Article

AI product companies: get paid for your records

By TrustPast

Updated November 7, 2026

AI product companies are young as companies go, and most of them think of their history as too short to matter. It usually isn’t. A team of 40 that has shipped a product for four years has written down nearly every decision it made: in the tracker, in the chat, in the experiment log, in the incident reviews after the bad weeks. This page is for AI product companies of that kind, with 30 or more people now or at their peak.

The thing such a company tends to overlook is that its operating record is a different thing from its product and its customers’ data. A buyer licenses records like these to train AI models, and what it pays for is how the company worked: how a feature went from a customer complaint to a release, how a quality problem was found and fixed, how pricing was argued. The plain version of that idea, worked through four trades, is on what data buyers want from a trade business. It applies here with very little change.

Below: the records, the buyer’s own words on what it wants, the things that stay out, the price, and the programme.

The records an AI product company keeps

These are our examples, not a buyer’s list, and the choice of what is licensed is micro1’s.

The tracker. Every ticket, from “responses slow for enterprise accounts after 5 p.m.” to “add audit export.” The description, the estimate, the comments where two engineers disagreed about the cause, the date it closed, the link to the release. A product team of 25 can write tens of thousands of tickets in four years. Each records a small decision.

The experiment log. Fast-moving product teams test constantly: a new ranking rule, a new onboarding step, a new default. A good log says what was tried, why, what the team expected, what happened, and what it decided. The experiments that failed are as much a part of the record as the ones that shipped, because each one is a guess that met evidence.

The quality review sheets. Teams that ship software whose output varies keep some way of judging it: test sets, scoring rubrics, reviewer notes on a sample of outputs each week, the threshold a release had to clear. The rubric’s edit history shows what the team decided “good” meant, and how that changed after customers complained.

The incident reviews. The day the product gave a customer something badly wrong. The timeline, the first theory, the real cause, the fix, the change to the release process afterwards. Read in order, the reviews show a team’s first theory losing to the evidence, again and again.

The chat. The engineering channels, the launch channel, the thread with a large customer’s team during a hard rollout, the late-night debate about whether to roll back. Slack or Teams is where the quick calls got made. The chat archives page explains which channels show how the work was done and which are only chatter.

The roadmap and its revisions. The quarterly plan, the cuts made at mid-quarter, the notes on why one bet was dropped for another. Put the plan next to what shipped and you see how the company decided under pressure. Our project history page goes through why the plan beside the outcome is the valuable part.

The sales and pricing record. The CRM of trials, pilots and contracts, with the notes on why a pilot converted or didn’t. The pricing page’s history, and the spreadsheet that argued for each change. For a young product, pricing is decided again every few months, and each decision is written down.

The onboarding plans. Every large customer gets some version of one: the kickoff notes, the integration checklist, the list of what broke in week one and who fixed it. A product company that has onboarded 80 customers holds 80 small projects, each with a plan and an outcome. The later ones are shorter, and the difference between the first plan and the fortieth is what the team learned.

The architecture decisions. Some teams keep a written log of technical choices: the option picked, the two rejected, the reason, the date. Teams that don’t still have the same debates in pull request comments and design documents. Either way, a choice that was reversed a year later, with the reason written down both times, is reasoning a buyer can read.

The support queue and the knowledge base. The questions customers asked, the answers the team settled on, the articles written so the same question would stop coming. A support ticket that went from “it’s broken” to “it was a permissions setting” to a new help article is a full problem-to-answer cycle.

How a complaint became a release is the story these records tell together. A large customer reports that results got worse after an update. Support logs it and escalates. The chat thread argues about whether it is real; someone pulls examples. The quality review sheet confirms a drop on one kind of request. An experiment tests a fix. It helps that case and hurts another, so a second experiment follows. The release goes out two weeks later with a note in the change log, and the rubric gains a new check so it doesn’t happen again. Each system holds one piece: the ticket, the argument, the test, the score. Only together do they show the decision.

