Packaged to buyer standards
We know what AI researchers look for because we've delivered it and bought it. Your records are prepared to the standard data vendors and labs expect.
For companies with 10–350 employees
AI companies now pay for real records of how business work gets done. Most of those deals go to large companies. Ground-Set.AI brings companies like yours together, turns your records into ready-to-use training data, and shares the proceeds with you when it sells.
No cost to join. Nothing is shared until you opt in to an offer.
Founded by an early Scale AI employee who led the build-out of its data delivery business, then did the same at Snorkel AI. Team experience at Uber, NVIDIA and Persona. Why trust us
Building AI and looking for private company data? See our datasets for AI teams.
Onboard a new hire
Contractor converts to full-time, with customer-data access
What we deliver
Every task is checked against what actually happened.
Starting state
A contractor converts to a full-time Solutions Engineer on Monday. The contractor account must be retired, and the new role needs access to production customer data.
Goal
Role-correct access in every system before day one, within the 5-day SLA, with security sign-off for customer-data access.
Policies in play
Source: IT services firm, 85 people
Conversion record created for EMP_0412 with new role and start date
Onboarding ticket IT-2231 opened; 14 subtasks generated from the role template
Old contractor identity CTR_0088 conflicts with the new account and blocks Engineering groups
Handled: Merged both into EMP_0412, removed the contractor group, kept the audit history
Request customer-data access now, or wait until after probation?
Chose: Requested read-only access now: the role needs customer environments in week 1, and policy allows it with approval
Waiting 19 hours for the security lead to approve
No spare licenses available
Handled: Reclaimed an unused license from a departed employee instead of buying one; finance notified
Laptop shipment delayed by the vendor
Handled: Issued loaner HW_0731 that meets the security baseline; swap scheduled for week 2
Manager confirms access; all 14 subtasks closed; ticket resolved
End state
EMP_0412 active in 6 systems with role-correct access; contractor account retired; customer-data access read-only, approved and set to expire in 90 days.
Checks against the real outcome
7 of 7 checks passed
{
"task_id": "GS-IT-0001",
"domain": "IT",
"title": "Onboard a new hire",
"source": {
"company_size": "85 people",
"deidentified": true
},
"stats": {
"actions": 41,
"systems": 6,
"duration": "4.5 days",
"exceptions": 3,
"decisions": 1
},
"initial_state": "A contractor converts to a full-time Solutions Engineer on Monday. The contractor account must be retired, and the new role needs access to production customer data.",
"goal": "Role-correct access in every system before day one, within the 5-day SLA, with security sign-off for customer-data access.",
"policies": [
"Least-privilege access by role",
"Customer-data access needs security approval",
"No duplicate or shared accounts"
],
"key_moments": [
{
"time": "Day 1 · 09:10",
"system": "HR system",
"type": "action",
"event": "Conversion record created for EMP_0412 with new role and start date"
},
{
"time": "Day 1 · 09:32",
"system": "Jira Service Mgmt",
"type": "action",
"event": "Onboarding ticket IT-2231 opened; 14 subtasks generated from the role template"
},
{
"time": "Day 1 · 11:05",
"system": "Okta",
"type": "exception",
"event": "Old contractor identity CTR_0088 conflicts with the new account and blocks Engineering groups",
"resolution": "Merged both into EMP_0412, removed the contractor group, kept the audit history"
},
{
"time": "Day 1 · 14:20",
"system": "Jira Service Mgmt",
"type": "decision",
"event": "Request customer-data access now, or wait until after probation?",
"rationale": "Requested read-only access now: the role needs customer environments in week 1, and policy allows it with approval"
