What Is Turing's Project Lazarus? Pay, Terms and Process

Project Lazarus is Turing's programme for companies' working records: Turing prepares a company's chat, email, tickets, documents, CRM, databases and code at its own cost, licenses the record to AI labs on non-exclusive licences, and pays the company from each sale, on terms set in its agreement, with "Nothing upfront." It takes active, winding-down and wound-down companies from most regions, with the strongest demand in the US, UK, Canada and the EU. For the first 90 days after you sign, only Turing sells your record, then you can renew or opt out, and if nothing sells you owe nothing. What Turing does not publish is the number that decides what you earn: how each sale is split, which your agreement sets after Turing has priced a real sample, reviewed within 5 business days.

This guide is for the founder, CFO or counsel of a company that has heard Turing pays for company data and wants to know what it would actually be signing. It answers the questions in the order they come up: what Project Lazarus is and how it has changed, which companies it takes, what records it wants, how it pays (the revenue share, the 90-day selling right, renewal and opt-out), what happens to employees' personal data, how long it takes, how it compares with selling for a fixed price to micro1, Mercor, Handshake AI or Scale AI, and how to apply or be introduced. It is about the programme at lazarus.turing.com; Turing's other businesses come in only where they bear on it. Every term is Turing's own published wording, cited to its record; where a line is our reading rather than Turing's words, it says so.

A note on who is writing. AIData.Broker is in Turing's referral programme: Turing pays us for each company we introduce, and the company pays us nothing. We are not Turing's agent or representative, and Turing's own review decides whether it takes a company. A holder we introduce to a buyer pays us nothing: the buyer pays us its published referral payout, out of its own money; Turing's is up to $50K, paid if the company signs and its record sells. In our own sale the holder pays us a fee fixed in the mandate, only if a sale closes. Either way the holder itself licenses or sells its data to the buyer, and we do not buy a holder's data, take ownership of it, or license or sell it in our own name.

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

Contents

1. What Project Lazarus is
2. Which companies it takes
3. What records it wants
4. How Lazarus pays
5. What happens to employees' personal data
6. How long it takes
7. How it compares with selling to micro1, Mercor, Handshake AI or Scale AI
8. How to apply or be introduced
9. Questions people ask
10. What this means for you

What Project Lazarus is

Turing is one of the large suppliers of human data to AI labs. It sells custom and off-the-shelf datasets, RL environments and human experts to frontier labs in software engineering, enterprise knowledge work and science, and it reported around $300 million in ARR when it raised a $111 million Series E at a $2.2 billion valuation in March 2025. An RL environment is a simulated workplace where an AI model practises tasks and is scored on the result; companies' real records are the raw material that makes such environments realistic, which is why Turing wants them. Turing also builds training tasks and evaluations to order: ServiceNow Research's EnterpriseOps-Gym benchmark states that each task Turing made for it cost approximately 100 USD.

Project Lazarus is the line of that business that takes records a company already holds. Turing's own words for what it does are short: "We prepare your company's record, license it to labs, and pay you every time it sells." The company keeps the data; Turing gets the right to prepare the record and license it, and ownership does not change hands. Turing's FAQ says the buyers are frontier AI labs and companies building AI models ( its home page). In plain terms, Lazarus is a reseller paid by revenue share: it carries the cost of cleaning and packaging the record, sells licences to it, and passes part of each sale back to the company.

How the programme changed in 2026

Lazarus has changed its basic offer more than once in a few months, so an article, a post or an email about it may describe a programme that no longer exists. It began with code. A Turing newsletter of March 2026 priced a modern codebase at $1,000 per PR, up to $10K per codebase, a legacy one (COBOL, .NET, Java) at $10K to $100K+ depending on complexity, and company documents such as tickets and specs at $50 per artifact when bundled with a codebase. In September its CEO told The Information it had bought five to ten failed-startup codebases, paying an average in the tens of thousands of dollars each.

By October the page had widened to a company's whole working record, offered for a single payout of up to $1M, with repositories that typically pay $10k to $100k+ and a promise of one payment, with no recurring commitment. Its process page was then telling holders that most went from first conversation to payment in under two weeks. By 2 November 2026 all of that was gone: the single payout became a revenue share on every licence sold, nothing upfront, with Turing alone selling the record for its first 90 days. And when we read the site again on 7 November, its FAQ had added a second way it pays: when Turing builds training tasks, evaluations or environments on a company's record, the company is paid for that work too ( Turing's FAQ).

The practical lesson is to read the agreement, not the last article you saw. A figure quoted for Lazarus from before November (the single payout, a price per repository or per pull request) describes an offer Turing has withdrawn.

