.

I went through this live yesterday. You can watch the full episode here.

Deciding is the easy part. The hard part is making your tools obey that decision. The settings that do it are inside each platform’s admin area, and few teams switch them on.

Do this and your team keeps the AI they want, your confidential data stays in-house, and when your annual review comes you can show which kinds of data your team may send to an AI tool, and which never go near one. Below, I walk you through that setup for both Claude and ChatGPT, one row of your Exit Map at a time. Upgrade to get every step. By the end, your team’s AI tools will be locked down so your private data can’t slip out, and you’ll have the records to prove it if you’re audited.

Read more

Yesterday’s live session opened with a diagram from a 2015 Google paper called ‘Hidden Technical Debt in Machine Learning Systems’. That paper made an argument that changed how teams build with AI: writing the model is the fast, cheap part. Running it inside a company, and keeping it working over time, is where the cost and the risk sit. By the authors’ own estimates, a mature AI system is around 5% model and 95% that surrounding infrastructure.

That 95% is where AI projects fail, and it’s what you get asked to approve as a leader, even though it’s the hardest piece for a non-technical person to assess. So on the live, I gave you a way to break any AI project into five layers, and for each layer, the one question that tells you whether it’s a good investment or a risk. Together, the five answers show you where a project is strong and where it will cause trouble, before you commit budget or scale it. Those five layers are model, compute, data, orchestration, and governance. Work through them in order, and you can sit in a leadership meeting and say, with evidence, which layer is ready and which one needs work before the company invests further.

If you missed it, watch the recording here. It includes a link to the deck and the Google paper. This article picks up where that session ended.

What you’re building today

On the live, you learned how to read an AI project across the five layers. Today you turn that into something you own. You’ll build a Claude Project that takes about 30 minutes of notes from a working session with your team and returns a leadership-ready assessment that shows where the project is solid, where it’s exposed, and what to resolve before you scale. Build it once, and you reuse it on every AI project after this, without relearning the framework or writing the assessment yourself.

Below this line is the full build. You get the questions to work through for each layer, the scoring system that shows where the project is solid and where it’s exposed, the instructions you paste into Claude to set the tool up in 10 minutes, and a worked example that turns one messy set of notes into an assessment you could put in front of your CEO that same afternoon.

Read more

It’s more difficult than it seems because no one documented the costs of running a workflow before the tools were introduced, leaving you without earlier figures to compare today’s performance against. Usage goes up, people report saving time, but none of it converts into dollars anyone can put in a budget.

This article gets you that figure for one workflow, and gets it accurate enough to put in front of your finance team.

The method is a Friction Ledger, which is three columns you fill in for a single workflow before you spend anything. Column one is frequency, which is how many times a week the task happens. Column two is proximity to cash, which is how many steps sit between that task and cash entering or leaving your company. Column three is handoff cost, which is how many minutes a person spends on the output after your AI produces it. Column three decides whether the workflow returns dollars or costs them, because every minute your reviewer spends is a minute subtracted from what the AI saved.

Thursday’s live session ran that through a worked example. You can open the slides I used for that here.

Getting from the method to your own figure takes about 90 minutes, spread across a week.

Every workflow has a review time above which it costs your company more than it returns. Cross that line and your AI spending is a loss, however good the tool is. Almost nobody knows their own figure, and one formula produces it in under a minute.

The rest of this article gives you four things:

1. A spreadsheet with the break-even formula already in it, where you enter your own workflow’s numbers and read the annual return off the bottom.

2. Nine questions to ask the person who usually manually runs the workflow, so your timings come from watching rather than remembering.

3. A prompt that turns those numbers into a one-page summary you can send to your finance team.

4. Six slides you can copy for the meeting where leadership decides whether to keep paying for your AI tools.

Read more

First, the four regulations reshaping enterprise AI in 2026:

• The June 2nd federal executive order on frontier model pre-release review.

• State-level frontier AI laws, led by California TFAIA and New York RAISE.

• State-level automated decision-making laws, led by California ADMT, Colorado SB 26-189, and Texas TRAIGA.

• Federal procurement guidance from the White House Office of Management and Budget.

Second, why enterprise AI adoption is no longer a software-access problem. The review functions your company already runs still apply, but each one now has to evaluate categories of exposure that traditional enterprise procurement was never designed to handle.

Third, the four Adoption Gates every AI project now clears before deployment: vendor selection, use-case screening, monitoring, and documentation.

