Key takeaways
- Company data is worth $100K to $2M or more per dataset to AI labs and data vendors, and the largest company archive so far, Spirit Airlines', drew bids of $10M and $12.5M.
- Records that let a buyer follow each job from the first request to the final result, including how it turned out, raise the price most; the number of files on its own matters least.
- The value of company data is set by the market: the real number is a buyer's priced offer, and an online estimate is only a starting point.
- An archive license pays once for records you already have; workflow deals pay for new tasks every month, about $900 a task at micro1.
01
Why AI labs and data vendors are buying company data
The short answer
AI is moving from answering questions to doing jobs. To learn a job, a model has to see that job done, and the only record of how work really gets done sits inside companies. That record is what AI labs and data vendors are paying for.
How an AI model learns
An AI model learns by studying a very large number of examples and finding the patterns in them. Those examples are called training data.
What a model can do depends on what its training data contains. A model that has studied millions of written emails learns how to write a good email. To learn how to do a whole job, such as handling a customer's repair from the first phone call to the final invoice, a model needs examples that show that whole job being done, step by step, along with how it turned out.
What the public internet can't teach
For the past several years, most training data came from the public internet. That includes websites, books, news articles, forums and open-source code. It is why tools like ChatGPT and Claude write well and can answer questions on almost any subject.
But the public internet mostly shows finished results. A business publishes its product pages, its reviews and its blog posts. It doesn't publish the day-to-day work that produced them.
Why that gap matters now
The goal of AI companies has moved from AI that answers questions to AI agents. An AI agent is software that carries out a job step by step, the way an employee would.
A model can only learn a job by seeing that job done. The record of it exists only inside companies, in their help desks, team chat, email, shared drives and CRMs (customer relationship management software, such as HubSpot or Salesforce, where sales teams track every customer and deal).
What a business records that the public never sees
Every task a business handles leaves a trail of records behind it, and that trail is far richer than anything the business shows the public. Service calls at an HVAC company, a business that installs and repairs heating and air conditioning, are a good example. Publicly, two calls about the same problem might end in two one-line reviews: two stars from one customer, five stars from the other.
Inside the company, each review sits at the end of its own trail of records, and each record captures a different part of the job. The support ticket is the note the office creates when a customer calls, describing the problem in the customer's own words, such as “the AC is running but not cooling.”
The HVAC company's dispatcher, the person who schedules its technicians, then records which technician was sent to help the customer and which equipment and parts they took along. The technician's own notes, written during or after the visit, describe what they found and what they repaired. Finally, the company's email history shows whether anyone checked back with the customer once the work was done.
The diagram below puts the two sequences side by side, and it shows plainly what caused each review. That is what a model needs to learn a job: not just how it ended, but what led there, so it can repeat what works.
Who is buying
Two kinds of buyers pay for these records. Most business owners deal with a data vendor and never with the lab directly.
- AI labs
- Companies that build and train AI models, such as OpenAI, Anthropic and Google DeepMind.
- Data vendors
- Companies that buy records from businesses, anonymize them (replace names and other personal details with placeholders) and turn them into training material for the labs. micro1, the data vendor we typically refer companies to, is one of them.
How buyers use the records
AI labs and data vendors put company records to work in two ways.
Use 1
Practice environments
The vendor builds a realistic copy of a company's software, such as an inbox, a ticket queue and a CRM, filled with anonymized records. An AI agent practices tasks in that copy thousands of times and is scored on each attempt, until it does the task reliably.
Think of it as a flight simulator for AI. The industry calls these reinforcement learning (RL) environments.
Use 2
Worked examples
A complete record of a task done well, such as a support ticket from the first message to the fix, is used like an answer key. The model studies it to learn what good work looks like at each step.
A ticket that stops before the fix shows the model the problem but not how it was solved.
We explain both in more depth, with examples, in why AI companies want your data.
How much buyers are spending
micro1 has committed $1B to buying enterprise data over the next 12 months. More than 50 companies now sell data and training environments to AI labs.