Which companies it takes

Turing publishes no minimum size, age or industry for a company. Its FAQ answers "Do we have to be shutting down?" with "No. Most companies we work with are still active. Winding-down and wound-down companies follow the same process." Its intake form asks which systems the company used and how many years of records it has, and states no cut-off on either ( the form). That sets it apart from the fixed-price programmes, each of which names a floor: micro1 wants 30+ employees, Handshake AI 20+ full-time employees and 3+ years in operation, and Scale AI says 40+ employees is its preferred starting point.

On place and language, Turing says companies can be based in most regions, with strongest demand in the US, UK, Canada and the EU; English-language records sell best, and it will flag early any jurisdiction with export rules. It is one of only 7 buyers of company records on our board that say they take holders from anywhere or work with holders abroad. Its general terms of service forbid use that breaks US export control laws administered by BIS or OFAC, with no country list, so a company in or owned from a sanctioned country should expect to be refused. A company outside the US should also read selling data to AI companies from outside the US.

Operating, winding down or wound down

For a company that is still trading, Turing works from a copy, on the company's timeline: systems keep running and the originals stay where they are. For one that is closing, time matters more than anything else on the page. Turing's intake warns that records depreciate fast after operations cease (Slack 30 days, CRM 60 to 90 days), and its home page tells winding-down companies to start before subscriptions lapse. For a company that has already closed, the page says the record can still sell if the company kept exports or still has admin access ( Turing, company status).

Whoever applies must be able to sign for the data: Turing can only buy artifacts from someone with admin authority. In a closed company that is often the hard part. A dissolved company's assets may belong to its shareholders, a liquidator or a trustee, and the former CEO may no longer have authority to license anything. Read who owns a dissolved company's data and how long to keep business records after closing before applying, and the estates page if a trustee or administrator is involved.

What records it wants

Turing wants the record of how a company actually worked, not a curated dataset. Its page groups it in four parts: how the team communicated, how the work got built ( "Source code, reviews, tickets, and the iteration history behind what shipped."), how the business ran (wikis, runbooks, schemas) and how decisions were made. Its intake form lists the systems it asks about: chat, email, tickets, wikis and docs, CRM, support desk, databases, source code and design files ( the form). It says it is strongest in software and engineering, finance and accounting, customer support, sales and CRM, legal and contracts, and operations and logistics.

What raises the value is connection. Turing says "Records that span several systems from the same team are worth more." A ticket that links to the pull request that fixed it, and to the chat thread where the fix was argued, shows a model the whole path of a piece of work; a folder of loose documents does not. Other buyers say the same: the head of Sunset, which buys closing companies' data, told Forbes that a Jira ticket tied to a specific code commit is worth more than a standalone document, and micro1's chief revenue officer said that without Slack, CRM or email records the rest of a corpus is worth less. Turing accepts old systems: its intake asks for a rough count per system, estimates are fine, and any historical system, Lotus Notes or Bugzilla included.

Code, with or without the rest

Code is welcome but not required. Turing's FAQ says source code is not required: communication, tickets and docs often matter more. Its main page used to say code was never bought on its own; when we read the FAQ on 7 November it answered "Can I offer just source code?" with yes, adding that code earns most with its commits, pull requests, reviews, tickets and docs ( Turing's FAQ).

Code goes through a screen before anyone sees it. You run Turing's evaluation script on your own machine and upload only the CSV it produces, so the code stays with you for the first look. The script counts a pull request only if it changes at least one test file, has a test that fails before it and passes after it, and an earlier test that still passes, and only if its linked issue is closed, with a body of 20 to 2,000 words. It drops pull requests from bot accounts and those whose title and body are not in English, and it runs a check for signs of AI-generated code whose consequence its guide does not state. A codebase without tests yields little on this screen, however large it is. The guide to selling a codebase covers code buyers in depth; for chat and tickets, see the guides to Slack data and Jira and issue history.

How Lazarus pays

A share of each sale, not a price

Lazarus pays through revenue share. Its FAQ: "Each time we license your record to a lab, you're paid from that sale. The exact terms are in your agreement, and we walk you through them before you sign." Turing prepares the record at its own cost, and preparation costs are covered from sales, never charged to you; the agreement sets how each sale is split. Turing's argument for the model is that "A fixed-price buyer pays you once, then keeps reselling your data." That is a fair description of the fixed-price programmes: micro1's published licence, for one, is perpetual, irrevocable and fully paid-up, with the right to sublicense through multiple tiers.

