How it works / Buyers
Real operating history. Listed early, owner-confirmed before anything moves.
rl.data.cool sells licenses to email, documents, CRM, tickets and call recordings from real small and mid-sized businesses, for training data and RL environments. Our buyers are AI labs, environment builders and enterprise AI teams. Here is every step from browsing to accepted delivery, what data.cool checks for you, and what you can rely on.
$0
To browse and request samples
3
Prior-sales signals on every deal sheet
10%
Held until you accept the export
4%
Buyer fee, at close
01 / Who buys
Three kinds of buyer, one exchange.
We call everyone who licenses data here a buyer. Most are one of three kinds, and each usually reaches for a different license. Usually, not always: every deal is open to offers on any license type the owner allows.
01Usually Exclusive
AI labs
In-house environment teams training general agents. Seed replicas of business apps, write tasks from real requests, build graders from real outcomes.
An environment reveals what you're teaching your next model, so labs tend to want the corpus to themselves for a window.
02Usually Slots
Environment builders
The companies labs hire to build environments: Mercor, Surge, Scale, Turing, Mechanize and others. Fill a replica with years of believable work, fast.
You buy inputs every project. A slot license is cheaper than an exclusive and still caps how many others hold the same corpus.
03Usually Non-exclusive
Enterprise AI teams
Teams tuning agents to one kind of work, like legal intake, accounts payable or a support desk, often through reinforcement fine-tuning.
You need many businesses doing the same job more than you need exclusivity. Non-exclusive is the lowest price.
02 / The flow
From deal sheet to accepted export.
Humans and agents follow the same steps. Everything here works from the web, the REST API or the MCP server: see agents.
01
YouBrowse deal sheets, free
Every deal is an anonymized deal sheet: vertical, systems in scope, volume and date ranges, license type and live demand. No account needed to read. The same data is in deals.json, the REST API and the MCP server, so your agent can search it directly.
- Numbers are in ranges. Every published combination matches at least three businesses, and listing dates are rounded to the week.
- A lot (L-) pools several similar businesses into one corpus: you get scale, and no single business can be picked out.
- Filter or sort by readiness to see only deals that have reached a given rung, like a reviewed sample.
- Can't find it? Post a buy order. We source it through our own deal flow and our broker network, guided by the demand heatmap.
02
data.coolSee how ready it is
Every deal shows a readiness rung. We list leads early on purpose: a buyer asking about a lead is often what brings the owner to the table. Each rung is a check data.cool ran, and it only counts when every rung above it is met.
- Lead: the business answered our intake questions. Vetted: automated checks found a real, established business.
- Owner committed: the owner authorized the sale. No sample is released and no bid is accepted before this.
- Data measured: volumes and date ranges read from the owner's systems. Sample reviewed: a de-identified sample passed the Clean Room.
- Where a deal starts depends on who brought it. data.cool's own intake can list a business as early as Lead. A broker's listing only appears once the owner has confirmed, so it starts at Owner committed.
- Interest in a lead goes to whoever holds the relationship with the owner. You never contact the business yourself.
Readiness
- Lead
- Self-reported
- Vetted
- Business checked
- Owner committed
- Mandate verified
- Data measured
- Counts and dates
- Sample reviewed
- Under NDA
03
data.coolRead what's been checked, and by whom
Rights checks are set by data.cool only, never by the lister. Hover any check on a deal sheet for its method and date. Brokered listings show nothing as verified until the desk has reviewed them.
- Mandate: the owner verified an email on the company domain and confirmed, or signed a mandate with data.cool. A broker's letter alone never counts.
- Registry, PII removed and privilege screen: desk checks, against the registry and Clean Room evidence.
- Screening: registry, sanctions, litigation and web presence. A sanctions match takes the deal off the exchange.
- Data proof: the owner connected Google Workspace or Microsoft 365 read-only, and we measured counts and date ranges.
- Domain checks: a separate automated score for domain age, mail provider, SaaS stack and web history.
Rights checks
- Mandate
- Verified
- Registry
- Verified
- PII removed
- Verified
- Privilege screen
- n/a
- Domain checks
- 100
04
OwnerBrokerdata.coolCheck it hasn't been sold before
Under Prior sales, every deal sheet answers that question three ways. See how each signal works below.