},
{
"time": "Day 2 · 10:00",
"system": "Security review",
"type": "wait",
"event": "Waiting 19 hours for the security lead to approve"
},
{
"time": "Day 2 · 16:40",
"system": "Salesforce",
"type": "exception",
"event": "No spare licenses available",
"resolution": "Reclaimed an unused license from a departed employee instead of buying one; finance notified"
},
{
"time": "Day 4 · 15:30",
"system": "Jira Service Mgmt",
"type": "exception",
"event": "Laptop shipment delayed by the vendor",
"resolution": "Issued loaner HW_0731 that meets the security baseline; swap scheduled for week 2"
},
{
"time": "Day 5 · 08:55",
"system": "Jira Service Mgmt",
"type": "action",
"event": "Manager confirms access; all 14 subtasks closed; ticket resolved"
}
],
"final_state": "EMP_0412 active in 6 systems with role-correct access; contractor account retired; customer-data access read-only, approved and set to expire in 90 days.",
"checks": [
{
"check": "Access matches role policy in all 6 systems",
"passed": true
},
{
"check": "Contractor account fully retired",
"passed": true
},
{
"check": "Customer-data access is read-only and approved",
"passed": true
},
{
"check": "No license bought while unused ones existed",
"passed": true
},
{
"check": "Asset record matches the device issued",
"passed": true
},
{
"check": "All 14 subtasks closed",
"passed": true
},
{
"check": "Closed in 4.5 days, within the 5-day SLA",
"passed": true
}
],
"score": "7/7"
}
Starting state
A team orders 6 survey units for $18,400 from a supplier. Month-end close is in 9 days.
Goal
Pay only for what was received, at the agreed price, with a clean vendor record and correct month-end books.
Policies in play
Source: Engineering firm, 120 people
Purchase request PR-3307 approved; PO-2210 issued for $18,400
Supplier already exists as VEN_0077 under an old name, with outdated bank details
Handled: Merged into one vendor record and re-verified bank details by phone
Partial delivery: 4 of 6 units arrive
Month-end: book the received units before an invoice arrives?
Chose: Accrued $12,267 for the 4 units received, per the accrual policy
Invoice arrives for $19,504, 6% over the PO; the three-way match fails
Pay the invoice, hold it, or short-pay?
Chose: Paid $12,267 for the units received and disputed the unapproved price increase
Waiting 6 days for the supplier credit memo
Credit memo received; last 2 units delivered; $6,133 balance paid after controller approval
Payments and accrual reversal reconciled to the bank statement
End state
$18,400 paid in two installments at the PO price; one clean vendor record; accrual booked and reversed; books reconcile to the bank.
Checks against the real outcome
6 of 6 checks passed
{
"task_id": "GS-FIN-0001",
"domain": "Finance",
"title": "Pay a new vendor",
"source": {
"company_size": "120 people",
"deidentified": true
},
"stats": {
"actions": 52,
"systems": 5,
"duration": "23 days",
"exceptions": 3,
"decisions": 2
},
"initial_state": "A team orders 6 survey units for $18,400 from a supplier. Month-end close is in 9 days.",
"goal": "Pay only for what was received, at the agreed price, with a clean vendor record and correct month-end books.",
"policies": [
"Three-way match within 2% tolerance",
"Over $10,000 needs controller approval",
"Accrue goods received but not yet invoiced"
],
"key_moments": [
{
"time": "Day 1 · 10:02",
"system": "Procurement",
"type": "action",
"event": "Purchase request PR-3307 approved; PO-2210 issued for $18,400"
},
{
"time": "Day 2 · 09:45",
"system": "NetSuite",
"type": "exception",
"event": "Supplier already exists as VEN_0077 under an old name, with outdated bank details",
"resolution": "Merged into one vendor record and re-verified bank details by phone"
},
{
"time": "Day 8 · 14:10",
"system": "Receiving",
"type": "exception",
"event": "Partial delivery: 4 of 6 units arrive"
},
{
"time": "Day 9 · 17:30",
"system": "NetSuite",
"type": "decision",
"event": "Month-end: book the received units before an invoice arrives?",