Fourth, three practical steps to prepare: build a use-case intake before you build any pilot, ask vendors for evidence instead of assurances, and reduce your company’s dependency on any single AI vendor’s release schedule.

📌 A quick note before we start. My paid subscription price is going up at the end of day July 31st from $5/month or $50/year to $10/month or $100/year. If you upgrade before then, or if you’re already a paid subscriber, you’ll stay at the current price for as long as your subscription remains active.

What this piece gives you

The live gave you the framework. This piece gives you the workflow.

By the end, you’ll have six artifacts your team can install this week:

• A one-page use-case intake form.

• A screening guide that maps every AI project to the specific state laws that apply to it, across the seven U.S. states with active or upcoming AI regulations.

• A vendor evidence checklist covering eleven categories of documentation your legal team should require before any AI vendor evaluation begins.

• A monitoring log schema your data engineer can implement without additional infrastructure.

• A documentation template designed to survive a regulatory audit.

• A build order that shows your team how to put all five into place, in the sequence that saves the most rework.

Read more

Today, you’re going to be that someone. You’re going to build the thing Microsoft and AWS charge millions for, at your own desk.

What you’re building

By the end of this build, you’ll have a working, autonomous agent that watches your team’s support intake queue in Notion, reads a written playbook for how your team already answers common questions, drafts a response in your team’s voice, and hands that draft to a person for approval before anything goes out.

It checks the queue every 30 minutes on a set schedule; no one has to open a chat window and trigger it. I set mine to check only during work hours, since a person still needs to be around to review what it drafts, but you can widen that window to whatever hours your team covers.

I ran this against my own company’s queue, and every screenshot in this piece comes from that run. My team gets a handful of the same questions from new users, over and over. That’s the type of request this build is for.

You’ll use Claude Code to run the agent, and Notion to store incoming requests and the playbook that agent follows. If your team tracks requests somewhere else, the same setup works with whatever you’re already using.

Read more

That decision happens dozens of times a day. Every ticket follows the same steps and uses the same types of information, even though the tickets themselves are different. High volume and a repeatable process are what make a decision a good candidate for AI to handle.

Every team has decisions like this. Support triage. Sales escalation. Refund reviews. Vendor evaluations. Once you can name these decisions, you can build AI tools that handle the parts that don’t need human judgment.

The way you identify these decisions on your own team is a 4-step audit. Recognizing these opportunities shows your team’s proactive approach and builds confidence in your ability to leverage AI effectively.

Step one names the decision that recurs.

Step two names who does each step today.

Step three names the information each step needs.

Step four circles the steps an AI model could handle.

Those four steps together give you a clear specification you can hand to your Revenue Operations team, or use to build the tool yourself in Claude Projects, empowering your team to innovate and improve efficiency.

Subscribe now

What’s below

The full audit walkthrough applied to customer support triage. A worksheet template you fill in for any decision on your team. The four outputs of a completed audit, and what each one tells you about the tool to build.

Then the Claude Project: the system prompt with six triage criteria, three sample tickets to calibrate the criteria before you connect anything, the connector setup for your support inbox and CRM, the approval flow that keeps a person in the loop until your team trusts the scoring, and four other decisions you can adapt this same setup to inside your team.

Setup takes about an hour once you’ve picked the decision you want to automate. If a support team of five is spending 45 minutes per day each on manually triaging, this build gives them back about 18 hours a week combined. That’s roughly half a full-time work-week of hours, returned to the work your team does best.

Read more

She didn’t have to search for the trend herself.

That advertiser’s workflow is one example of what Pinterest changed in June. The same pattern is rolling out across its advertiser tools, and operators outside ad-tech should look closely at how Pinterest is using its data to take repetitive analysis off the human side of the workflow.

What Pinterest Shipped a Week Before Cannes

On June 17, Pinterest announced four AI products ahead of Cannes Lions: Business Assistant, Pinterest MCP, a new Performance+ creative model, and Ask Pinterest. The announcement reads like an ad-tech story. Underneath the ad-tech framing, Pinterest is using a decade of taste and intent data to automate the analysis and selection work advertisers used to do themselves.

In the announcement, Pinterest’s Chief Business Officer Lee Brown said, ‘The future of discovery won’t be driven by keywords alone. It will be shaped by context, taste, and trusted recommendations.’

