How Much Is Your Company's Data Worth to AI Companies?

The honest answer is a range. Buyers of company records publish prices that run from $10K for a closed startup to $5M for a large operating company, and the middle half of the public offers for one company's records sits at $174K to $693K per company. The few sales that actually closed and were made public sit lower, with a median of $28.3K per company, while the largest single archive on record, bankrupt Spirit Airlines' corporate records, drew a $10 million winning bid from Google in a court-supervised auction. Where your company lands inside that span depends on three things: what you hold, how linked and how long your records run, and whether the rights are clean.

This guide is for the founder, CEO, CFO or counsel of a company that keeps working records (email, chat, tickets, documents, CRM, finance) and has started getting calls, or wondering whether it should. It lays out what buyers publish, what has actually been paid, how a valuation of company records is built from the market's own records, what raises and lowers a price, why quoted and realised prices differ, and three worked examples you can compare yourself against. Every number links to the record it comes from.

A word on who is writing. AIData.Broker keeps the price board, the deal ledger and the buyer board this guide draws on, and it also acts as the agent of holders who license or sell their data. That makes us a party in this market, so check any figure you plan to rely on against its source.

This guide is research from public sources, not legal, tax or investment advice.

Contents

1. The short answer, by kind of company
2. What buyers publish
3. What has actually been paid
4. How a valuation is built
5. What moves value up
6. What buyers turn away
7. Why quoted and realised prices differ
8. Three worked examples
9. Questions people ask
10. What this means for you

The short answer, by kind of company

Company records are the working history of a business: its procedures and runbooks, its tickets and project histories, its documents, its customer records, its email and chat. AI companies buy them because a model that is meant to do office work learns best by practising tasks inside a realistic copy of an office, and a copy is only realistic if real records fill it. That is why the main buyers are the firms that build those training environments for AI labs ( micro1, Mercor, Handshake AI, Turing, Scale AI and others), plus a few labs buying direct, such as Google in the Spirit auction.

Price is set per company, not per gigabyte or per message. There is no list price, and the same archive can draw very different numbers from different buyers. Still, the public record is now large enough to show bands. The table below groups them by the kind of company, using the buyers' own published tiers and the few disclosed deals. Read each row as a band that a real company of that kind has been offered, estimated at or paid, not as a promise.

Two things in that table matter more than any single number. First, size moves the price, but much less than people expect: the market's own estimators show that doubling a company's head count multiplies the figure by only about 1.47 times, and doubling its years of records by about 1.31 times. Second, published tiers sit above what closes. The middle of the buyers' offers is $316K per company, while the median of the sales on record is $28.3K per company. Both facts come up again below, because they explain most of the surprise holders feel when a real offer arrives.

What buyers publish

Around 24 buyers on the board say on their own pages that they license or buy a company's records, and many of the buyers we track publish some figure for what they pay. Those figures come in three shapes: a range or tier (from one sum to another, often with a plus sign at the top), a floor or a ceiling ("commonly ... or more", "up to ..."), and an estimator, a calculator on the buyer's site that returns a range once you type in your head count and years. All three are invitations to talk, not offers: none of them has seen your data.

The table gathers what the main buyers and intermediaries publish, each linked to its record and the page it was read from. The buyers' pages change often, so the date matters; the buyer board keeps the current reading, and the guide to company data buyer programs compared sets their licence terms side by side.

Read that table with one fact in mind: the ranges overlap so heavily that they say less about price than about who each buyer wants. Replay and SimpleClosure publish low floors because they take small and closed companies. micro1 and Handshake publish high floors because their programs are built for operating companies above a size: micro1's program states " 30+ employees US Prioritized Primarily English", and Nyne's offer-beating promise applies to US operating companies with at least 20 full-time W-2 employees and three years in business. A company outside a buyer's rules is not a cheap version of that buyer's deal; it is not that buyer's deal at all.

The estimators, and the formula three of them share

The calculators are where most holders first see a number, so it is worth knowing how they work. micro1's, Handshake's and Nyne's estimators run the same formula: a fixed $140,000, plus $2,333 per person at the company's largest, plus $11,667 per year of records, times a location factor. For a US company of 100 employees and 10 years, micro1 shows $379,290 to $888,088 and Handshake $383,083 - $896,969. The location factor is steep: the same company in Europe gets $284,468 to $666,066 from micro1 and elsewhere $113,787 to $266,426.