See your estimate

What counts, in the buyer’s own words

There is no micro1 page written for this category. Its general pages are specific enough to read against a product team’s systems.

Start with the eligibility question in micro1’s FAQ. The answer: “We’re looking for established companies with 30+ full-time employees, mature business operations, and substantial historical operational data. This can include internal communications, SOPs, project histories, CRM records, support tickets, and other business documentation.” Two conditions follow it: “Data should be primarily in English. We’re currently prioritizing US companies, followed by other Western markets.” A product team that writes its tickets in English and sells in the US meets both.

What micro1 is after is the work, not the output. Its FAQ: “We are looking for data that reflects how work is actually performed within an organization.” A little further down: “The most valuable datasets capture real operational expertise, decision-making processes, and repeatable workflows.” An experiment log is a decision-making process written down.

The partnerships page also sorts its examples into functions, and three of them read like a product company’s org chart. Software and engineering: “Source code documentation, engineering workflows, bug tracking, product planning, and technical knowledge bases.” Customer support: “Support documentation, ticket workflows, QA processes, customer operations, and knowledge management.” Sales and CRM: “CRM data, sales processes, playbooks, customer lifecycle documentation, and revenue operations.” The tracker, the help desk and the CRM map onto those three, one each.

Young companies worry about size. micro1’s FAQ sets no floor: “There’s no fixed minimum number of files, documents, or terabytes.” What it weighs instead is “the historical depth, breadth, quality, and relevance of the available data,” and it adds that “Companies with several years of records across multiple business systems are generally a stronger fit.” Breadth is where a product company is usually strong, because it runs a separate system for nearly every function.

micro1’s announcement of 9 October 2026, a commitment to spend $1 billion over the next 12 months on licensing enterprise operational data, explains the interest: “Real business operations involve incomplete information, competing priorities and exceptions that require judgment.” An experiment that helped one customer and hurt another, and the call the team made about it, is such an exception.

What is never included

Our page on what is never included lists what stays out for every company. A product company has a few places of its own to check.

What customers send through the product. The documents, messages and files a customer puts into your product, and what the product returns, belong to the customer under your contract. As our own advice, treat them as out of scope, and have counsel read your customer terms before you describe anything. The record of how your team worked is a separate thing from your customers’ content.

Data you licensed from someone else. Datasets bought or licensed from third parties, and anything collected under someone else’s terms, are not yours to license again. As our own advice, list them separately so nobody counts them by mistake.

Personal details inside the operating record. Customer contacts in the CRM, user emails in support tickets, staff names in the chat. micro1 says on its data partnerships page that before any participation the two sides confirm “De-identified before sharing,” meaning “Personal and identifying information is removed or transformed before your data is shared with third parties.”

Security material. Keys, credentials, customer environment details, penetration test findings. As our own advice, these stay out entirely, and you say so on the first call.

Staff files and payment details. Payroll, contracts, reviews, card and bank details. Not how the product was built.

How de-identification works

The partnerships page lists four commitments, confirmed together before anything starts. Data goes to “U.S. and allied buyers only.” There is “Data segregation.” Data is “De-identified before sharing.” And there are “Defined retention periods”: “Raw data is deleted according to an agreed schedule.”

For a company that sells security to its own customers, the handling detail matters. micro1’s FAQ on protection says “Personally identifiable information (PII) is removed when relevant, access is restricted to a limited number of authorized team members, and customer data is processed through isolated pipelines that prevent it from being commingled with other engagements.” It also says “employees with access to data systems use company-managed devices and follow established security controls and training requirements.” Expect your own security team to ask for those terms in writing.

Two more steps come after the transfer, by micro1’s account. Some records are reworked: “certain datasets may undergo synthetic rewrites that preserve workflow structure while reducing direct ties to original records.” And you see a sample first: “companies are given the opportunity to review representative samples before prepared data is used.” When the work ends, “Upon request or at the end of an engagement, data is deleted according to the agreed retention and deletion terms.”

Engineers tend to want to clean things before they send them. micro1 says that is not needed: “No, micro1 has processes for de-identifying and transforming sensitive information.” The general steps, for any company, are on our page about how de-identification works.

