Before AI, one picture of the business

AI, applied · 01 · September 29, 2026 · 7 min read

What a large AI vendor sells first

In the second quarter of 2026, Palantir Technologies reported that its U.S. commercial revenue "grew 149% year-over-year and 28% quarter-over-quarter to $764 million." Total revenue for the quarter, which ended June 30, 2026, "grew 93% year-over-year and 19% quarter-over-quarter to $1.935 billion." The results were released on August 3, 2026 and filed with the SEC as an exhibit to a Form 8-K.

Palantir is usually discussed as an AI company. The more useful detail for an owner is what sits underneath the AI. The company's own documentation describes it before it describes any model.

The same page says that "in many settings, the Ontology serves as a digital twin of the organization." Another page in the documentation calls the Ontology "the central system" that lets customers use AI safely in their enterprises. Strip away the vocabulary and the order is plain. First, one maintained picture of the whole company. Then the AI, working from that picture.

This is not an argument for buying anything Palantir makes. It is a reference point. A company that sells AI to large organizations puts the picture of the business first, and that order holds at any size.

Things, links and changes

The documentation describes the picture in three parts. An object type "defines an entity or event in an organization." A link type "defines the relationship between two object types." An action type "defines how an object type can be modified."

In plain terms, objects are the things a business deals with: customers, jobs, quotes, invoices, trucks, parts. Links say which things belong together: this invoice is for this job, which is for this customer, at this address. Actions are the changes people make: book the job, send the invoice, mark it paid. The documentation compares the whole arrangement to a spreadsheet: an object is like a row, a property is like a column, and a link is like a join between two tables.

None of this is exotic. It is the discipline of deciding what counts as one job, one customer and one payment, and then making every system agree on it. The software is built on top of that decision. It does not make the decision for anyone.

Scattered pieces, scattered answers

An AI tool can only reason about what it is shown. Give it the calendar and it knows when. Give it the invoices and it knows how much. Ask which jobs lost money last quarter and it needs both, matched correctly, plus the change order that was agreed over the phone. If the pieces live in different places, under different spellings of the same customer's name, the answer is a guess formatted as a report.

The federal government's AI risk framework names a version of this problem. Among the risks that NIST lists as new or increased compared with traditional software is that "the data used for building an AI system may not be a true or appropriate representation of the context or intended use of the AI system." The framework also warns that datasets "may become stale or outdated relative to deployment context." NIST wrote those lines about building AI systems. The same logic applies to a small office handing a tool a partial view of its own work.

The same problem at five people

A five-person business rarely has anything it would call a data problem. It has an estimate in one app, the schedule in another, invoices in a third, and the customer history in someone's head. Take a small remodeling crew as an illustration. The quote is built in one tool. The job goes on a shared calendar. The invoice goes out from the accounting software. The note that the customer wants the side gate latched because of the dog lives in a text thread on the owner's phone. Each tool works. None of them knows the others exist.

The cost shows up when someone asks a question that crosses the lines. What does a bathroom job actually earn after the second trip for a missing part? Which customers from two years ago are due for the next project? Why did last month feel busy and pay thin? The answers exist, spread across four places. Assembling them is a Saturday with a spreadsheet, so it usually does not happen.

That is the problem the Ontology documentation describes, at a smaller size. The fix is smaller too. It does not require a platform. It requires the same decision about what one job is.

Where small firms stand in the numbers

The Census Bureau tracks AI use through its Business Trends and Outlook Survey, which it describes as providing "a biweekly, nationally representative view of AI implementation." In a story published May 26, 2026, covering data collected from December 14, 2025 to May 3, 2026, the Bureau reported that overall AI usage "hovered between 17% and 20%," and that "between 20% and 23% of businesses expected to be using it in the next six months." As of May 3, 2026, the national rate was 19.8%.

The split by size matters most here. In the Bureau's words: "Between December 2025 and May 2026, AI use increased among firms with at least 20 employees but didn’t change significantly among firms with fewer than 20 employees." And: "Less than 20% of firms with four or fewer employees reported using AI."

A Census working paper on the survey's AI supplement, covering November 2025 to January 2026, adds the forward view. For firms with 1 to 4 employees, expected use "sits at roughly 21%, only slightly exceeding their current use rate of 18%." Firms with 250 or more employees projected 40% expected use against a 31% current baseline.

Among firms not planning to use AI, the most common reason given was that "AI is not applicable to the business (65% of firms, firm-weighted)." For many businesses that is a fair answer. The survey's list of possible reasons also includes "Lack of required data" as its own option.

The groundwork nobody counts

Economists have a name for the lag between buying a new technology and getting paid for it. Erik Brynjolfsson, Daniel Rock and Chad Syverson wrote in 2021 that "general purpose technologies (GPTs) like AI enable and require significant complementary investments," and that "these investments are often intangible." They called the resulting pattern the Productivity J-curve: returns look low while the groundwork is being laid, and arrive later.

The Census working paper found little of that groundwork so far. It reports that "over half of AI-using businesses (64%) report no institutional adjustments." Changes to data management and storage practices, along with complementary capital investments, "take place in a smaller fraction of firms (in the 7-8% range)." The same paper found "a robust positive correlation between firm commercial performance and the breadth of AI integration." A correlation is not proof that one causes the other. It is consistent with the idea that the value sits in the plumbing, not the tool.

The order that works

The sequence that holds up, at five people or fifty thousand, has three steps.

• List every place the work lives. Every app, spreadsheet, inbox, paper folder and phone, including what only one person knows. Write down what each one is the authority for: the calendar owns dates, the accounting software owns money.
• Connect them so one job is one record. Pick one name and one number for each customer and each job, and make every tool use it. When the estimate, the schedule, the invoice and the notes all point at the same job, the business has its picture.
• Then automate what repeats. Reminders, follow-ups, reorders, the monthly look at which jobs made money. With one record per job these become simple rules instead of weekly effort, and any AI tool asked about the business has the whole picture to read.

The first two steps are unglamorous. Often they can be done inside the software already in use, with settings and a naming habit rather than a new purchase. They are also the part that lasts. The tools will change several times over the next few years. A clear account of what the business is made of carries over to whichever tool comes next.

Palantir's documentation puts the picture before the AI. A five-person shop can draw its own version on a whiteboard this week, and it will be worth more than the next app.