"I have still never seen a data contract get turned down by a top lab if it's good quality data for budget reasons."
A data-industry insider, on what labs spend
02
How much is company data worth to AI buyers?
The short answer
Company data is worth $100K to $2M or more per dataset to AI labs and data vendors, and the largest company archive so far, Spirit Airlines', drew eight-figure bids. Where a company lands in that range depends on how completely its records show each job from start to finish, how many years and software systems they cover, and how rare the work is.
What "a dataset" means in a deal
A dataset, in a company data deal, is the set of records a company agrees to license in one contract. It usually covers several years of the company's team chat, email, documents, help desk tickets and customer records, exported from the tools the team works in.
A license lets the buyer use the records for agreed purposes while the company keeps ownership. The price is paid for that whole dataset, not per file or per gigabyte, which is why the same amount of data can be priced very differently from one company to the next.
What companies are paid per dataset
Bids for the largest company archive so far, Spirit Airlines'
Per finished task, for ongoing workflow work (micro1)
How micro1 breaks the range down
Data vendors publish the full $100K to $2M+ range themselves: Polyshares, another data vendor, says "Typical deals run $100K to $2M and above." micro1 publishes how its own payments split across that range on its data partnerships page. micro1 says payments typically run from $150K to over $2M, and that compensation is based on "the quality, uniqueness, and value of the data provided".
| micro1 tier | Payment | What it describes |
|---|---|---|
| Qualified enterprise data | $100K–$500K | An established company licensing its everyday operating records |
| Large-scale or multi-team | $500K–$2M | Large operational datasets, or ongoing work across several teams and business functions |
| Highly unique | $2M+ | Proprietary operating data with significant value for the most advanced AI models |
Public deals that show what companies are paid
Very few company data deals are made public, because most are private licensing contracts. The two reported below show buyers bidding millions for an airline's everyday email and chat, and a small company's owner earning six figures a quarter grading AI models.
- Spirit Airlines: eight-figure bids for internal recordsGoogle won the bankrupt airline's employee emails, Microsoft Teams messages, spreadsheets and calendars at auction with a $10 million bid, ahead of a $7.5 million bid from the AI data company Mercor, Reuters reported in August 2026. micro1 then made a late $12.5 million bid, Fortune reported. The court had not approved a sale as of October 2026, and the records are to be de-identified (stripped of names and other personal details) before transfer.
- A 14-person HVAC company: about $1,000 per taskThe owner of a 14-person HVAC company is paid about $1,000 per task to grade AI models, checking and scoring how well they handle his firm's routine work, The Information reported. He is on pace to earn more than $150,000 a quarter. micro1 links the story from its own data partnerships page.
The Spirit auction also shows that several buyers will bid against each other for the same records, since Google, Mercor and micro1 all made offers. Company data has no list price. The final number depends on how many buyers want the records and how complete they are, far more than on how many files there are.
Why company records aren't priced like content deals
The best-known AI data deals are content licenses, where a publisher or website lets an AI lab use what it has already published. Reddit's deal with Google was reported at about $60 million a year, and News Corp's deal with OpenAI was valued by The Wall Street Journal at more than $250 million over five years.
Those deals pay for millions of posts or articles that the public could already read. A company data deal pays for something different, a private record of how a job is done from start to finish, which never appeared on the public internet at all. That is why a 60-person company with well-kept records can be paid six or seven figures for data no one outside the company has ever read. For more of these deals, dated and sourced, see data monetization examples.
03
What sets the price of company data
The short answer
The price of company data is set by how useful the records are for teaching an AI model a job. Records that can be followed from start to finish, show how each job ended, cover several years and systems, come from an unusual line of work and are clearly yours to license are worth the most. Volume on its own matters least.
Buyers pay for what a model can learn
AI labs and data vendors judge a dataset by one question: how much will it improve a model at real work? Every factor that moves the price is a version of that question.
That is why the factors below are about completeness, rarity and permission rather than size. A model learns a job from records that show the whole job, and it learns something new only from work it hasn't already seen.