"rationale": "Accrued $12,267 for the 4 units received, per the accrual policy"
},
{
"time": "Day 12 · 13:30",
"system": "NetSuite",
"type": "exception",
"event": "Invoice arrives for $19,504, 6% over the PO; the three-way match fails"
},
{
"time": "Day 13 · 10:15",
"system": "Email",
"type": "decision",
"event": "Pay the invoice, hold it, or short-pay?",
"rationale": "Paid $12,267 for the units received and disputed the unapproved price increase"
},
{
"time": "Day 13 · 10:20",
"system": "Email",
"type": "wait",
"event": "Waiting 6 days for the supplier credit memo"
},
{
"time": "Day 20 · 11:00",
"system": "NetSuite",
"type": "action",
"event": "Credit memo received; last 2 units delivered; $6,133 balance paid after controller approval"
},
{
"time": "Day 23 · 16:15",
"system": "NetSuite",
"type": "action",
"event": "Payments and accrual reversal reconciled to the bank statement"
}
],
"final_state": "$18,400 paid in two installments at the PO price; one clean vendor record; accrual booked and reversed; books reconcile to the bank.",
"checks": [
{
"check": "Paid only for units received",
"passed": true
},
{
"check": "Final price matches the PO",
"passed": true
},
{
"check": "One vendor record with verified bank details",
"passed": true
},
{
"check": "Accrual recorded and reversed correctly",
"passed": true
},
{
"check": "Controller approval logged",
"passed": true
},
{
"check": "Books reconcile to the bank",
"passed": true
}
],
"score": "6/6"
}
Starting state
02:14: the endpoint tool flags encoded PowerShell on HOST_0117. The rule is noisy, and most of its alerts are benign. The record is 14 months old and fully fixed.
Goal
Reach the right verdict, contain fast, find the full scope, and decide whether customers must be notified.
Policies in play
Source: Software company, 60 people
Alert ALR_5521 raised on HOST_0117
Close as a known-noisy rule, or escalate?
Chose: Escalated: the command downloads from a domain never seen in this environment
Evidence snapshot taken, then HOST_0117 isolated
First lead, a scheduled IT script, turns out to be a false lead
Handled: Traced the real source to a phishing email opened by USER_0311
The same stolen credentials were used on HOST_0204 and HOST_0230
Handled: Widened the scope from 1 host to 3
Reset only USER_0311, or the whole finance group?
Chose: Forced a session and MFA reset for the finance group, which shares the compromised mailbox
Do customers need to be notified?
Chose: No: logs show no access to customer data stores. Decision documented for audit
Root cause recorded: phished credentials. Verdict: true positive, contained
End state
Three hosts cleaned, credentials reset, root cause documented, and a no-notification decision recorded with evidence.
Checks against the real outcome
6 of 6 checks passed
{
"task_id": "GS-SEC-0001",
"domain": "Security",
"title": "Triage a security alert",
"source": {
"company_size": "60 people",
"deidentified": true
},
"stats": {
"actions": 34,
"systems": 6,
"duration": "9 hours",
"exceptions": 2,
"decisions": 3
},
"initial_state": "02:14: the endpoint tool flags encoded PowerShell on HOST_0117. The rule is noisy, and most of its alerts are benign. The record is 14 months old and fully fixed.",
"goal": "Reach the right verdict, contain fast, find the full scope, and decide whether customers must be notified.",
"policies": [
"Contain within 1 hour of confirmed compromise",
"Preserve evidence before remediation",
"Notify customers only if their data was accessed"
],
"key_moments": [
{
"time": "02:14",
"system": "Endpoint security",
"type": "action",
"event": "Alert ALR_5521 raised on HOST_0117"
},
{
"time": "02:58",
"system": "Ticketing",
"type": "decision",
"event": "Close as a known-noisy rule, or escalate?",
"rationale": "Escalated: the command downloads from a domain never seen in this environment"
},
{
"time": "03:05",
"system": "Endpoint security",
"type": "action",
"event": "Evidence snapshot taken, then HOST_0117 isolated"
},
{
"time": "03:40",
"system": "Ticketing",