Other estimators use other formulas and land far away from those. For a US company of about 100 people and about 10 years, the public estimators read on 16 October 2026 showed Grepped at $37.5K to $54K, Ooak at $99k to $139k, micro1 at $379,290 to $888,088 and Troveo at $800K to $1M: the low ends differ by about 21 times. Mode's estimator, read on 17 October, shows $290K to $570K for 100 desk employees and 10 years in the US, and says it is " Indicative only; subject to a full review of your internal tools and data volume".

The lesson is not that one calculator is right. Each is a buyer's or broker's opening position, tuned to what it wants and labelled as indicative. The only estimate worth anchoring on is one built from every public price at once, moved to your company's facts, which is what the valuation method below does.

What has actually been paid

Offers and estimates are plentiful. Completed sales with a published price are rare, because buyers and holders sign confidentiality clauses and most deals are private contracts between private companies. The public record of what was actually paid for a company's records comes from four kinds of source: bankruptcy dockets, which are public by law; press reports in which a party described a deal; intermediaries' own accounts of deals they closed; and buyers' statements about how many deals they did, usually without prices.

Spirit Airlines: the one fully public auction

The best-documented sale is the court-supervised auction of bankrupt Spirit Airlines' de-identified corporate records: email, Teams, OneDrive, SharePoint, CRM, revenue management and workflow data, an archive of about 100 million emails and 500,000,000 Teams items. Google opened at $5M. Three qualified bidders then took the price to Google's winning bid of $10M in about two and a half hours, with Mercor's $7.5M standing as the backup. Twenty days after the auction, micro1 filed a $12.5M proposal, a 25 percent premium over the winner. From the first bid to the last, competition moved the price 2.5x.

Three details of that sale matter for anyone pricing records, and none of them is the headline number. First, payment waits for certified de-identification: the agreement pays the price only on delivery of de-identified data, certified by an independent agent before the first and every later transfer. Second, rights carve-outs did not cut the price: after the auction, Spirit and Google agreed that passenger transaction, refund, inflight purchase and WiFi sales systems would not be transferred in any form, and the deal is still described at the same price. Third, employee data is what slows it: the sale hearing has been adjourned five times and now sits at 6 November 2026, while the parties negotiate with objectors over employees' personal information, and members of Congress asked the companies to halt the transfer of former-employee data. As of this writing the sale is agreed, not approved. The guide to selling customer data in bankruptcy follows the case in detail.

Closed startups: many small deals

The other well-documented end of the market is the closed startup. SimpleClosure, which handles wind-downs, told Forbes its test marketplace had processed nearly 100 deals and recovered about $1M for founders, which works out to about $10K each against a quoted range of $10K to $100K. By October 2026 its asset marketplace had handled over 150 deals, with 47% of its deals now including code or data sales. One founder described his own result to MIT Technology Review: roughly $25,000 net for his code base and a pending amount, likely around $15,000, for the workspace data in Google, Notion, Figma and Slack.

Larger closed companies fetch more. The transcription company cielo24, closed after thirteen years of Slack messages, Jira tickets, emails and Google Drive files, sold its data assets for mid-six figures. Turing told The Information it had bought five to ten failed-startup codebases, paying an average in the tens of thousands of dollars each. And one startup whose team was hired away received multiple offers, up to $300K, for its code and communications, and declined them all.

What intermediaries say they closed

Several intermediaries now publish lists of deals they say they completed. They are useful as a picture of the kinds of company that sell, but nobody independent has confirmed them, and the intermediaries are describing their own success, so they count as claims. Red Deer Data lists $500K for a logistics firm's dispatch, messages and procedures, $280K for an accounting group's workpapers and communications, and $1.2M for an industrial distributor's records. Polyshares lists perpetual exclusive licences at $300,000 for a general contractor, $450,000 for a roofing company, $275,000 for a solar company and $110,000 for a marketing agency. Shfa.ai says it paid $943,000 to one healthcare customer, and Corpus says its largest licence to date was $2M.

Buyers' statements add volume without prices. micro1 told the Spirit court it had completed more than 50 data transactions in the preceding 45 days, without naming the holders or prices, and it has committed to spend $1 billion over the next 12 months acquiring and licensing enterprise operational data. A buyer-side network told its members in October 2026 that one lab buyer had just bought five corpuses at once, including a software startup's email, Drive and Slack over 3+ years, 65 private repositories and 13.7K tickets, a creator marketplace's CRM, sales calls, support tickets, Drive files and Slack, and an industrial parts firm's 11 years of CRM, accounting, inventory and orders. No price was given. What those lots show is the shape of what sells: several linked systems, years of history, real work.