Users don’t want to type queries anymore. They want the platform to already know what they’re looking for, and Pinterest is rebuilding its advertiser tools around the same idea.

Subscribe now

Which Advertiser Decisions Pinterest’s AI Now Handles

Each of the four products replaces a task someone on the advertiser side used to do manually.

• Business Assistant automates the trend monitoring a campaign manager used to do by scanning dashboards.

• The Performance+ creative model automates the asset selection that a creative lead used to do through A/B testing.

• Pinterest MCP gives an agency analyst direct access to Pinterest campaign data and analytics from inside the analyst’s own working tools, instead of having to log into Pinterest separately to pull the same numbers by hand.

• Ask Pinterest automates the multi-step planning a shopper used to do across several searches.

The people doing the work stay in place. What changes is the part of the job they spend time on.

A campaign manager defines what a winning campaign looks like and reviews the recommendations Business Assistant surfaces. A creative lead sets the brand voice and approves which AI-selected variant runs. An agency analyst advises the client using insights MCP delivers automatically. The strategy and judgment stay with the human. The retrieval, comparison, and selection now sit with the model.

Advertisers spend less time on retrieval and comparison and more time on strategy, judgment, and client relationships.

Subscribe now

Why This Matters for the Team You Run

Pinterest spent two years building infrastructure to automate the repetitive analysis its advertisers used to do themselves. The Performance+ model improved click volume by 7.5% in Pinterest’s own testing. Pinterest MCP is letting agencies like PMG, Pacvue, and Omnicom’s Jump450 plug Pinterest insights directly into their AI workflows.

The pattern is borrowable. The technology is open. Pinterest MCP runs on the same protocol your team can run.

Technology is the easy part. The harder part is auditing your team’s decision flows in enough detail to see which steps an AI tool could handle on its own. Without that audit, applying AI to your team stays abstract.

Many operators can name three or four obvious time sinks. Status meetings. Slack triage. Pulling numbers for the weekly review. Those are the symptoms a team sees.

The deeper layer is the work nobody tracks. Context switching. Re-reading yesterday’s threads to remember what was decided. The seven-minute Slack derail that costs forty minutes of regained focus. The constant low-grade pattern-matching across dashboards that no one would call work but everyone does anyway.

Until a team can name those untracked tasks specifically, no AI tool can handle them. The audit has to come first.

In case you missed last week’s Data Letter sprint, it covered the same problem from a different angle. Samsung is rolling out ChatGPT, Gemini, and Claude to all 260,000 of its employees and measuring the rollout by full-time-equivalent.

• Samsung Gave 260,000 Employees AI. Few Will Use It Well. explains why access to AI doesn’t change how the work gets done.

• Build the AI Scoring System Samsung’s 260,000 Employees Aren’t Getting is Wednesday’s live recording, where I built a working scoring system from scratch.

• I Built an AI Lead Scoring System That Reads My Inbox and Writes to HubSpot. Here’s How. is the full implementation guide: the Claude Project setup, the system prompt, and the HubSpot connector flow.

One Thing to Try This Week

Pick one decision your team makes every week. Something like ‘which support tickets get escalated this morning’, ‘which leads sales should prioritize, or ‘which marketing campaign deserves more budget this week’. The decision should recur often, have a clear input, and produce a clear output.

For each step in that decision, write down two things: the person who currently does the step, and the data or context they pull to do it. For a ‘pipeline coverage review’, for example, the person is the sales manager and the data is the current week’s deal stages from the CRM.

Many teams find that around 60% of the cognitive effort involves retrieval, comparison, or pattern matching. Work an AI tool can do faster than a tired manager at the end of the week.

If an AI tool handled the 60% that’s retrieval and comparison, the hours your team spends on that work would free up for the 40% that requires judgment, strategy, and client time.

Pinterest finished that audit for its advertisers two years ago. The same exercise is overdue on many teams.

Next Step

This Thursday, July 2nd, I’m going LIVE to walk through how AI is redistributing cognitive labor across organizations. What the pattern looks like, where it’s working, and the framework operators can use to map their own teams.

On Thursday, I’ll release the full build: Map Where Your Team’s Mental Energy Goes, and Build an AI to Redistribute It. The Pinterest story above explains why this redistribution matters. Thursday’s build walks through a full example: a team’s weekly decision flow, the audit that identifies the 60%, and the AI tool built to handle it. You can run the same approach on your own team’s workflow.