Three numbers decide what a revenue share is worth, and Turing publishes none of them. The first is your share of each sale. The second is how much preparation cost comes back out of sales, and whether it comes off before your share is calculated: if it does, the first sales may pay Turing back before they pay you. The third is how many labs buy, and at what price. Turing gives no estimate from a form, because "any number from it is a guess"; it prices from a sample instead. So until a sample has been reviewed, a company considering Lazarus has no figure from Turing at all, and after it has one, the figure is a price per sale, not a promise of sales.

There is also no published payment date. The agreement sets the terms, so the question to ask is how soon after a lab pays Turing your share arrives, and what statement shows you each sale. For how every buyer's payment trigger compares, read when AI data buyers actually pay.

The 90-day selling right, renewal and opt-out

Turing's one firm term on exclusivity is short: "For the first 90 days after you sign, Turing is the only one selling your record. After that, you can renew or opt out." Its how-it-works page adds that if nothing sells you owe nothing, and after 90 days you can opt out and take the record elsewhere. This is exclusivity of the selling channel, not of the licence: the labs Turing sells to get non-exclusive licences, so the same record can go to more than one lab. For those 90 days you cannot offer the record through another buyer or broker; after them, renewing (in our reading) keeps Turing as the only seller for a further period, and opting out frees you to sell it elsewhere.

What the page leaves to the agreement is what opting out leaves in place. Licences Turing has already sold to labs will presumably continue, since a lab that paid for a licence will not give it back, and the agreement should say so, along with whether you keep being paid when a lab renews. Ask too whether a sale Turing has started but not closed when you opt out still counts as its sale. The market prices full exclusivity at 2.0 to 5.8 times a non-exclusive licence, and two builders of training environments told Epoch AI exclusive deals cost roughly 4 to 5 times non-exclusive ones (a reported figure). Turing's 90 days are much narrower than that, and, in our reading, free to Turing: no part of the split is published as the price of the window. The guide to exclusive and non-exclusive licences explains what each kind is worth.

Ownership, the licence and a later sale of your company

Turing's FAQ is plain on ownership: "You do. You give Turing the right to prepare your record and license it. Ownership doesn't change hands." It is equally plain that the agreement can matter later: "A buyer's diligence may ask about it, so loop in your counsel before you sign." A licence that keeps running after your company is sold is something an acquirer will weigh, so know how Turing's right ends and whether it passes to a new owner. The page says nothing about how long Turing's right to license lasts, what a lab may do with the record (train only, or pass it on) or whether a lab may sublicense it. Those terms are in the agreement alone, and the guide to data licensing agreements walks through where each one sits.

What to ask Turing before signing

Our offer check reads a written offer against the published clauses of the buyer that made it, and the negotiation page explains where a holder should stand on each.

What happens to employees' personal data

A company's working record is full of people: names, messages, performance reviews, salary threads, sick notes in a direct message. Turing says it handles this on its side, before any lab sees anything. Its intake page lists the safeguards: personal data and secrets are removed before anything reaches a model, processing is isolated and air-gapped, access is limited to a named team under NDA, and you approve the scope before anything moves. Its FAQ answers "What about employee and customer data?" by saying personal details are removed before anything reaches a model and that it goes through exactly what is covered, system by system, on the call or by email ( Turing's FAQ). A mutual NDA is signed before any data moves, the sample included, and its SOC 2 Type II report is available under NDA on request.

Two things follow. First, Turing de-identifies after the copy leaves your systems, in its own environment. That is the common model (micro1 and Mercor also remove personal data themselves), but it is not the only one: Scale AI says it anonymises inside your own systems before anything is extracted. A company that wants personal data never to leave its hands should ask Turing whether it can work that way, or remove the most sensitive material itself first, with our de-identification standard and de-identification tool as a checklist. Second, Turing's scrub does not take the legal decision off the company. The company is the one licensing its records, and the law on its staff's messages applies to it.

The company's own obligations

In the US, federal law generally lets a company authorise access to messages stored on a system it provides to its own staff: the Stored Communications Act's access offence does not apply to conduct authorised "by the person or entity providing a wire or electronic communications service". State law is where it tightens. California's privacy law exempted employee data until the exemption "shall become inoperative on January 1, 2023", and the state's privacy agency says consumers now include "California residents that are employees or job applicants". Data counts as de-identified under that law only with all three of reasonable technical measures, a public commitment not to re-identify, and contracts binding every recipient. A company with staff in the EU or UK is under the GDPR, where employee consent is rarely a valid basis; read licensing data for AI training under the GDPR.