The realised side, in one table

Put the disclosed outcomes side by side and two clusters appear: closed startups in the tens of thousands, and operating or large closed companies in the hundreds of thousands, with one airline far above. The weighted average of what a small company's data sold or was priced at, across the small-company records in the ledger, is $26,000.

How a valuation is built

If every public figure is an opening position and real sales are scarce, how can anyone put a defensible range on one company's records? The method behind the valuation on this site answers that by refusing to start from any single buyer's formula. It starts from all of the market's records at once: the 109 records in the price ledger of what a whole company's records sold for, were offered or were estimated at. Each one is moved to your company's facts, weighed by how much it should count, and read as three cases. The method is described here in plain words, step by step, because the same reasoning works on paper and because a range you cannot trace is not worth much.

One principle sits under every step: a figure no record states is not used. The few rules the method is given rather than reads (how much a sale counts against an offer, how far apart the outer cases sit) are named and shown with every result, and a new record in the ledger, or one dollar changed in an existing one, changes the valuation.

Step 1: choose the comparables

Every per-company record in the ledger is a candidate comparable, with two exceptions. A record whose buyer publishes a rule your company does not meet is left out, because that buyer would not deal with you at that price or any other: a figure from a program that requires 30+ employees says nothing about a company of twelve. And a record that states only words ("six figures") or an open-ended figure ("up to" a sum, "from" a sum) sets nothing on its own, so it is listed with the reason it was left out rather than quietly dropped.

Step 2: move each comparable to your company

A record about a company of 100 people with 10 years of records in the US does not price a 40-person company with six years. So each comparable is moved along the dimensions where it differs from yours, by slopes fitted from the buyers' own estimators: pairs of outputs from one estimator that differ in a single input show how the market prices that input alone. The slopes in force today:

The last row needs a word of explanation. Buyers name the kinds of records they want on their own pages: conversations (email, chat), documents, tickets and project histories, code, customer records, finance, calls and meetings, operations (orders, inventory, scheduling), product data and engineering drawings. Each kind weighs as often as buyers name it. A company whose records cover the kinds buyers name most often covers more of "what buyers ask for", and that coverage moves the price far more than raw volume does. It is the arithmetic behind what micro1's chief revenue officer said in plain words: " If a company won't sell its Slack, customer relationship management, or email records, the rest of the corpus is less valuable."

Step 3: weigh each comparable

Not every record deserves the same say. Each comparable is weighed three ways. By what it is: a sale that happened, a buyer's published offer and an intermediary's figure count 1, 0.5 and 0.25 against each other. By its age: the newest records count fully and older ones progressively less, because this market moved fast in 2026. By how far it had to be moved: a record that needed large adjustments to fit your company counts less than one that was already close. A record that states no company size at all counts only as far as its price agrees with what the sized records say for your company, and is left out when it lies far outside them.

Step 4: turn offers into expected sales

Here is the correction that most separates a valuation from a calculator. Offers and estimates sit above what closes (the section on quoted and realised prices below explains why). So each offer and estimate among the comparables is scaled by the ratio between what the sales and the offers among those same comparables came to, both moved to your company. A sale is taken as it stands. If no sale is among the comparables, the offers are taken as stated and the result says so.

Step 5: read three cases

The weighted, moved comparables form a spread of prices for a company like yours. The pessimistic case is the price a tenth of that weight falls below; the optimistic case is the price a tenth lies above. Once enough sales sit among the comparables to state a spread of their own, the sales bound both cases, so neither edge can run past what real sales show.

Step 6: place the base case with a score

The base case sits between the two edges, and where it sits is decided by six criteria that buyers name in their own words, each weighted by how many buyers name it:

Each criterion is scored from 0 to 1, from measurements of your exports where you provide them and from your answers where you do not; a criterion with neither sits in the middle, because nothing places it either way. The higher the combined score, the closer the base case sits to the optimistic edge.

Step 7: apply the gates

Some problems do not lower a price; they remove part of the lot. The gates check who can sign (someone with authority to bind the company, such as an officer or the fiduciary of an estate), the rights, sensitive categories, people's identifiers and credentials. A part that fails a gate leaves the value by its share: finance records holding people's bank or card details leave the lot and the sale, and code with passwords or keys in its history is held at nothing until they are removed and changed. The rest of the gates decide whether a sale can go ahead at all.

