# How AI labs use company data

> How AI labs use company data: what arrives, how names get swapped, what labs build with it, what you keep and what you can't take back. Buyer by buyer.

By Matt Burns, Vaultfish. Canonical: https://vaultfish.com/guides/how-ai-labs-use-company-data  
Published: 2026-09-30  
Updated: 2026-10-14

AI labs use company data to train and test software agents. Your Slack threads, email chains, tickets and documents show how a real task moved from request to finished deliverable. Public text can't show a lab that path.

The data usually arrives as an export or a connected account. A buyer swaps the names. A lab then builds practice tasks and scoring tests from it.

> **Data licensing means a lab gets to use your records for a while, under rules.** Think of a library. The lab borrows real work to learn how work gets done. You keep the originals.

You can't get a trained model back out of a lab. Most questions below trace to that.

The public record on what happens inside a lab is thin. Buyers describe their own pipelines on their own pages, so those claims are self-reported. Where a buyer says nothing, we say so.

## What data actually arrives?

Three collection methods show up across the buyers we read.

[Mercor](/buyers/mercor) connects to your accounts. [Its page](https://www.mercor.com/data/) lists 100+ supported tools and says no data leaves your systems until you approve it.

[Troveo](https://www.troveo.ai/business-data) asks for an export and says it needs no OAuth (a login link) into your systems. [micro1](/buyers/micro1) starts with a form that lists your tools and rough sizes. Its engineers then help your IT people with the transfer.

The content is workplace records. [micro1 lists](https://www.micro1.ai/data-partnerships) SOPs, knowledge bases, CRM data, project histories and QA processes. [Troveo](/buyers/troveo) lists chat, email, documents, work tools, CRM and support, finance, and code on its business data page.

micro1's CFO says labs want the chain: a request in chat, the spreadsheet that answered it, and the approval that followed. The CFO used a marketing budget request as the example in an undated recorded forum talk.

Each buyer has its own guide, linked at the bottom. Start with [how micro1 handles it](/guides/how-micro1-uses-your-company-data) or [how Mercor handles it](/guides/how-mercor-uses-your-company-data).

## How do the companies change the data before a lab sees it?

Every buyer says it removes or swaps personal details. The methods differ, and so does what is left behind.

Troveo shows the cleanest example on its [business data page](https://www.troveo.ai/business-data). A sentence naming a customer and an email address becomes a sentence with PERSON-014 and EMAIL-019 in their place.

micro1's David Revland, who works on human data, describes a similar swap in an undated recorded forum talk. A name becomes the same stand-in on every platform, so the role and the cross-tool trail survive.

[micro1's page](https://www.micro1.ai/data-partnerships) says some datasets get synthetic rewrites. micro1 staff said in the same talk that companies can't clean the data first, because blanking names destroys the trail it needs.

[Mercor](https://www.mercor.com/data/) says its extraction and anonymization is automated. [Handshake AI](https://joinhandshake.com/ai/data-partnerships/) says it applies a structured process to remove direct and indirect identifiers.

A swapped name can still point to a person. The [UK Information Commissioner's Office](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/pseudonymisation/) says pseudonymised data is personal data in the hands of anyone who holds the extra information. [Handshake AI](/buyers/handshake-ai) says on its own page that no process of removing names eliminates re-identification risk entirely. The detail is in [micro1's name transformation](/answers/micro1-pii-transformation) and [pseudonymization versus anonymization](/answers/pseudonymization-vs-anonymization).

Figure: steps Export or connect -> Lab trains or tests -> Model keeps what it learned. Settle what the model may keep before you export.

## What do labs build with it?

Public reporting names three uses: reinforcement learning environments, agent training and evaluations.

An [RL environment](/answers/what-is-an-rl-environment) is a practice space. A model attempts a task and an automated grader scores it. Time tied Google's Spirit Airlines bid to environments like that. Google told Time only that the data "can be helpful in improving our products and AI models."

Fortune, reporting on micro1's offer for Spirit's records, wrote that an agent needs to learn how to choose actions, use tools and recover from mistakes. The record of those steps is a [trajectory](/answers/what-is-a-trajectory-in-ai).

The third use is evaluation. micro1's workflows track has employees grade a general-purpose agent on their real daily work. micro1 receives the grade plus the agent's trajectory.

The gap: no lab has published how it used a specific purchase. We have buyers describing plans. The longer read is [what AI labs do with company data](/answers/what-do-ai-labs-do-with-company-data).

## What do you keep?

The companies split on this in their own words.

Figure: comparison of What each company says about your data rights: micro1 (You retain ownership), Troveo (Ownership never transfers), SimpleClosure (emphasized: Perpetual and irrevocable, Perpetual license). Two say you keep ownership. A perpetual license never ends, so check the contract.

[micro1's page](https://www.micro1.ai/data-partnerships) says companies retain ownership of their underlying data. [Troveo's licensing guide](https://www.troveo.ai/resources/exclusive-vs-non-exclusive-data-license) says a license is permission and you keep ownership. [SimpleClosure's terms](https://simpleclosure.com/tos/) take a non-exclusive, transferable, sub-licensable, perpetual and irrevocable license to the assets you submit.