What decides the price

The money comes in three published bands. The first, “$100K-500K” (indicative, not an offer), is labelled “Qualified Enterprise Data Partnerships.” The second, “$500K-3M” (indicative, not an offer), is for “Large-scale operational datasets or ongoing participation across multiple teams and business functions.” The top band, “$3M+” (indicative, not an offer), is for “Highly unique, proprietary operational data.”

Which band a company reaches is decided case by case. micro1’s FAQ names the inputs: “Factors considered include dataset size and volume, workflow complexity, data quality, domain expertise represented, and the uniqueness of the operational knowledge contained within the dataset. Each partnership is evaluated individually.” The calculator on its payout page carries a caveat to the same effect: “Indicative estimate only. Final payout is determined based on data sources, volume, quality, access terms, and due diligence.”

Here is our reading of those factors for a product company, not micro1’s. Workflow complexity is the release and review process. Domain expertise is the reasoning in the experiment log and the quality sheets. Uniqueness is the part only your team could have written, because nobody else ran your experiments. Due diligence is where the customer-content question comes back, and in our reading, records that are plainly the company’s own should be the simpler ones to approve. Our page on what decides the price goes through each factor.

The timing is published too. Step four of micro1’s process reads: “We pay in full within 30 days of acceptance.” The FAQ breaks that into two parts: “Once your company transfers its data, micro1 has 30 days to review and approve it. After the data is approved, we’ll ask your company to submit an invoice, and payment will be made within 30 days of invoicing.” Plan the invoice into your own month-end.

Which programme fits

The headcount decides the programme. A company with 30 or more people, counted now or at its peak, goes to micro1, and so does a company of any size based outside the US. On micro1’s referral page, the Enterprise Data Partnership card lists “30+ employees,” “US and Western markets,” “English” and “All industries.” Our estimate asks for your headcount and names the programme that fits; the programme page for 30 or more has the detail.

Plenty of young product companies close, and the record survives them. The same referral page has a second card, the Defunct Company Data Partnership, with one line under it: “We are acquiring data from companies that are no longer operating.” That card asks for “30+ employees,” “10+ Countries,” “English” and “All industries,” and lists company compensation of “$100K-$3M+ for approved data packages” (indicative, not an offer). Who can sign for a closed company, and what its customer contracts said at the end, come first; our page for closed companies goes through them.

Ownership stays put either way. In micro1’s words: “Companies retain ownership of their underlying data.”

Three other kinds of company keep records shaped much like these. Game studios run the same tracker, playtest and post-mortem cycle, with players in place of customers. Video production studios keep version after version of the work with the client’s notes on each, much as a product team keeps release notes. And consumer brands and photo studios test what customers want and keep the pricing sheets that argue each change.

TrustPast

Questions and answers

Does the buyer want our customers’ data?

No. What a customer sends through your product is the customer’s, under your contract with it. micro1 says personal and identifying information is removed or transformed before data is shared with third parties; as our own advice, keep customer content out of scope and ask counsel.

Which of our records count?

micro1 lists internal communications, SOPs, project histories, CRM records, support tickets and other business documentation. Under software and engineering, it names source code documentation, engineering workflows, bug tracking, product planning and technical knowledge bases.

Is there a minimum amount of data?

micro1 says there’s no fixed minimum number of files, documents or terabytes, and that companies with several years of records across multiple business systems are generally a stronger fit.

Do we need 30 employees?

Thirty is micro1’s published line: its FAQ asks for “30+ full-time employees,” along with mature business operations and substantial historical operational data. Count everyone, now and at the peak; our estimate asks for that number and names the programme that fits.

How much does micro1 pay?

micro1’s published bands start at $100K-500K and run past $3M+ for the most unique data (indicative, not an offer). micro1 says it evaluates each partnership on its own.

Do we keep ownership?

Yes. micro1’s FAQ answers the question in one line: “Companies retain ownership of their underlying data.”