Step 8: show deal terms beside the range, never inside it

The same records fetch different prices under different terms. An exclusive licence fetches about 4 to 5 times a non-exclusive one, by what founders who sell training environments told Epoch AI, and an evaluation-only licence about 0.2 to 0.5 times a full commercial one. These are shown next to the main figure and never blended into it, because they depend on a decision you have not made yet.

What it looks like on a real set of answers

Take the valuation's starting answers: a company of 30 to 49 people at its largest, 6 to 10 years of records, operating, based in the United States, some systems linked, common work such as software, marketing or general services. On today's records it returns $74,000 to $590,000. Change only two answers, to a company that has closed and whose files are mostly loose, and the middle of the range falls by 81%. The range is wide on purpose: it rests on answers alone. Measuring your exports narrows it, a checked de-identified sample narrows it further, and only a buyer who has seen the data can turn it into an offer.

What moves value up

The valuation's criteria come from what buyers say they pay for, and the buyers' own pages are blunt about it. Across the kinds of data we have catalogued, six qualities come up again and again as the things that move a price, and each one can be checked in your own systems before a buyer ever sees a sample. They are listed in the order of how many kinds of data each one moves.

What they have in common is that none of them is size. Nyne puts it in one line: " More records do not automatically mean a higher offer." A buyer is paying for what a model can learn from your records that it cannot learn anywhere else, and a mailbox of loose emails teaches less than a smaller set of tickets, each tied to the conversation that opened it and the change that closed it.

Linkage, in numbers

Linkage is the quality holders most often underrate, and the clearest public evidence for it is a buyer's own estimator. Ooak's calculator values an operating company of 51 to 100 people with 4+ years at $85k to $121k with six tools (Google Drive, Gmail, Slack, Jira, GitHub, HubSpot), but at $16k to $22k with only Jira, GitHub and HubSpot; Handshake says its highest payouts include data from 8 or more named applications. The chief executive of Sunset, a buyer of closing companies' data, told Forbes that a Jira ticket tied to a specific code commit is worth more than a standalone document, and that healthcare and finance carry a premium.

The practical consequence is about how you export, not only what you hold. Records exported system by system, with the identifiers that tie them together stripped out, lose most of their linkage on the way out. Records exported with those identifiers intact, and de-identified in a way that keeps the same person or ticket consistent across systems, keep it. The Spirit agreement wrote that requirement into the contract: certification to the legal standard while preserving referential integrity.

Specialised trades

Which trades count as specialised? Buyers say so in their own words. In October 2026 Business Insider reported buyers saying that law and accounting see the strongest demand, and Mode's chief executive said value rises with the number of white-collar employees. Sunset named healthcare and finance as carrying a premium. HUMXN Research Labs is built around skilled trades and manufacturing, and The General Data Company around industrial companies' operational records. The common thread is work whose judgment a lab cannot buy from a crowd of contractors. The holder pages for each kind of company say what buyers ask each one for.

What buyers turn away

The mirror image of the drivers is a list of things buyers say they do not want, or want much less than holders expect. Some of these are commodities a lab already has in bulk; some are records a buyer cannot legally take; one is a category buyers exclude by their own terms. Knowing the list saves a holder from building a sale around the part of the archive that carries the least value.

Beyond that list sit the records a buyer would want but you may not be able to sell. Other companies' confidential material inside your email and drives is the most common: in the Spirit sale, five third parties objected over their own data inside the archive, three of them moving to enforce confidentiality agreements. Employee communications raise their own questions, covered in the guide to employee emails and chats in AI training. And software terms can stop an export: the Slack, Microsoft 365 and Zendesk pages set out what each system's terms say about exporting data and training AI. None of these makes an archive worthless. Each one, found late, takes a part out of the lot after the price was agreed, which is the most expensive time to find it.

Why quoted and realised prices differ

Every holder who has compared a buyer's calculator with a written offer has met the gap. Across the price board, the middle of what buyers offer per company is $316K per company, and the median of disclosed sales is $28.3K per company. Part of that gap is mix (sales are disclosed more often for small closed companies), but most of it has causes a holder can see coming and, in part, manage.

The first three rows run one way and the last runs the other. That is the whole story in four lines: headlines and calculators overstate what a single buyer pays, and competition between buyers is what closes the gap, sometimes from below.