Owning your files and the rights a buyer holds over its copy are separate questions. Read the contract for the second one. Our page on [what reps and warranties a buyer asks for](/answers/what-reps-and-warranties-does-an-ai-data-buyer-ask-for) shows what you will be asked to promise.

## What can't you take back?

Deleted files don't un-train a model. [Troveo's guide to licensing deals](https://www.troveo.ai/resources/ai-data-licensing-deals) says models already trained survive the license ending. A [law firm checklist](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes) from Frankfurt Kurnit Klein and Selz says data that has trained a model cannot practically be pulled back out.

A regulator can reach further than a contract can. In February 2024 the [FTC's order against Avast](https://www.ftc.gov/news-events/news/press-releases/2024/02/ftc-order-will-ban-avast-selling-browsing-data-advertising-purposes-require-it-pay-165-million-over) required deletion of the browsing data sent to Jumpshot and of the algorithms derived from it. That was a consumer privacy case and not a company-archive deal. It still shows that a model is not always safe from an order.

What a buyer will delete is a contract term. micro1 says data is retained only for a limited period under agreed retention policies, with deletion on request. Mercor's page says nothing on retention or deletion. [Deleting after training](/answers/can-i-delete-my-data-after-an-ai-company-trains-on-it) covers the three separate questions inside this one.

_Deleting a file is easy. Un-teaching a model is a different job, and nobody has the button._

## Where can it leak?

Four places, in the order we would worry about them.

1. **The buyer's own security.** [Mercor said](https://www.mercor.com/blog/update-on-mercor-security-incident/) a malicious release of the open-source LiteLLM library in March affected it, and that the impact on customer information was very limited. [The Next Web](https://thenextweb.com/news/meta-mercor-breach-ai-training-secrets-risk) reported about 4 TB exposed, including source code, a user database and interview recordings. I found no report of company-archive data exposed. The facts are on [the Mercor data breach page](/answers/mercor-data-breach).
2. **Re-identification.** Swapped names plus context can point back to a person.
3. **Rights you did not hold.** Your customers' and employees' messages are inside the archive. You promise the buyer you had the right to share them.
4. **The people who can see files before the swap.** Ask who that is, in writing.

Who pays if a buyer is hacked is usually a contract term. See [who is liable if the data buyer is hacked](/answers/who-is-liable-if-the-data-buyer-is-hacked).

## Which companies explain their handling?

Six publish enough to read. Each guide covers what the company asks for, how it collects, what it says it does to the data and who gets it. It also covers ownership, retention, security, how it pays and what to ask.

- [How micro1 uses your company data](/guides/how-micro1-uses-your-company-data)
- [How Mercor uses your company data](/guides/how-mercor-uses-your-company-data)
- [How Troveo uses your company data](/guides/how-troveo-uses-your-company-data)
- [How Handshake AI uses your company data](/guides/how-handshake-ai-uses-your-company-data)
- [How SimpleClosure uses your company data](/guides/how-simpleclosure-uses-your-company-data)
- [How Defined.ai uses your company data](/guides/how-defined-ai-uses-your-company-data)

The [buyers directory](/buyers) lines them up side by side. Before any call, run the [rights check](/rights-check) and read the [risk map](/risks).

Talk to a privacy or commercial lawyer in your state before you sign.

micro1 may pay us a referral fee if a company we introduce goes ahead with it. [How we are paid](https://vaultfish.com/disclosure)

[Apply to micro1's data partnership (paid link)](https://req.micro1.ai/post/74520e8b-4673-4da5-b855-53609aaf6710?referralCode=f82e519e-5d64-4e0e-9574-760c84371cf2&utm_source=referral&utm_medium=share&utm_campaign=data_referral) · [micro1's own data-partnerships page](https://www.micro1.ai/data-partnerships)

_Paid link. [How we are paid](https://vaultfish.com/disclosure)_

_Vaultfish helps you run the process. We are not lawyers and this is not legal advice. If you need counsel who knows this area, we can point you to one._

## Sources

- [Its page](https://www.mercor.com/data/)
- [Troveo](https://www.troveo.ai/business-data)
- [micro1 lists](https://www.micro1.ai/data-partnerships)
- [Handshake AI](https://joinhandshake.com/ai/data-partnerships/)
- [UK Information Commissioner's Office](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/pseudonymisation/)
- [Troveo's licensing guide](https://www.troveo.ai/resources/exclusive-vs-non-exclusive-data-license)
- [SimpleClosure's terms](https://simpleclosure.com/tos/)
- [Troveo's guide to licensing deals](https://www.troveo.ai/resources/ai-data-licensing-deals)
- [law firm checklist](https://technologylaw.fkks.com/post/102o455/selling-company-data-for-ai-training-what-to-know-before-you-say-yes)
- [FTC's order against Avast](https://www.ftc.gov/news-events/news/press-releases/2024/02/ftc-order-will-ban-avast-selling-browsing-data-advertising-purposes-require-it-pay-165-million-over)
- [Mercor said](https://www.mercor.com/blog/update-on-mercor-security-incident/)
- [The Next Web](https://thenextweb.com/news/meta-mercor-breach-ai-training-secrets-risk)