Connected records are worth more than loose files
Connected records are records that can be linked to each other, so a buyer can follow one piece of work across every tool it passed through. A help desk ticket that leads to a team chat thread, then to an invoice, then to a follow-up email, all for the same customer, is one connected job.
The same four records without a shared case number (the reference number the help desk gives each job) or customer name are four unrelated pieces of text. The diagram below shows the difference.
Outcomes turn a record into a lesson
An outcome is the recorded result of a piece of work: whether the deal closed, whether the ticket was solved, whether the customer came back. A record that shows how the story ended teaches a model which actions lead to good results and which don't.
Sales records in a CRM show why outcomes matter. A lead (a potential customer) that the sales team followed all the way to a won or lost deal shows a model which steps led to a sale. A lead whose history stops halfway shows the steps but not whether they worked, so it is worth far less.
Every factor that moves the price, up or down
Buyers weigh about eight factors when they price a dataset. The diagram below sets the same factors side by side, showing what lowers an offer on the left and what raises it on the right.
What each factor means
- Scarcity
- Work that AI models rarely see is worth more. A niche trade, an unusual piece of industry software or an uncommon process teaches a model something it can't learn anywhere else.
- Clean rights
- Buyers pay more when it is clear the company is allowed to license the records. Customer personal details, other companies' confidential material and the terms of service of the apps the records came from can all limit what can be licensed.
- Exclusivity
- An exclusive license means the company licenses the records to one buyer only. Buyers pay more for data their competitors can't get, but the company gives up the chance to license the same records again.
- Years of history
- More years of records contain more finished jobs, and they show how the business handled busy and slow seasons, new staff and new tools. Several years across several systems is the strongest case.
- Number of systems
- Each tool holds a different part of the same work. Chat shows the discussion, email shows the customer, the CRM shows the deal and the help desk shows the fix, so together they show the whole job.
- Exclusions
- A company can leave out categories such as HR or health records. Excluding what is truly sensitive is normal; excluding whole departments removes the work buyers came for and lowers the offer.
- Volume
- More files raise the price only when they add more complete, connected jobs. A large pile of unrelated documents adds little.
Clean rights and exclusions are explained fully in is it legal to sell company data?, and the terms that come with exclusivity in data licensing vs. selling.
04
How to value company data: data valuation for AI buyers
The short answer
Data valuation is the process of putting a money value on a set of data. Economists group the methods into three families. The cost method asks what the data cost to create, the income method asks what income it earns, and the market method asks what a buyer will pay for it. For company data licensed to AI labs and data vendors, the market method decides the number, because the value is whatever a buyer offers for those specific records.
Why the value of data is hard to pin down
Most companies have never put a number on their data, because it was never something they could sell. Data is created as a by-product of running the business and used internally, so it rarely shows up as an asset with a price on it.
Economists who study the value of data say the same. A 2022 review by Diane Coyle and Annabel Manley at the University of Cambridge's Bennett Institute notes that there are relatively few observable market prices for data, because most data is created and used inside the organisation that holds it.
The three data valuation methods
The same review divides data valuation methods into three broad approaches. Each answers a different question, and each gives a different number for the same records.
| Method | The question it answers | What it tells a company licensing data to AI |
|---|---|---|
| Cost-based | What did it cost to create, store or replace this data? | It gives a cautious minimum, which the review calls a reasonable lower-bound estimate. The staff time spent writing tickets and emails says little about what a buyer will pay. |
| Income-based | What income does the data earn, or what would it cost to license instead? | It becomes useful once a deal exists, because a license fee or monthly workflow pay is the income this method measures. |
| Market-based | What are buyers actually paying for data like this? | This method decides the number, because a priced offer from an AI lab or data vendor is the market value of your records. |
Why the market method wins for AI data
A buyer doesn't care what the records cost the company to produce. The buyer cares how much the records will improve a model, which depends on the factors in the previous section.
That is why the most reliable way to value company data for AI is to ask a buyer. Before that step, a company can narrow its own range with the estimate below, which prices its industry, staff, years of records and location the way a data vendor does.