The one large public sale of a company's internal records shows where this bites. Spirit Airlines' bankrupt estate agreed to sell its de-identified business records to Google, and the court-appointed privacy ombudsman's mandate covers consumers only, leaving employees to objections and the court. Employees did object, members of Congress wrote asking the companies to halt the transfer of former-employee data until protections are in place, and the approval hearing has been adjourned again and again, most recently to 17 November 2026. A company that settles its employee-data question before applying, through notice, policy and a decision about which systems and channels to leave out, is in a far better position than one that leaves it to the buyer's scrub. The guide to employee emails and chats in AI training sets out the law state by state.

How long it takes

Turing publishes its steps in order. You answer a few questions about your systems and team, sign the mutual NDA online (nothing moves before it is signed), give rough numbers per system, share samples against a sampling request Turing builds around your systems, and then its QA team reviews the samples and comes back with your agreement within 5 business days ( Turing's process). The first three steps take minutes; the samples go at your pace, which in practice means how fast you can export them and decide what to leave out.

So the clock to the agreement can be short, and the clock to money is not Turing's to promise: it depends on labs buying. Its FAQ answers "How soon will it earn?" with a usual period counted from scoping, not from signing, which is a description of past sales rather than a term of the agreement. The earlier promise that most holders reached payment in under two weeks belonged to the single-payout model and has been taken down. A closing company should count backwards from the dates its subscriptions lapse, given how fast Slack and CRM records lose value once operations stop, and export first.

How it compares with selling to micro1, Mercor, Handshake AI or Scale AI

The large programmes that buy company records split into two models. Fixed price: the buyer reads a sample, names a price, and pays it for a licence it can use and, usually, sell on. micro1, Mercor and Handshake AI work this way, and Scale AI prices a continuing partnership the same way, delivery by delivery. Revenue share: a reseller prepares the record and sells licences to it, and the company is paid from each sale. Turing moved to this model in November; Troveo, a smaller marketplace, also pays a share of every sale. Neither model is better in general. A fixed price moves the risk of the record not selling to the buyer and gives the company money it can count; a revenue share keeps the upside if the record sells many times and leaves the company with nothing if it does not.

The live version of this table, with every buyer's source and the date it was read, is the buyer board; the full comparison of these and a dozen smaller programmes is micro1, Handshake, Turing, Mercor: data programs compared, and micro1 has a guide of its own.

What the revenue share is worth against a fixed price

A revenue share beats a fixed price only if the record sells often enough, at a high enough price, with a large enough share. Non-exclusive data can sell more than once: Troveo advertises an average of 4 sales per dataset (labelled illustrative), and SuperAnnotate pays again on each relicense. But headline numbers in this market run well above what deals settle at: Turing's own code purchases averaged in the tens of thousands of dollars, and SimpleClosure's test marketplace recovered about $1M across nearly 100 deals. The honest comparison is between a fixed offer in hand and an estimate of how many labs will license your record, which only Turing can give you, from its own sales.

The track records are thin on every side. Our board records that no holder has confirmed a Project Lazarus deal, that none has confirmed a sale of company records to Handshake and that none has confirmed one to Mercor. micro1 told the Spirit Airlines court it had closed more than 50 data transactions in 45 days, a self-reported figure no counterparty has confirmed. Ask any of them for a reference you can call.

The order you sign in matters

The terms of these programmes interact. Signing with Turing first means 90 days in which only Turing may sell the record. Signing with micro1 first means a promise that you have not already licensed the same data on the same or similar terms to anyone else, which a Lazarus agreement covering the same records may break. Neither is a reason to avoid either buyer. Both are reasons to have every offer in front of you before you sign any of them, and to decide which records go to whom: a company can license its code to one buyer and its support history to another, if each agreement describes exactly what it covers.

Worked example: a 45-person software company winding down

Example. A US software company with 45 staff, seven years of history and Slack, Jira, GitHub, Google Workspace and HubSpot is closing in two months. On published rules it fits Turing (no size rule; winding-down companies welcome), micro1 ( 30+ employees, English), Handshake AI ( 20+ full-time, 3+ years), Mercor's listing for a software company licensing its private code, at $50k to $500k, and Scale AI's preferred 40+ employees for as long as most of its staff are still there.

Its first decision is time. Turing's intake says Slack records lose value within 30 days of operations ceasing, so the company exports every workspace with its own admin tools now and keeps an officer with authority to sign after the closing date. Its second is what it can show without a sample: micro1's estimator, read in October, showed $203,016 to $376,750 for 30 employees and five years and $285,896 to $604,463 for 50 employees and ten, while Turing names no figure before a sample. Those estimators are indications, and micro1's runs the same formula as two other buyers', so they are one anchor, not several valuations.

