Know a company with years of records? Refer it and earn up to $100K
Data type guide

Employee Screen Recordings for AI Training: A Company Guide

How a company can license recordings of staff using CAD, ERP, claims or accounting software to computer-use AI developers: consent and monitoring law.

Markov states sold
15,000+ hoursto frontier AI labs
micro1 agent trajectories
$1–2M a yearqualifying trajectories, with approval
Notice laws read
NY Civ. Rights 52-c; CT 31-48d
A dispatch clipboard and barcode scanner in a warehouse office
Short answer

A company can license screen recordings of its staff doing real work in professional software, such as CAD, ERP, claims or accounting systems, to developers of computer-use AI agents. A useful recording pairs the objective, the screen, every click and keystroke, and the system record confirming the result. Markov and Fig collect this data, and micro1 licenses agent trajectories. Employees need notice and consent, and customer data on screen must be removed or kept off.

  • Computer-use agents learn from people operating real business software, step by step.
  • A useful recording has four parts: objective, screen video, input actions, and a system record of the result.
  • Individual programs pay people for their own computers and exclude employer data, so company work needs a company arrangement.
  • New York and Connecticut require written notice of electronic monitoring; the federal Wiretap Act has consent and business-use exceptions.

Can a company license employee screen recordings for AI training?

Yes, if the recordings are made with notice and consent, cover work the company controls, and leave customer data out. The buyers are developers of computer-use agents: AI systems that operate software the way a person does, by reading the screen, clicking and typing.

The bottleneck is examples of experts at work. Markov, a Y Combinator company, says it collects "synchronized screen recordings, mouse/keyboard actions, and annotations from real workflows" and that the data "is especially scarce for expert workflows in domains like CAD, architecture, engineering, finance, and other professional software" (Markov on Y Combinator).

This is the company side. If you are an individual who wants to be paid for recording your own screen, see get paid to record your screen for AI training.

Who buys recordings of professional software work

What the main collectors and buyers say on their own pages, as read on 12 November 2026:

CompanyWhat it says it collects or buysSource
Markov (Y Combinator, Summer 2026)Screen recordings, mouse and keyboard actions and annotations; states it "Sold more than 15,000 hours of data to frontier AI labs"YC launch page
FigRealCUA v0.1 "spans over 600 hours, 274k actions, 27k app transitions, and 35 applications", captured from "real work performed by people on their own computers"Fig
Mercor (individual listing)Pays $50-$60 per hour for "completing and narrating real multi-app productivity and automation tasks" on contributors' own MacsMercor listing
micro1 (company program)Licenses trajectories from a company's own AI agents: "We license qualifying trajectories for $1–2M annually, with your approval."micro1

Individual programs are built around a person's own computer. Mercor's listing states: "Your work at Mercor will not involve access to confidential or proprietary information from any employer, client, or institution." (Mercor) Recordings of your staff in your systems therefore need a company-level license, not individual sign-ups.

What a useful recording contains: objective, screen, actions, system record

A video of a screen alone is weak training data. Buyers can use a recording when they know what the person was trying to do and whether it worked. In Origin's experience, the system record that confirms the result is the part companies most often forget to keep.

  • Long sessions matter. Fig reports that "Across more than 27,000 observed app transitions, we found that 93% were returns to apps used earlier in the same recording" (Fig). Real work jumps between apps; recordings cut to one tidy task lose that.
  • Failures and corrections are useful. A record of an error message and how the employee recovered shows the agent what to do when things go wrong.
  • Software that rarely appears in public video, such as CAD, ERP, claims and accounting systems, is the scarce part.
PartWhat it isExample in an ERP
Objective (the decision)The task as the employee received it"Post this supplier invoice and resolve the price variance"
ContextWhat was on screen and in other apps, and the documents usedPurchase order, receipt, invoice PDF, an email from the buyer
ActionsEvery click, keystroke and app switch, time-stampedOpens three-way match, edits a line, adds a note, routes for approval
Narration or notesWhy the person did what they did"Price differs by 4%; contract allows 2%, so I sent it to purchasing"
Recorded outcomeThe system record that shows the resultInvoice posted, approval log, payment run, no later reversal