The six reasons

1. Headlines are ceilings. "Up to" figures describe the best deal a buyer has done or hopes to do. SimpleClosure's averages, Ooak's smallest closed deal and Turing's per-codebase average all sit far below the same companies' headlines, as the record of headline ceilings against realised averages shows.

2. Estimators are not offers. They are tuned per buyer, differ about 21 times at the low end for the same company, and are labelled indicative. Nyne, which promises to beat a qualifying written offer, says plainly that calculator estimates and 'up to' amounts do not qualify.

3. The buyer reviews, and can reject. A published range is set before anyone looks at your records; an offer is set after the sample. Acceptance continues after delivery: micro1's published master terms give the buyer 30 business days after delivery to reject the data, and silence counts as acceptance. The same terms take a worldwide, perpetual licence with the right to sublicense through multiple tiers for the one fee: a resale right that a holder should price, not give away.

4. Rights and de-identification come off the top, or decide the winner. Some buyers pay for de-identification: Google and micro1 each bore all de-identification costs in the Spirit sale with no price reduction. Others net it out. And who controls the process is itself priced: in the Spirit auction Mercor offered $5.2 million with independent de-identification or $7 million with its own tools, and the estate chose the lower bid, saying privacy factors may be outcome-determinative. A higher number on a weaker structure did not win, and micro1's later $12.5M has not displaced Google for the same reason.

5. One buyer at a time never shows you the market. Buyers value the same archive on different theses about how many useful training tasks it yields, which is why their numbers differ so widely, and some press for speed: Business Insider reported small businesses receiving six-figure "exploding offers" that expire in 48 hours, with a broker quoting buyers as saying " I'm offering you double market. I need an answer by Friday." The Spirit auction is the counter-example: the same archive went from $5M to $10M once three buyers bid against each other.

6. Terms change the price of the same records. Exclusivity fetches about 4 to 5 times a non-exclusive licence, and a non-exclusive licence can be sold again: Troveo advertises an average of 4 sales per dataset (labelled illustrative). A quoted price for an exclusive, perpetual licence and a realised price for a two-year non-exclusive one are not the same thing, even when they describe the same archive. The guide to what a data licensing agreement contains walks through each term, and the negotiation page covers how to hold your position on them.

Three worked examples

The rules above are easier to apply to a real company than to read in the abstract. Below are three realistic holders, each described by its facts, each walked through the same questions: which buyers' rules it fits, which public records sit closest to it, what moves it up or down inside its range, and which rights questions decide whether the range holds. These are examples, not valuations of any real company, and the only numbers in them come from records. For your own range, the valuation runs the same method on your answers.

Example 1: a 40-person SaaS company with six years of Slack, tickets and CRM

The facts. A US software company, still operating, with about 40 people at its largest. It has six years of Slack, a ticketing system (support tickets in Zendesk and engineering work in Jira, many tickets linked to the GitHub pull requests that closed them), a HubSpot CRM with every deal's emails logged against it, and Google Workspace email and Drive. It sells a horizontal product, not one built for a specialised trade.

Which buyers fit. It clears the published size rules of the large programs: micro1's 30+ employees, US prioritised, primarily English and Nyne's at least 20 full-time W-2 employees and three years in business. It also sits in the bottom tier of Replay's size bands. That matters more than any estimate: a company that clears several buyers' rules can have those buyers compete.

The nearest public records. No sale of a company exactly like this one is public, so the closest evidence is offers and estimates for companies of similar size. They are spread widely, which is the normal state of this market:

What the valuation does with them. The valuation's starting answers describe almost exactly this company (30 to 49 people, 6 to 10 years, US, operating, some systems linked, common work), and on today's records they return $74,000 to $590,000. The buyers' estimators above sit toward the top of that range and the intermediary's calculator and Replay's tier toward the bottom, which is what a method that weighs every record, and scales offers by what sales among the comparables came to, should produce. Within the range, this company has two things that move a base case up. Its tickets point to code changes and its deals to their emails, which scores on linkage; and resolved tickets and closed deals show outcomes. What holds it back is that the work is common software work, not a trade buyers name as scarce.

What decides whether the range holds. Three rights questions. First, how the Slack history comes out: Slack's terms on bulk export and AI use (see the Slack rights page) decide which export route a buyer will accept. Second, customers' personal data inside the tickets and CRM must be removed or de-identified before any sample leaves the building (the de-identification standard sets out how), and customer contracts may limit what the company can do with records about each customer, which the guide on whether it is legal to sell customer data covers. Third, the code is a separate asset with its own buyers and price logic, set out in the guide to selling your codebase; selling the communications without deciding the code's fate can cost the code its best buyer.

