Are AI Companies Buying Data From Small Businesses?

Are AI companies buying data from small businesses?

Yes, from some. AI developers now pay for records of how real work gets done: a technician's notes on a repair, an accountant's review comments, a dispatcher's decisions on a bad day. Many of those records sit in firms with a few dozen people, because that is how most trades and professional services are organised.

Most of the money in the public record still goes to large archives. Reddit disclosed data licensing contracts worth $203 million, and News Corp's deal is reported at more than $250 million (Quartz). Small sellers appear too. A founder winding down a startup that was "just a few years old" netted roughly $25,000 for its code base, with about $15,000 more pending for its workspace data (MIT Technology Review), and Forbes reported Slack and email deals for closed startups, "with some deals reportedly reaching six figures" (Newser).

The honest answer for most of the 34,752,434 small businesses in the United States (SBA Office of Advocacy) is no: they do not hold records an AI developer would pay for. This page is about the minority that do, what they look like at 10 to 50 people, and what the rest can do instead. Which record types qualify in general is covered in what company data you can sell to AI, and the per-head numbers are in company data value per employee.

What small businesses have been paid: examples from 10 to 50 people

These are the deals Origin helped place with companies of 10 to 50 people, anonymised. Each one licensed a de-identified copy of records the business already kept for its own work. None of them created data for the sale.

• The pattern is the same in every row: trained staff writing down decisions, and the outcome recorded in the same system.
• Two firms with the same headcount can differ by several times, because years of history, record depth and regulation move the price more than size does.
• Full case studies, including larger companies, are on the case studies page.

What a 10 to 50 person firm realistically holds

Owners usually underestimate volume. The figures below are Origin's rough arithmetic for typical firms, not quotes. Use them to check whether your own history is in the same range before you ask anyone for a number.

• A 12-technician HVAC or plumbing contractor that runs four jobs per technician per day writes roughly 12,000 work orders a year. Five years in a field service system is about 60,000 jobs with notes, photos and invoices. See exporting ServiceTitan data.
• A 15-person accounting and tax firm with 300 clients produces several thousand sets of workpapers, review notes and close checklists over five years, mostly in a document or practice management system and in the books themselves. See exporting QuickBooks data.
• A 20-person specialty contractor finishing 40 projects a year has 200 closed projects in five years, each with RFIs, change orders, daily logs and photos that tie the plan to what was built.
• A 25-person trucking firm with 20 trucks generates continuous ELD logs, dispatch records and maintenance histories; three years is hundreds of thousands of log entries linked to loads and repairs.
• A 30-person software or IT services firm holds years of commits, tickets, incident notes and chat. Email and shared documents sit beside them; see exporting Google Workspace data.

What disqualifies most very small firms

There is no legal minimum. The SBA counts any business with fewer than 500 employees as small (SBA Office of Advocacy). The practical floors come from what a buyer needs: enough cases, written by enough people, over enough years. Below about ten people, most firms miss at least one of them.

• Too few writers. When one or two people do all the work, the records show one way of working, and much of the reasoning never gets written down.
• Records that live in heads, phones and paper. Text messages to customers, a paper job book or unscanned files do not add up to a dataset.
• Short history. IRS guidance says to "keep records for 3 years" in the ordinary case, and employment tax records for at least 4 years (IRS). A firm that deletes on that schedule rarely has the five or more years that make an archive stand out.
• Records in someone else's system: a franchisor's platform, a client's portal, or files a contract says you hold in confidence.
• The wrong kind of data. Customer lists, point-of-sale totals and website analytics show what sold, not how skilled work was done, and contact lists are personal information that model builders do not want.
• Records full of health, financial or children's data that cannot be reliably de-identified.

How a small business data deal usually pays

One-time license fees are common: a payment on signing or delivery for a prepared, de-identified copy of the records. They are not the only structure. Some deals pay per record delivered; one public filing shows "$100 per delivered record, invoiced monthly" (Tempus Form 8-K). Others pay fixed amounts each year of a multi-year term, as in Reddit's two-to-three-year contracts (Quartz), or a share of revenue, whose value "depends on a definition of revenue the seller cannot easily verify" (Latsoudis + Associates). Yearly updates of new records can add later payments.

Whatever the structure, it does not replace the business. For an owner thinking about retirement or a sale, the records can be a separate asset to realise before or alongside it; see what is my business worth.

Expect work before money. Personal details must come out of every record first. California law treats data as deidentified only if, among other steps, the business "contractually obligates any recipients of the information to comply" with the rule (Cal. Civ. Code 1798.140). In the largest recent court sale, the buyer said personal information would be scrubbed "by a third party before receipt" (Axios). The method is in how to de-identify company data before licensing.

System access: what is normal and what is a red flag

Some deals deliver an export; others use a connection to the system that holds the records, because exports from field service and accounting tools can lose notes and attachments. A connection is not a warning sign by itself. How it is scoped, and when it is granted, decides whether it is safe.

Normal looks like this: a license is signed first; the connection is read-only; it covers only the named systems, record types and date range in the license; it uses its own credentials that you can revoke at any time; and the data it reads is de-identified before the buyer receives it. A de-identification provider sets the same order for the buyer's side: a written commitment, before transfer, "to hold the data for internal, access-controlled use" (Tonic.ai).

The red flags are unscoped access (an admin password, a staff login, or "all company data"), any access requested before a license is signed, and upload portals or price quotes offered before anyone has seen a description of your records. Slow down and get the terms in writing; the data licensing due diligence checklist lists the checks.

If your records do not qualify: the alternatives

Recording programs pay people, not companies. AI and robotics developers buy video of hands-on work, with "cameras capturing the hands-on details of preparing food, stocking shelves and repairing equipment", and gig platforms offer contributors "paid video-recording tasks to help train robots" (Fortune). The BBC describes "hundreds of thousands of workers across the world paid to record their activities", including "small business work like pottery, garment-making, embroidery, candle-making" (BBC). The pay is per hour of footage: one Los Angeles program "pays $80 for two hours of footage" (Los Angeles Times). For an owner it is side income for staff who opt in, with consent and privacy questions of its own, not a sale of the company's records.

Staying sellable is the other route. A firm that is too young or too thin today can qualify in a few years if it keeps its work in one system, stops routine deletion and writes fuller notes. The guide to keeping company data sellable covers the retention settings to change now.

First steps for a small business owner

You do not need to export anything to find out whether your records qualify.

• Write down which systems hold your work records, how many years they cover, roughly how many jobs, clients, projects or tickets there are, and how many people write them.
• Note anything that belongs to clients or a franchisor, or carries health, financial or children's data.
• Stop routine deletion of old records until you know their value.
• Compare your answers with the examples above and what companies have been paid.
• If an offer has already arrived, read how to review a data licensing offer before you reply.

The bottom line

AI companies do buy data from small businesses, but only from those whose records show skilled work over several years and whose owners can license them cleanly. At 10 to 50 people that is common in trades, accounting, construction, logistics and software. Below that, most firms are better served by keeping their records well now than by looking for a buyer today.

This page is general information to take to your lawyer and accountant. It is not legal, tax or investment advice.

Origin Data Partners is an independent data partner that helps companies get their data bought. A buyer may pay Origin when a deal closes.

Not sure whether a buyer or an offer is sound? Origin checks who the buyer is, what the license allows and how the price compares with published deals before anything is signed. Get a free estimate and a second opinion.