What companies are paid for this data

Published per-hour rates are for individuals recording their own computers, such as the Mercor listing above. For companies, micro1 publishes $1–2M a year for qualifying trajectories from a company's own AI agents, and company-wide ranges for all operational data:

ProgramPublished range or ruleEligibility stated
micro1 Company Data PartnershipsPayments "typically ranging from $100K to over $3M", paid within 30 days of acceptance (micro1)Established companies with 30+ full-time employees and substantial historical operational data
Handshake AI data partnerships"$100K to $4M+ payouts", described as an indicative estimate, not an offer (Handshake AI)20+ full-time employees and 3+ years in operation
Mercor data monetizationNo range published; compensation "is based on data volume, the tools you connect, and the depth of your data" (Mercor)Target industries listed on the page

These are company-wide ranges for all the records a company licenses, not prices for one record type, and no program publishes a price per record for the data on this page. Read them as the size of deals programs say they make, not as a quote. No program we read publishes a company price per hour of employee screen recording.

Customer data on screen

The hardest part is usually not the employee but what is on the screen: customer names, account numbers, claims, patient details, prices and contracts.

  • Record in a sandbox or training environment with synthetic customers wherever the software allows it. That avoids most of the problem.
  • Financial records. Financial institutions may not disclose nonpublic personal information "to a nonaffiliated third party" without the notice and opt-out steps the Gramm-Leach-Bliley Act sets (15 U.S.C. 6802). Banks, lenders, insurers and their service providers should keep live customer accounts off camera.
  • Client confidentiality. Accounting, law and claims work may be done under client contracts that bar disclosure. Record only work on your own books or synthetic files unless the client agrees.
  • Redaction after the fact is possible but must cover video frames, typed text and narration, and it must be checked by a person.
  • For methods, see how to de-identify company data before licensing.

How to prepare a screen recording program

A short checklist for an operations leader:

  • Pick two or three workflows in professional software that your team does well and that rarely appear in public, such as month-end close, a claims review or a CAD change.
  • Check each software license for any limit on recording or sharing its interface.
  • Write the employee consent: what is recorded, who receives it, how it is used, how to opt out, and whether participants are paid.
  • Set up a sandbox with synthetic data, or a redaction process that a person reviews.
  • Capture the objective at the start and export the system record that confirms the result at the end.
  • Already have SOPs, training videos or audit logs? Those can be a first sample before any new recording. See license workflow data for AI agents.

Next step

Describe the workflows and software your team uses before recording anything. That description is enough for a first conversation, and it keeps every recording decision with you.

Does your company hold records like these? See what your data is worth.

Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes (how Origin is paid). See what your data is worth.

Common questions

Can a company sell screen recordings of its employees to AI companies?

It can license them when employees have been told and have agreed, the work is the company's own, and customer data is kept off screen or removed. Buyers want recordings of professional software such as CAD, ERP, claims and accounting systems.

What makes a screen recording useful for computer-use AI?

The objective of the task, the screen video, every click and keystroke with timestamps, ideally the person's reasoning, and the system record that confirms the result. A video with no objective or outcome is much weaker.

Do employees have to consent to screen recording?

Notice is required by law in some states: New York Civil Rights Law 52-c requires written, acknowledged notice of electronic monitoring and Connecticut 31-48d requires prior written notice. The federal Wiretap Act has a one-party consent exception. Licensing for AI training goes beyond monitoring, so a specific consent is the cleaner basis; your counsel decides.

Who buys computer-use training data?

Collectors such as Markov, which states it has sold more than 15,000 hours to frontier AI labs, and Fig, which published a 600-hour dataset of real work. micro1 licenses trajectories from a company's own AI agents for $1–2M a year.

Can employees join individual screen recording programs with work data?

Generally not. Mercor's listing, for example, says the work will not involve confidential or proprietary information from any employer or client. Company systems and data need a company-level arrangement.

Sources and further reading

What is your company’s data worth?

AI companies pay $20K to $5M for company records. Six questions, no files.