Example 2: a closed marketplace

The facts. A two-sided online marketplace that shut down last year. At its peak it had about 60 people and five years of records: Slack, Google Workspace email and Drive, support tickets with their resolutions, a CRM for its merchants, payment records, analytics events and its code. A former officer still has admin access to most systems, and the subscriptions are being cancelled one by one.

Which buyers fit. Closed companies are a market of their own. Wind-down marketplaces take them routinely, and micro1 runs a separate program for them that asks for 50+ peak employees, substantial Microsoft Teams or Slack data. There is direct evidence that labs buy exactly this kind of archive: one of the five lots a lab buyer bought at once in October 2026 was a creator marketplace's 96 repositories, 313K CRM contacts, about 83K sales-call recordings, 57K support tickets, about 200K Drive files, 120K+ Slack messages and 58M analytics events.

The nearest public records. For closed companies the realised side is better documented than for operating ones: SimpleClosure's $10K to $100K range with an early average of about $10K each, cielo24's mid-six figures for a much longer history, and one founder's roughly $25,000 for code plus around $15,000 expected for workspace data. The market's estimators price a closed company at 0.65 times the same company operating, and in the valuation, a closed company whose files are loose has its middle cut by 81% against the starting company. This marketplace is not loose: its tickets have resolutions and its CRM links to sellers, which keeps it at the better end of the closed band.

The part buyers will not take. A marketplace's most voluminous records are about its users: shoppers' and merchants' names, messages to each other and payment details. Those are consumers' personal data. The Spirit sale shows the market's answer: it " does not contemplate the transfer of the personal consumer data or e-mails of Spirit consumers", and the passenger systems were dropped from the sale in any form. Payment records holding people's card or bank details leave the lot entirely under the valuation's gates. What sells is the company's own work: the support playbooks and resolved tickets (de-identified), trust and safety decisions, internal discussion, documents and the operating history of running a marketplace.

What decides whether the range holds. Two clocks and one signature. The clocks are the systems' deletion periods after cancellation: Jira and Confluence Cloud keep data 60 days, Microsoft 365 180 days at most, Google Workspace states no recovery window, and a Slack workspace downgraded to free loses history older than one year, on a rolling basis. Every subscription cancelled before export takes a kind of record out of the lot for good; the deletion clock counts them down. The signature is about authority: in a closing company the right officeholder signs, and in a US bankruptcy the sale goes through the court with a consumer privacy ombudsman where personal data is involved. The estates page and the guide to selling customer data in bankruptcy cover both, and the holder page for closing companies and estates lists what buyers ask them for.

Example 3: a large professional services firm

The facts. A US accounting and advisory firm, operating, large enough to sit in the top size band of Handshake's estimator ( 200+ employees), with twelve years of records: Outlook email and Teams, SharePoint holding workpapers, review notes and templates, a practice-management system with time and billing, and a CRM. Its people are almost all desk workers, and most of its work is written down.

Which buyers fit. Nearly all of them on size and place. More useful is what buyers say about the trade: buyers told Business Insider that law and accounting see the strongest demand, and micro1's $500k+ tier is defined as "large-scale operational datasets or ongoing participation across multiple teams and business functions", a description that fits a firm with several service lines.

The nearest public records. At this size the published figures move into seven digits, though with the same caveats as before:

What the valuation does with them. Size helps less than the jump from the first example suggests: each doubling of head count multiplies the price by about 1.47 times, not by two. What lifts this firm is the criteria: specialised work in a trade buyers name, history across twelve years, and reasoning visible in review notes and workpapers that show why a judgment was made, not only what it was. The firm's English text and its desk-heavy staff count for it too.

What decides whether the range holds. In a services firm, much of the archive is about clients, and that is where the rights work concentrates. Client confidentiality under engagement letters and professional rules decides what can leave even in de-identified form. Privileged material comes out as a matter of course: the Spirit agreement excludes attorney-client privileged material and requires its return if transferred by mistake. Third parties' confidential documents in email and SharePoint need finding before a sample, as the Spirit objections showed. And a living firm has a choice the closed marketplace does not: it can license non-exclusively for a fixed term and keep selling. Disclosed AI licences run two to three years in several of the best-known deals, and for a living company a two to three year term with renewal and refresh rights keeps the holder's leverage. The holder page for professional services firms lists what buyers ask that kind of firm for.