Subscribe now

The way Samsung plans to measure whether the rollout works is the part of the story every operator and manager needs to understand, because every other large enterprise AI rollout has hit the same wall: access without a system for using that access well.

What Samsung Just Did for 260,000 Employees

Samsung Electronics has around 260,000 employees globally. By the end of 2026, every one of them will have access to ChatGPT, Gemini, and Claude. The DX Division, which builds phones, TVs, and home appliances, gets access first. Every other Samsung subsidiary, from Samsung Display to Samsung SDI to Samsung Biologics, follows after.

To get there, Samsung is running a top-down training program. About 50 senior leaders are training first. 2,300 executives follow through August. Then everyone else, in waves, through the rest of the year.

The metric Samsung is using to measure whether the rollout worked is full-time equivalent, or FTE. One FTE equals one person working a 40-hour week. Samsung will measure AI success by how many people’s worth of work a single employee can do with AI. If a marketing manager using ChatGPT produces what used to take three people, that’s a 3.0 FTE.

Samsung is the first of Korea’s four largest business groups, called chaebols, to fully adopt external AI tools across all its companies. The other three are SK, Hyundai, and LG. Together, these four groups account for a large share of Korea’s industrial economy. Whatever Samsung does, the other three watch. Whatever Samsung measures, the other three are likely to measure too.

The FTE metric is worrying some Samsung employees, who said AI adoption ‘could be perceived as a measure for ultimately determining how many jobs can be cut rather than being used as a means of improving work efficiency.’

Here are some recent popular articles you may have missed:

I Built an AI Agent That Sends Me My Numbers Every Monday Morning → A step-by-step n8n build for an AI agent that reads your live metrics, remembers context, recovers from failures, and runs on its own every Monday.

Build Your Own Local AI Stack: A Session on Models, Hardware, and Quantization → How to choose an open-weight model that fits your local machine, runs without a cloud subscription, and can’t be shut down by a vendor.

3 Pieces of Free Software That Install a Private AI → The full step-by-step build for a private AI on your laptop in under an hour. Free, offline, yours.

I sat down with for her Career Pivot Playbooks series to talk about how I got from data science to building AI for productivity. Read the interview here.

Why Giving Employees AI Doesn’t Change How Work Gets Done

Samsung’s rollout is the largest one announced this year. It isn’t the first to hit the same problem.

In May, Marc Benioff said on the All-In podcast that Salesforce will spend $300 million on Anthropic tokens this year, primarily for coding. He described the efficiency gains across service, support, and marketing as ‘unprecedented.’ Salesforce’s support team went from 9,000 people to 5,000 over the past year as AI agents took over more of the work.

Later in the same interview, Benioff said Salesforce still needs a smarter routing system to connect its employees with the models they use. Right now, every request from an employee goes to the same top-tier model.

Benioff said simpler requests should be sent to smaller, cheaper models, and only the complex ones should reach a frontier model like Claude. Salesforce is spending $300 million a year on Anthropic, and the routing system Benioff wants doesn’t yet exist at his company.

Two weeks before this appearance on the All-In podcast, Microsoft started canceling Claude Code licenses for the engineers who build Windows, Microsoft 365, Outlook, Teams, and Surface. The cancellations will finish by the end of June. The official reason was that Microsoft wanted to consolidate its own GitHub Copilot CLI. The reason underneath was cost. Claude Code usage at Microsoft grew faster than the team’s budget could absorb.

Subscribe now

Each of these companies invested heavily in access to AI. None of them built the system that sits between the employee and the model and decides what the model should be asked to do, with what inputs, and scored against what criteria. That system is what’s missing. Without it, $300 million in token spend produces uneven output. Without it, access to frontier models becomes a budget problem. Without it, 260,000 employees with three AI tools each will fall back to the few uses they already know.

What a Scoring System Is

The operators I work with don’t ask the model to make the decision. They write down the criteria a good decision would meet, hand the model those criteria, and ask the model to score the option against them.

Let’s take a simple example: qualifying inbound sales leads. A new lead arrives through a contact form on your website. The sales manager writes down what makes a lead worth a call: job title fit (30 points), industry match (25), company size (20), intent signals in the message (15), tech stack overlap (10).

Each lead gets scored out of 100. The model reads the lead, scores it against the criteria, and returns the number, a tier (hot, warm, cold), and a written explanation of how it arrived at the score.