Data as an asset on the company's own books
Treating data as an asset means knowing what records exist, where they live and who may use them. A company that keeps that inventory is ready for a buyer's first questions and less likely to delete something valuable by accident. How data is recorded for tax or accounting is a question for the company's accountant; the figures here are what AI buyers pay, not the value an accountant would record on the balance sheet.
Estimate
How much is my company's data worth? Estimate your range
Pick your industry, enter your staff and years of records, and choose where your team is based. The range updates as you go. It is a starting point, not an offer.
| Desk employees | Estimated range | Possible refresh pay over 24 months |
|---|---|---|
| 25 | $170K–$700K | +$140K–$560K |
| 50 | $210K–$830K | +$170K–$660K |
| 100 | $270K–$1.09M | +$220K–$870K |
| 250 | $470K–$1.87M | +$380K–$1.5M |
| 500 | $790K–$3.18M | +$630K–$2.54M |
| 1,000 | $1.45M–$5.8M | +$1.16M–$4.64M |
| 2,000 | $2.76M–$11.03M | +$2.21M–$8.82M |
This estimate is our own, based on what data vendors pay for company records. It isn't an offer: the real number comes from the buyer, 1–2 days after an intro call. A full-page version, with the method explained, is at the data value calculator.
05
How to read your estimate, with worked examples
The short answer
The estimate gives a starting range from the things a buyer checks first: industry, desk staff, non-desk staff, years of records and where the team is based. It can't see how connected, complete or rare the records are, and those are what move a real offer above or below the estimate.
What the estimate measures
Desk employees, the people who work at a computer most of the day, count the most, because they write most of the chats, emails, tickets and documents a buyer wants. Each year of records adds steadily to the range. Legal, medical, accounting and finance firms get 30% more, and teams outside the US get less.
Three worked examples
The same estimate gives very different ranges as those inputs change. These examples use the calculator exactly as it runs on this page, for companies based in the US.
| Company | Inputs | Estimate |
|---|---|---|
| HVAC contractor | 30 office staff, 60 technicians and 8 years of records | $190K–$750K |
| The same contractor, longer history | The same staff, with 15 years of records | $240K–$940K |
| Law firm | 40 desk staff, 5 non-desk staff and 10 years of records | $270K–$1.08M |
| Software company | 300 desk staff and 10 years of records | $550K–$2.19M |
The HVAC contractor's range rises by about a quarter without hiring anyone, because seven more years of records add seven more years of finished jobs. Its 60 technicians add less than its 30 office staff do, because people who work away from a computer leave fewer records behind. The law firm has fewer staff than the HVAC contractor but a higher range, because buyers pay more for legal records.
What pushes a real offer above the estimate
A real offer can land well above the estimate when the records are unusually strong. The most common reasons are records that link each job from start to finish, outcomes recorded for most jobs, a niche industry, an exclusive license and clean, documented rights.
A real offer can also land below it. Heavy exclusions, records that are hard to export, unclear rights to customer data and a single disconnected tool all lower the number.
06
One-time archive payment vs. ongoing workflow pay
The short answer
Companies are paid for data in two ways. An archive license pays once for a snapshot of records the company already has. Workflow pay is ongoing. The company's team does its real work using the newest AI models, and the data vendor pays for each finished task, about $900 a task at micro1.
The archive: one payment for what already exists
An archive license covers the records a company has built up over the years, exported once, anonymized and delivered to the data vendor. The data vendor uses the archive as seed data, the realistic starting material for the practice environments described above.
The archive is paid as a single fee, usually a flat amount set in the contract. micro1 says it pays in full within 30 days of accepting the data. This is the payment the $100K to $2M+ range describes.
Workflows: pay for new work, every month
A workflow deal pays a company's team to do its real work using frontier AI tools, the newest models from the major labs. Each finished task becomes a trajectory, a step-by-step record of how a skilled person used the tools to get the job done, and the data vendor pays for each one.
micro1 said in a public webinar that it pays about $900 per finished task, and that a task can take a couple of hours. Its data partnerships page says it licenses qualifying trajectories for $1–2M a year, with the company's approval.