What the three examples show together

Put side by side, the examples make three points that hold for most holders. Fit comes first: which buyers' rules you meet decides how many buyers can compete, and competition, not size, is what moved the one fully public price. The criteria move the base case: linkage, outcomes, history, specialised work and visible reasoning decide where in the range a company sits, and each can be checked in your own systems this week. Rights decide whether the range survives: consumer data, client confidentiality, privileged material and deletion clocks each take parts out of a lot, and they do it most expensively when found after a price was agreed.

Questions people ask

How much is my company's data worth? It depends on what you hold, how linked and how long it runs, and whether the rights are clean. Public offers for one company's records have a middle half of $174K to $693K per company, disclosed sales a median of $28.3K per company, and the one fully public auction reached $10 million. The valuation puts a range on your own company from those records.

How much is a company's Slack worth on its own? Usually less than the same Slack sold with email and CRM: micro1's chief revenue officer said that without Slack, CRM or email records the rest of the corpus is less valuable, and the reverse holds too. Depth counts: one buyer said one company's deep Slack can beat another's whole archive. Check Slack's export terms on the Slack rights page before promising a buyer anything.

Do AI companies pay per employee? Three buyers' estimators use a per-person term: $140,000 plus $2,333 a person plus $11,667 a year of records, times a location factor. Across the whole market, though, doubling a company's head count multiplies the price by only about 1.47 times. Head count is a proxy for how much work your records describe, not the thing being bought.

Is a closed company's data worth less? On average yes: the market's estimators price a closed company at about 0.65 times the same company operating. But closed companies sell all the time, through wind-down marketplaces and programs such as micro1's for companies with 50+ peak employees, and the bigger risk is losing records to cancelled subscriptions before they are exported.

Is the number on a buyer's calculator an offer? No. The calculators are labelled indicative (" Indicative only; subject to a full review of your internal tools and data volume"), they differ by about 21 times for the same company, and Nyne's offer-beating promise excludes calculator estimates and 'up to' amounts. Only a written offer after a sample prices your data; the offer check compares one against the public record.

Can I sell the same records to more than one buyer? Yes, if each licence is non-exclusive, and non-exclusive data can sell more than once. An exclusive licence fetches about 4 to 5 times a non-exclusive one, so the choice is a price decision; price exclusivity as a multiple, and never give it by default.

How fast do buyers pay? Of the ten buyers the selling guide compares, 5 publish how soon the holder is paid. Where de-identification must be certified, payment follows delivery: the Spirit price is paid only on delivery of de-identified data.

What is the smallest sale worth running? The weighted average of what a small company's data sold or was priced at, across the small-company records in the ledger, is $26,000, and that is the smallest sale we take on. Below it, listing directly with a marketplace or a wind-down service is usually the better route, and the valuation names the ones whose published rules fit.

What this means for you

A range is a tool for two decisions: whether a sale is worth the work, and whether an offer in front of you is far below what the public record supports. It is not an asking price. Used that way, these steps take a holder from "what is it worth?" to a position it can act on.

1. Inventory before you price. List each system, its years of history, how many of your people appear in it, and whether its records link to each other. The readiness check walks through it, and the company records page shows what buyers ask for kind by kind.
2. Get a range from all the records, not one calculator. The valuation starts from your answers alone, with every comparable behind the result; measuring your exports narrows it. The rates page and the price board show the market figures it reads.
3. Check the rights before any sample leaves. Consumer data, client confidentiality, privileged material, other companies' confidential documents and software terms are the usual problems. The rights hub and the datasheet tool help build the rights file a buyer will ask for.
4. Read every offer against the record. An exploding deadline or an "up to" figure is a signal to slow down. The page on offers in your inbox and the offer check say what to look for.
5. Decide the terms you will sell on. Exclusive or not, how long, which uses: each changes the price of the same records. The licence terms and the negotiation page cover them.
6. Let buyers compete. The one fully public price in this market rose from $5M to $10M because three buyers bid. The guide to selling company data to AI labs explains the process, what labs buy shows where demand sits, and direct or through us sets out when to go to a buyer yourself.

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its data to the buyer, and we do not take ownership of it. Only the holder pays us, a fee fixed in the mandate and paid only if a sale closes, and we run sales only for holders based in the United States.