The sales team trusts the output because they can see exactly how the model arrived at the number, and they can change any single point value within 30 seconds if they disagree. The model applies the same scoring across hundreds of leads a week, faster than any one person could.

The same scoring approach works for almost any decision your team makes regularly:

• Evaluating vendor proposals

• Triaging customer escalations

• Reviewing inbound resumes

• Flagging unusual expenses

• Prioritizing bug reports

A scoring system is what your team would write down if they had unlimited time to document how they make these calls. It’s the criteria, the point values, and the explanation requirement, written down once and applied consistently thereafter. AI handles the volume. Your team keeps control over the criteria.

Samsung’s training program covers how to use ChatGPT, Gemini, and Claude. From everything Samsung has announced publicly, it doesn’t cover how to build the system that turns those tools from a chat window into a decision engine the team can trust.

Subscribe now

Your Next Step

Two days is how long it’ll take you to build the system that Samsung’s 260,000 employees aren’t getting.

This Wednesday at 8:30 AM EST, I’m going live on Substack to build a working scoring system from scratch. Here’s what I’ll be covering:

• How to take the way your team already makes a decision and turn it into a written scoring system

• The structure of a scoring prompt that returns a usable number every time

• How to write the explanation field so your team trusts the output

• What to do when a domain expert on your team disagrees with the model’s score

• Why a short, plain scoring system outperforms a complicated AI agent

Thursday’s article is the full implementation guide of what we built on the live. Every line of the scoring prompt, the workflow setup, the structured output format, and the integration into your existing tools.

By the end of this week, you’ll have a working scoring system running on your team’s own decisions, in your own tools, with your own criteria. You’ll be the person on your team who built it.

Subscribe now

The second is the one a smaller group has already started: what does our team own, and what runs regardless of whether others are having an outage today?

This article is for people moving into the second conversation. The full step-by-step build for a private AI on your laptop in under an hour. Free, offline, yours.

Read more

Some people spent last week building with Fable 5, but you’re the person on your team who already runs AI tools locally, so a frontier model going down didn’t affect you.

For those of you who aren’t this person or that team yet, this was hopefully the wake-up call you needed to start owning more of your AI stack.

The good news is that the tools you need to build with are free.

Work through the builds below in order, or pick whichever sounds closest to what you’re working on this week.

Subscribe now

Run a model on your own machine

n8n. local. A private AI agent on your laptop, no API keys.
How To Fine-Tune and Deploy an LLM Without an ML Engineer. A support assistant grounded in your own docs, running locally.

Use it on a recurring job

I Built an AI Agent That Sends Me My Numbers Every Monday Morning.
Build a RAG System with NotebookLM in Under an Hour.
Build Your Own AI Productivity App with Codex.

Give it memory so it stops losing your context

I Built a Second Brain to Help Me Run My Company.
Every Time You Start a New Chat, You Lose Everything. Here’s How I Fix That.
How We Use Agentic AI to Help Us Make Big Product Design Decisions.
How to Use AI to Do the Executive Function Stuff Your Brain Skips Over.

What this means for how work is changing

AI Agents Are Redesigning Workflows. Your New Skill Is Staying in Motion.
Agentic AI Matters More to ADHD Brains Than Getting Faster Answers.

This week on The Data Letter: Wednesday’s live session is free and open to everyone. We’re going deep on local model selection, hardware sizing, and quantization, so you walk out knowing which model fits your machine.

Three pieces of free software install a private AI on your laptop gets published Thursday, with the full build ready to deploy.

Read it here:

Every build I’ve published over the past few weeks has given you one piece of this. The local agent build put a private model on your laptop with n8n, Ollama, and Docker. No API keys necessary. The Monday metrics build taught the agent to answer from wherever you’re storing your data instead of inventing numbers, and to say ‘I don’t have that’ when the answer wasn’t there. The RAG build in NotebookLM showed how retrieval lets a model answer from your documents. The prompt engineering piece made the case for treating your system prompt as an asset, versioned and tested, not a throwaway string.

Today, those pieces become one thing you own: a support assistant that runs entirely on your own machine, answers from your team’s documents, speaks in your team’s voice, and refuses to make things up. This is the same system I built for Asaura AI. You’ll build your version of it in the next half hour, for free, and when you’re done, the model is yours. No vendor update can take it out from under you, because no vendor is in the loop.

Below the paywall is the complete build, every step from an empty workflow to a working assistant you’ve tested and trust, plus the downloadable files ready to paste in.

Read more