Which one fits a company
Archive
Best for years of records
A company with several years of connected records in several systems gets the largest single payment from an archive license. The team's work doesn't change, and the deal is done once the data is accepted.
Companies that have shut down can license an archive too, if their records survive.
Workflows
Best for skilled teams with time
A company whose staff do specialist work, and can set aside hours for it, earns ongoing income from workflows. The pay depends on how many tasks the team finishes.
A company can do both, licensing its archive first and taking on workflow tasks after.
Other ways to earn from data, from selling insights to building data products, are compared in data monetization models and strategy.
07
Why AI data prices are high now, and whether they will last
The short answer
AI data prices are high because AI labs and data vendors want far more company records than companies are offering. Few owners know their records can be licensed, so buyers pay a premium for what they can get. As more companies come forward, prices are likely to fall.
Demand is far ahead of supply
On the demand side, the buyers have large budgets. micro1 has committed $1B to enterprise data acquisition over the twelve months from October 2026, financed with capital from Citi and Hercules Capital. Other data vendors are competing for the same records.
On the supply side, most business owners don't know their records have a price. A market with many buyers and few sellers pays sellers more.
Licensing is cheaper for buyers than the alternative
AI labs also pay because taking data without permission has become expensive. Anthropic agreed in 2025 to pay $1.5 billion to settle a lawsuit by authors, about $3,000 for each of an estimated 500,000 books, the AP reported.
A signed license with a company that owns its records, anonymized before anyone sees them, carries far less risk of a lawsuit. Buyers value that certainty, which is one reason clean rights raise the price.
Why prices are likely to fall
Prices this high depend on the gap between demand and supply, and that gap should narrow as more companies learn they can license their records. micro1 has said publicly that it is paying a premium now and that commercial conditions will probably change.
Our own view is that prices will level off, possibly within a year. That is a forecast, not a fact. What it means in practice is that a company with years of connected records loses nothing by finding out its number now.
08
Mistakes that lower what company data is worth
The short answer
The costliest mistakes are deleting old tools before exporting them, sending data samples before a lawyer has reviewed the deal, excluding more than needed, and accepting a rushed offer. Each one either destroys records buyers would pay for or weakens the company's position in the deal.
Mistakes before the deal
- Deleting old tools and workspacesCompanies cancel a chat tool or a help desk and let years of records disappear. Export the history with the tool's own admin export before closing any account.
- Assuming more files mean more moneyVolume is the factor that matters least. Spending weeks gathering every file adds less than showing that the records connect.
- Scrubbing the data yourselfDeleting names or whole threads by hand breaks the links between records. The data vendor replaces personal details with consistent placeholders, so the same customer is "Customer 14" in every record and the work still makes sense. See how data is anonymized.
Mistakes during the deal
- Sending a sample before legal reviewA sample shares real company records with the buyer before any contract protects them. Have a lawyer review the agreement first.
- Excluding whole departmentsLeaving out HR or health records is normal. Leaving out sales or support removes the work buyers pay for and lowers the offer.
- Accepting an exploding offerOffers that expire in 48 hours are common in this market. A serious buyer will still be there after your lawyer reads the contract.
- Exporting through app APIsAn API is the connection an app offers other software for pulling data out of it. Slack, HubSpot and Zendesk restrict using data pulled through their APIs for AI training. The download tools built into each app's admin settings are the usual route.
Comparing buyers can also change the number, as the competing bids from Google, Mercor and micro1 for Spirit Airlines' records showed. The options, from AI labs and data vendors to marketplaces and brokers, are compared in where to sell data.
09
How to find out what your company's data is worth
The short answer
To find out what company data is worth, list the systems the team has used and for how many years, use the estimate for a starting range, then apply to a data vendor for a priced offer. With micro1, the priced offer arrives 1–2 days after an intro call, and no confidential data changes hands before a contract is signed.