- The owner's answer, given at confirmation. An owner whose data is under an active exclusive elsewhere can't list.
- The broker's answer and the agreement they hold with the business: engagement letter, LOI, NDA or verbal.
- Our relisting scan, which checks every new sample for watermarks from copies already sold here.
Prior sales
- Owner
- Never sold before
- Broker
- None known
- Agreement
- Signed LOI
- Relisting scan
- Clear
05
Youdata.coolOpen a buyer account and sign the NDA
Sign in with a work email. The desk runs KYB on your first access request and replies within one business day. Once you're vetted, access requests you make signed in or with your API key are approved automatically.
- Access is under the Mutual NDA, accepted with your typed name, timestamp and IP address.
- Use materials only to evaluate the deal. Don't contact, identify or try to re-identify anyone behind a listing.
06
data.coolYouGet a Clean Room sample
Samples are de-identified extracts: names, emails, phones and other personal data are replaced with consistent per-deal pseudonyms, so threads and records still join up. Each comes with a measured residual-risk report, not a promise.
- Download links are short-lived and every download is logged.
- Your copy carries watermarks and canary records unique to your organization. Don't remove or alter them.
- Pseudonymized data can still be personal data under GDPR and India's DPDP Act. The report tells you what's left.
Sample
- Identifiers
- Pseudonymized
- Residual risk
- Report under NDA
- Link
- Expiring, logged
- Watermark
- Yours alone
07
YouOwnerdata.coolBid or buy now
Prices aren't published: every deal is open to offers. The owner keeps a private reserve, bids below it are declined, and the desk may counter. Choose slots, an exclusive or a non-exclusive license.
- A bid is binding once a human at your organization confirms it, or when it falls within an API key's spending cap.
- Agents can bid over MCP or REST. Anything above the cap waits for an email confirmation.
Licenses
- Slots
- A fixed number of buyers. Default
- Exclusive
- One buyer, for a window
- Non-exclusive
- Open, lowest price
08
YouDeposit, escrow, invoice
On acceptance you pay a 10% deposit, credited to the price (within 3 business days for buy now). The balance invoice is due within 14 days by ACH or wire through Stripe. A 4% buyer fee is added at close.
09
data.coolYouOwnerDelivery, acceptance and the holdback
The full export is delivered against a fingerprinted manifest. You have the acceptance window in your license (10 business days if none is stated) to accept it, or reject it in writing for a material mismatch with the deal sheet or manifest. The owner then has 15 business days to cure. 10% of the price is held until you accept.
- A false prior-sales declaration counts as a material mismatch.
- If you neither accept nor reject in time, the export is accepted.
- When the license ends, delete the data and copies within 30 days and certify it.
03 / Corpus to environment
A real company is the world, the task and the answer key.
An RL environment needs three parts. A licensed corpus gives you the raw material for all three, already consistent with each other, because it all happened in the same business.
01 / The world
The systems in scope
Email, documents, CRM, support desk, accounting, chat, project management, internal databases, call recordings. The records that fill your replica so it looks lived in.
02 / The task
Requests inside the history
A customer asking for a refund, a ticket escalated to billing, a partner querying an invoice. Real work someone was actually asked to do, with the context they had.
03 / The grader
What happened next
The refund issued or refused, how the ticket closed, the invoice paid, disputed or written off. Real endings are an answer key nobody had to invent.
The outcomes are in the records themselves; labelling them for your grader is part of your build. New to environments? Start with Learn.
Read a deal sheet like a builder
Every field is a range or a grade, so no business can be picked out. Here is what each one tells you about the environment you could build.
- Systems
- Which apps your replica needs, and whether tasks can cross between them. Cross-application work is where models fail most.
- Volume
- Approximate work pieces (emails, documents, tickets, records), from under 10k to 10M+. How many tasks and how much world state you get.
- History
- From 1-3 years to 10+. Long histories carry policy changes, staff turnover and drift: the material for long-horizon tasks.
- Documentation
- How much of the business rules is written down. Mostly undocumented means the agent has to infer the rules, which makes harder tasks.
- Data proof
- Volumes and date ranges data.cool measured by connecting read-only to the source system, rather than what the lister declared.
- Lot (L-)
- Several similar businesses in one corpus. Variety across companies doing the same job is what helps a model generalize.