Its third decision is the model. If it signs micro1's terms, it is paid once, up to about 60 days from the transfer, and micro1 may sublicense the record for ever. If it signs with Turing, it is paid only if labs license the record, its share is whatever the agreement says, and for 90 days no one else may sell it; but if the record proves popular, it is paid again with each sale, after the company itself has dissolved, which raises a practical question of its own: who receives the payments once the company no longer exists. A closing company choosing the revenue share should name that recipient (a shareholder, a trust or the liquidator) in the agreement. It may still choose Turing. It will know what it gave up.

How to apply or be introduced

There are two ways in, and Turing's own review decides in both. Apply directly: Turing's get-in-touch page has a path for a company applying for its own data ("It's my company's data"), which asks about your systems and how many years of records you have, then sends you to the mutual NDA ( lazarus.turing.com/get-in-touch). Use an address at the company's own domain, from someone with authority to sign.

Be introduced: get a free valuation of your data and press Introduce me on it. We introduce your company to every referral partner whose published rules take your data, Turing among them where it fits, by filing the introduction in Turing's referral form with your contact's name; Turing then writes to your contact itself and says that we introduced you. From there you deal with Turing directly, under its NDA and its agreement, and you see the most any partner may pay us before we introduce you. An introduction does not change Turing's published terms, and the company pays us nothing for it. Where our own sale earns more for a holder based in the United States, the United Kingdom and Canada, the valuation says so, and every fitting partner is introduced as well.

Whichever way you go, prepare the same three things before the first call: a list of your systems with rough counts and years, a decision about which channels, mailboxes and repositories stay out (HR, legal advice, customer content you may not license), and the name of the person who can sign. The readiness check, the datasheet tool and the guide to preparing data for sale cover each of them.

Questions people ask

What is Project Lazarus by Turing? It is Turing's programme for licensing companies' working records to AI labs. Turing gets the right to prepare and license the record while the company keeps ownership, and pays the company from each sale. Since its FAQ was updated in November, it also pays when Turing builds training tasks or environments on the record.

Does Turing buy company data? Not any more, in the sense of paying once for it. Until late October its page offered a single payout of up to $1M; by 2 November it had become a revenue share on every licence sold, nothing upfront. The company keeps ownership and licenses through Turing.

How much does Project Lazarus pay? Turing publishes no figure. It prices from a real sample, reviewed within 5 business days, and what you receive is your share of each sale, set in the agreement, after Turing recovers its preparation costs from sales. Ask for the split in writing before you send more than the sample.

Do we have to be shutting down? No. Turing says most companies it works with are still active, and winding-down and wound-down companies follow the same process. A closed company can still take part if it kept exports or admin access and someone can sign for it.

What happens after the 90 days? For the first 90 days only Turing sells your record; then you can renew or opt out, and take the record elsewhere. The page does not say what happens to licences sold during the window, so get that in the agreement.

What does Turing do with employee data? It says it removes personal data and secrets before anything reaches a model, in isolated, air-gapped processing by a named team, on a scope you approve. That does not settle the company's own obligations to its staff, which depend on where they live: at a business California's privacy law covers, staff have had the full set of consumer rights over their personal information since the employee exemption lapsed on 1 January 2023.

Is Project Lazarus legitimate? Turing is an established company, backed by a $111 million Series E, and its programme publishes its terms. What no one has yet shown publicly is a completed Lazarus sale: our records hold no holder's confirmation of a deal. Sign the NDA before sending anything, and ask for a reference.

Can I sell just source code to Turing? Its FAQ now says yes, though code earns most with its history of commits, reviews, tickets and docs. It first screens code with a script you run on your own machine, which counts only pull requests that change tests and show the change works.

What this means for you

Project Lazarus is a real option for a company whose records are rich and connected, especially one below a fixed-price buyer's size rule or based outside the US. It is a different bargain from the fixed-price programmes, and the difference is all in numbers Turing does not publish. Before you sign:

1. Get the split in writing. Your share of each sale, the preparation cost that comes back out of sales, and how soon after a lab pays you are paid.
2. Decide the 90 days. For those 90 days only Turing may sell the record; sign when you are ready to wait that long, and settle what opting out leaves in place.
3. Settle the people in the record first. Decide which systems and channels stay out, and what your staff are told, before any sample leaves.
4. Put every offer side by side. A fixed price from micro1, Mercor, Handshake AI or Scale AI and a revenue share from Turing answer different questions; the buyer match shows which of them your company fits, and the comparison of every programme shows their terms together.
5. Know the number before the offer. A free valuation prices your records against public prices, and from it you can be introduced to Turing and every other referral partner that takes your data, at no cost to you.

For more on the companies behind these programmes, see Turing's page with everything recorded about it, the buyer board, how AI companies get training data and how to sell company data to AI labs.

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