Start with a list, not a download
A buyer can price a dataset from a description. The first thing to prepare is a list of every tool the team has used, such as chat, email, shared drives, the CRM, the help desk and project tools, with the years each one covers.
That list becomes the data manifest, a summary of what records exist, in which apps, over how many years and roughly how much. It contains nothing confidential. Whether the company qualifies at all is covered in does my company qualify?
From application to payment
Apply 2 minutes
The company fills in a short form with its headcount, years in business and the tools its team uses. The form asks for nothing confidential.
Intro call and data manifest
The company describes its systems, years of records and rough volume. No files change hands.
Priced offer 1–2 days
The data vendor sends a specific price based on the quality, uniqueness and value of the records. The company can turn it down.
Contract and transfer 1–7 days
The company's lawyer reviews the licensing agreement, and the records are then sent through an encrypted transfer. The data vendor anonymizes the records before anyone else sees them.
Payment ~30 days after acceptance
micro1 pays in full within 30 days of accepting the data.
Before you sign
Have a lawyer check the company's right to license the records, the exclusivity terms, how the buyer may use the data and what happens to it when the deal ends. Get every term in writing. This page is not legal or tax advice.
Questions
Frequently asked questions
How much is my company's data worth?
Company data is worth $100K to $2M or more per dataset to AI labs and data vendors, and the largest company archive so far, Spirit Airlines', drew a $10M winning bid from Google and a late $12.5M bid from micro1. Where a company lands depends on how connected its records are, whether they show outcomes, how many years and systems they cover, how rare its industry is and whether its rights to license are clean. A priced offer from a data vendor gives the real number.
How much do AI companies pay for company data?
AI labs and data vendors pay companies $100K to $2M or more to license a dataset of internal records, as a single payment. micro1, one data vendor, breaks that range into three tiers: $100K–$500K (qualified enterprise data), $500K–$2M (large or multi-team datasets) and $2M+ (highly unique data). Ongoing workflow work pays separately, about $900 per finished task at micro1.
What is data valuation?
Data valuation is the process of putting a money value on a set of data. Economists group the methods into cost-based (what the data cost to create), income-based (what income it earns) and market-based (what buyers pay). For company records licensed to AI labs and data vendors, the market-based value matters most, because the price is what a buyer offers for those specific records.
Do I need a lot of data to get paid?
No. Volume is the factor that matters least in pricing company data. A smaller set of records that shows work from start to finish, such as a support ticket linked to its fix or a sales lead followed to a closed deal, is worth more than a large collection of files that don't connect. Several years of records across several systems is the strongest case.
Is the online estimate the same as an offer?
No. The estimate on this page gives a starting range from industry, headcount, years of records and where the team is based. It can't see how connected, complete or rare the records are, which can move a real offer above or below it. A priced offer from the data vendor, 1–2 days after an intro call, is the real number.
Is it better to take a one-time payment or ongoing workflow pay?
An archive license pays once, usually the largest single payment, for records a company already has, and suits companies with years of connected records. Workflow pay is ongoing, about $900 per finished task at micro1, and suits teams with specialist skills and time to spare. A company can do both: license its archive first, then take on workflow tasks.
Will prices for company data stay this high?
Probably not. Prices for company data are high because AI labs and data vendors want far more records than companies are offering. micro1 has said it is paying a premium and that commercial conditions will probably change. Our own view is that prices will level off as more companies come forward, possibly within a year.
Does applying through The Corner lower my payment?
No. micro1 pays The Corner a referral fee out of its own money, and its published terms don't take that fee out of the company's payment. A company that wants this confirmed can ask on its intro call and get it in writing.
Bottom line
Prices are high now, and may not stay that way
We expect prices to level off as more companies start selling, possibly within a year. If your company has years of connected records, now is the time to find out what they're worth. micro1, the data vendor we typically refer companies to, prices the offer and handles everything after you apply.
Not legal, tax or financial advice. Figures come from micro1's published pages and public reporting as of November 2026.