- Emerging tier
- History and documentation are still being confirmed, so scope can be shaped to your spec before the export is cut.
04 / Prior sales
Nobody wants used wood in a new house.
If you build environments for a lab, training data a competitor already has makes you look bad to your customer. So we don't ask you to take one party's word for it.
01 / Owner, at confirmation
The owner's answer
Never sold, licensed exclusively before (ended), licensed non-exclusively, or not sure. An owner who says an exclusive is still active elsewhere can't confirm the listing at all.
Warranty under the Seller Terms
02 / Broker, at listing
The broker's answer
No prior sale known, sold before, or doesn't know, plus the agreement they hold: engagement letter, LOI, NDA or verbal. When it disagrees with the owner, the desk flags it.
Warranty under the Broker Terms
03 / data.cool, on every sample
The relisting scan
Every copy we release carries watermarks unique to its buyer. Each new sample is scanned for them, and a sample that carries them can't be approved.
Our own check
What the scan can't see. It catches data that left through rl.data.cool. It can't see a sale made entirely somewhere else, which is why the declarations exist and why a false one is a breach: you can reject the export as a material mismatch and be refunded from the holdback. When a field reads "Not declared", that is a signal too.
Before the sample is approved
- Relisting scan
- Runs on the Clean Room sample
- After approval
- Clear
05 / What protects you
What you can rely on.
01 / Mandate
Nothing moves without the owner
02 / Prior sales
Used data, three ways
03 / Remedy
Holdback and cure
04 / Privacy
Measured, not promised
05 / Watermarks
Your copy is yours
06 / Diligence
Blind or disclosed
06 / Procurement
What your legal and procurement team will ask.
The short answers, with the documents behind them.
- Who are we buying from?
- In a Blind deal, data.cool licenses the de-identified data in its own name as licensor of record, holds the owner's warranties and indemnity for your benefit, and enforces them for you. In a Disclosed-to-Counsel deal, your outside counsel can see the owner's identity once the owner approves them.
- Did the owner agree?
- Before anything binding, yes, on the record: the owner verified an email on the company domain and confirmed, or signed a mandate with data.cool, with the date logged. Until then the deal is labelled a lead. A broker's letter alone never counts.
- How long does it take?
- Weeks, not minutes. The owner commits, the sample is reviewed, the full export is prepared after your binding bid, then you have the acceptance window in your license. Readiness tells you how much of that is already done.
- Was it sold before?
- The owner's and the broker's prior-sales warranties, plus our relisting scan, are on the deal sheet. A false declaration is a material mismatch.
- Personal data?
- De-identified with consistent pseudonyms and a measured residual-risk report per sample. Pseudonymized data can still be personal data, so a Data Processing Addendum applies.
- What can we show a regulator?
- A dated, owner-confirmed license and a fingerprinted manifest for the corpus: the records you need when you document licensed private datasets, for example in an EU AI Act training-content summary.
- Vendor checks?
- The desk runs KYB on your organization on your first access request, within one business day. The Mutual NDA is accepted with a typed name, timestamp and IP.
- When it ends?
- Delete the data and copies within 30 days of the license ending and certify it. An independent auditor may check compliance on 15 business days' notice.
Documents: Buyer Terms, Mutual NDA, DPA, Subprocessors, Refund Policy.
07 / Money
What you pay, and when.
- Browsing and samples
- Free.
- Buyer fee
- 4% buyer fee on top of the winning bid, paid at close.
- Deposit
- 10% on acceptance, credited to the price. Buy now: within 3 business days.
- Balance
- Invoiced, due within 14 days by ACH or wire through Stripe.
- Holdback
- 10% of the price, held until you accept the full export.
- Refunds
- Rejection for a material mismatch, then cure or refund. See the Refund Policy.
Hard lines
What you agree not to do.
- 01Don't contact, identify or try to re-identify any business, owner, employee or customer behind a listing.
- 02Don't remove or alter watermarks, canaries or notices.
- 03Don't go around the exchange for 24 months after an introduction.
- 04Don't resell or sublicense licensed data beyond what your license allows.
- 05Don't use samples for anything but evaluating the deal.
The full rules are in the Buyer Terms and the Mutual NDA. The rest is in Help.