Skip to content

The data program

Coming later

Licensed,not handed over.

AI labs pay for real work. Butterfly is building two ways for your work to earn you royalties, while your records never leave your cloud.

The film · 1½ minutesWhat your closed cases could earn

We start reading your public website as you type.

01 · One-off or licensedLicensing tasks

Not an archive, handed over.A licence you control.

Multimillion-dollar sums are on the table for a large organisation’s data, in one-off deals: the archive goes, and the money comes once. Illustrative The data program can earn the same value, licensed task by task, without handing the archive over.

Data deals today

One-off.

  1. Hand over an archive.
  2. Take one payment.
  3. It’s gone for good.

In a bankruptcy auction, a large technology company bid $10M for one airline’s emails and Teams chats.38

The data program

Licensed, task by task.

  1. Your records stay in your own cloud.
  2. You choose the tasks, and approve each copy.
  3. Each licence would earn you royalties, and the same task could be licensed again.
  4. New tasks every day, a fresh pack every quarter.

02 · Your synthetic dataLicensing synthetic data

Data made from your recordshas value of its own.

Inside your own account, Butterfly Tasks would write a synthetic dataset from your records, with every personal detail made up. The dataset has value to AI labs on its own, so you could license it too, on terms you approve.

  1. What it is

    Synthetic examples written from your own records, inside your own account. Every personal detail would be made up, and each example checked six ways before anything leaves.

  2. Why it has value

    It keeps the shape of your real work, its fields, steps, rules and outcomes, with no real person in it: data of the kind AI labs train and test their models on.

  3. How licensing it would work

    You approve the licence terms: the licensee, the use and the term. Nothing leaves until a person you name approves it, and each licence would earn you royalties.

Project ButterflyIllustrativeHow licensing your data would work

03 · How a task or a dataset travelsBoth ways

Checked, packaged,and never emailed.

Before a task or a dataset is licensed, it is reviewed, and a person you name approves it. Then it goes straight into the lab’s own account.

Never leaves your account

  • Your records
  • Your held-back cases
  • Your readiness tests, never licensed
  • A recording or a person’s name

Leaves once a person you name approves it

  • A copy of a taskReviewed by experts, its grader tested, before any lab sees it.
  • A synthetic datasetEvery personal detail made up, each example checked six ways.

Personal data about your customers never leaves your account.

Trust & security

  1. 1 Checked

    Reviewed, and approved by a person you name, before any lab sees it.

  2. 2 Packaged

    It travels as a versioned package with a fingerprint, so both sides can prove what changed hands.

  3. 3 Delivered

    Straight into the lab’s own cloud account. Never by email.

For your engineers: the checks, the package and the delivery routes

The checks

  • The reference solution passes
  • A do-nothing agent fails
  • An exploit agent attacks the grader
  • A second expert solves it blind
  • At least three independent reviewers

Review is the bulk of the work: even heavily reviewed public tasks break, and 28 of Terminal-Bench 2.0’s 89 needed fixes within six months.15

The package

v1.2 · sha256 7f3a…e91c
  • The task folder instruction.md, task.toml, environment/, tests/, solution/ (the Harbor format, with an adapter for UK AISI’s Inspect)
  • A data card its source, its scope and its limits
  • A QA report with a pass-rate table across several models
  • Provenance and consent the chain from your closed case to the copy
  • A held-out split kept apart for testing

Every fix is a new release. Example

The delivery routes

  • Cross-account S3 grant
  • AWS Data Exchange grant
  • Private container or Harbor registry
  • Gated Hugging Face repository
  • Email never

A task’s licence, in three steps

From a sample under NDA to a master licence, then a contract for fresh tasks every quarter.

  1. 1A sample pack5–10 tasks, under NDA.
  2. 2A paid pilot batchThe lab tests the tasks on its own models.
  3. 3A master licenceUse (training or testing), exclusivity, term, no re-identification, deletion and audit.
  4. QEvery quarterA contract for fresh tasks, as models improve.

04 · What labs pay forLicensing tasks

Tasks today’s modelsstill get wrong.

A task is a test an agent can attempt again and again, with an automatic verdict. Labs pay for tasks in seven parts. Here is one from an insurer, part by part.

Task · Motor claimsExample

01 · Instruction

What a colleague would be asked to do.

Decide whether motor claim 4471 is covered, set the reserve and draft the letter to the policyholder.

A test item

Must stay secret, or it stops measuring anything.

Your own readiness tests are this kind. They stay in your cloud and are never licensed, so they keep telling you which AI is ready.

A training task

Teaches the model, so labs need many, at every difficulty.

With graders that resist shortcuts. These are the tasks the program licenses, kept apart from your tests.

Good candidates

Work that can be graded, and can legally leave you as a rewritten copy.

  • Insurance claim steps with no health data in them
  • Finance-operations reconciliations matching, coding and closing
  • Telecom troubleshooting a fault, the steps and the fix
  • Permitting checks validating an application against its rules

Kept for readiness testing

Tested inside your cloud, never licensed. Labs get no more than a synthetic twin of this kind of work.

  • Anti-money-laundering and suspicious-activity files
  • Patient data
  • Privileged legal work

05 · Your tasksLicensing tasks

Thousands of tasks.More every day.

A large organisation runs around 10,000 tasks, and real work makes new ones every day. We would find them, cost each one, and pick with you the ones worth more to your business, with real demand from labs.

What you receive · 01 of 05

Illustrative

A map of every task.

~10,000tasks in a large organisation, and new ones every day from real work

  • Matching a supplier invoice to its purchase orderhundreds a weekIn the film
  • Reconciling a bank accountdozens a weekAnother example

Models improve, so tasks expire. Once models pass a task reliably, labs retire it and ask for the ones models still get wrong.

Your licensed tasks, quarter by quarter

Example
Q1
Q2
Q3
Q4
  • In use
  • Fresh this quarter: what models still get wrong
  • Retired: models now pass them
62.9% → 76.0%

Terminal-Bench 2.0’s top score on the same tasks, in four months.

Source: Terminal-Bench, 6 May 2026 14

06 · Every quarterLicensing tasks

Models improve,so tasks expire.

Once models pass a task reliably, labs retire it and ask for the ones models still get wrong. Your work makes new tasks every day, so there is always a next pack, and your readiness tests never join it.

Your licensed tasks, quarter by quarter

Example
Q1
Q2
Q3
Q4
  • In use
  • Fresh this quarter: what models still get wrong
  • Retired: models now pass them
62.9% → 76.0%

Terminal-Bench 2.0’s top score on the same tasks, in four months.

Source: Terminal-Bench, 6 May 2026 14

2.5% → 21.25%

The top agent’s share of real freelance projects done, October 2025 to October 2026.

Source: Scale AI, Remote Labor Index, fetched 25 Oct 2026 13

The harder the work agents take on, the more labs need tasks like yours.

Agents may soon do a company’s harder work. But which business will trust an AI with its sensitive processes before it has been tested on them? Butterfly’s readiness testing would give you that confidence, and a return you can plan on.

07 · Who it suitsLicensing tasks

Work that can be graded,and can legally leave.

Some work makes a good licensed task. Some stays inside your cloud, tested for readiness and never licensed.

Good candidates

Work that can be graded, and can legally leave you as a rewritten copy.

  • Insurance claim steps with no health data in them
  • Finance-operations reconciliations matching, coding and closing
  • Telecom troubleshooting a fault, the steps and the fix
  • Permitting checks validating an application against its rules

Kept for readiness testing

Tested inside your cloud, never licensed. Labs get no more than a synthetic twin of this kind of work.

  • Anti-money-laundering and suspicious-activity files
  • Patient data
  • Privileged legal work

08 · Questions06 · Questions

What people ask first.

What exactly do we receive?

First, a map of your tasks, with the cost of each and how often it happens. Then a first set chosen with you, and royalties each time a lab licenses one of its tasks, on terms written into the program’s own contract.

How do the two ways differ?

A licensed task is a test an agent can attempt, with a grader and a reference answer. A synthetic dataset is a set of examples written from your records, every personal detail made up. Labs use both to train and test their models. You would approve each licence, and each would earn you royalties.

Does any of our data leave?

Your records never leave your cloud. When you approve a task, a copy of that one task leaves as a versioned package: the instruction, the starting state, the grader and the reference answer, with personal details rewritten. A synthetic dataset leaves the same way, once you approve it. Nothing you haven’t approved leaves.

Is it anonymous?

Every personal detail would be made up inside your cloud, and each example checked six ways before anything moves. We don’t call a task or a dataset anonymous on our own word: you would see the checks, and any residual risk written down, before you approve.

Who at Butterfly sees our cases?

Butterfly staff would have zero access to your systems or records. The work would run inside your own cloud account, the examples and their checks too, and you stay the controller.

Who are the licensees?

AI labs, and the vendors that supply them with tasks. Every licence names its licensee, and you see it before you approve.

What is the private register?

A private list of companies interested in a data deal, our clients and companies not yet working with us. We share it with AI labs under NDA, each entry as its company sent it and never a record, and tell each company which of its tasks or data labs ask for.

Does it touch our agent benchmark?

No. Your readiness tests are held apart. They never leave your cloud and are never licensed, so they keep measuring what they are meant to.

Can we stop?

Yes. It is opt-in, task by task and dataset by dataset. A licence has a term, with deletion and audit rights, and you can stop approving new ones at any time.

And GDPR?

GDPR applies, and the program is designed for it: you stay the controller, with a written lawful basis, a compatibility note and a DPIA in place before anything is licensed. It runs on its own contract, with Butterfly’s role defined in it.

09 · The private register07 · The private register

Join the private register.We’ll tell you what labs want.

For our clients and for companies not yet working with us. We share the register with AI labs under NDA, then tell you which of your tasks or data they ask for, in what volume and on what terms.

Interested in

Working with Butterfly already?
The kinds of workup to 4
Roughly how many tasks

The private registerShared under NDA

Kinds of work
Roughly
Interested in

Labs never see

  • Your records
  • People’s names
  • Your email

See the task map

Your proposal

Get your proposal.Then register your interest.

A readiness report made from your company’s public website: what your tasks cost, and the first 25 to test. If it reads right, register your interest: a person from Butterfly will call you to talk it through.

We start reading your public website as you type.

Or read on:Agent benchmarksHow it works