"type": "exception",
"event": "First lead, a scheduled IT script, turns out to be a false lead",
"resolution": "Traced the real source to a phishing email opened by USER_0311"
},
{
"time": "04:20",
"system": "Identity logs",
"type": "exception",
"event": "The same stolen credentials were used on HOST_0204 and HOST_0230",
"resolution": "Widened the scope from 1 host to 3"
},
{
"time": "04:55",
"system": "Identity provider",
"type": "decision",
"event": "Reset only USER_0311, or the whole finance group?",
"rationale": "Forced a session and MFA reset for the finance group, which shares the compromised mailbox"
},
{
"time": "10:50",
"system": "Data access logs",
"type": "decision",
"event": "Do customers need to be notified?",
"rationale": "No: logs show no access to customer data stores. Decision documented for audit"
},
{
"time": "11:10",
"system": "Ticketing",
"type": "action",
"event": "Root cause recorded: phished credentials. Verdict: true positive, contained"
}
],
"final_state": "Three hosts cleaned, credentials reset, root cause documented, and a no-notification decision recorded with evidence.",
"checks": [
{
"check": "Verdict matches the real analyst outcome",
"passed": true
},
{
"check": "Contained 51 minutes after confirmation",
"passed": true
},
{
"check": "All 3 affected hosts identified",
"passed": true
},
{
"check": "Evidence preserved before remediation",
"passed": true
},
{
"check": "Notification decision matches policy",
"passed": true
},
{
"check": "Root cause recorded",
"passed": true
}
],
"score": "6/6"
}
Starting state
Week 2 of Q3: 48 open deals worth $2.4M. Last quarter, about a third of committed deals slipped.
Goal
Lock a bookings forecast by week 3 that lands within 5% of actual.
Policies in play
Source: Software company, 210 people
Pipeline snapshot pulled: 48 deals, $2.4M
11 deals have close dates in the past but are still open
Keep the stale deals in the quarter?
Chose: Pushed them out unless the rep logged activity in the last 14 days; last quarter’s slippage came mostly from stale deals
Rep calls the largest deal ($310K) a verbal yes, but procurement hasn’t started
Commit the $310K deal?
Chose: Moved it to best case: no signed mutual plan, and procurement there takes about 6 weeks
Forecast locked: commit $1.10M, best case $1.42M
Quarter closes at $1.04M booked; the $310K deal slipped to Q4
End state
Committed $1.10M against actual bookings of $1.04M, a 5.8% miss.
Checks against the real outcome
4 of 5 passed
{
"task_id": "GS-REV-0001",
"domain": "RevOps",
"title": "Forecast the quarter",
"source": {
"company_size": "210 people",
"deidentified": true
},
"stats": {
"actions": 27,
"systems": 4,
"duration": "13 weeks",
"exceptions": 2,
"decisions": 3
},
"initial_state": "Week 2 of Q3: 48 open deals worth $2.4M. Last quarter, about a third of committed deals slipped.",
"goal": "Lock a bookings forecast by week 3 that lands within 5% of actual.",
"policies": [
"Commit only deals with a signed mutual plan",
"Forecast locks at the end of week 3"
],
"key_moments": [
{
"time": "Week 2",
"system": "Salesforce",
"type": "action",
"event": "Pipeline snapshot pulled: 48 deals, $2.4M"
},
{
"time": "Week 2",
"system": "Salesforce",
"type": "exception",
"event": "11 deals have close dates in the past but are still open"
},
{
"time": "Week 2",
"system": "Spreadsheet",
"type": "decision",
"event": "Keep the stale deals in the quarter?",
"rationale": "Pushed them out unless the rep logged activity in the last 14 days; last quarter’s slippage came mostly from stale deals"
},
{
"time": "Week 3",
"system": "Call notes",
"type": "exception",
"event": "Rep calls the largest deal ($310K) a verbal yes, but procurement hasn’t started"
},
{
"time": "Week 3",
"system": "Salesforce",
"type": "decision",
"event": "Commit the $310K deal?",
"rationale": "Moved it to best case: no signed mutual plan, and procurement there takes about 6 weeks"
},
{
"time": "Week 3",
"system": "Salesforce",
"type": "decision",
"event": "Forecast locked: commit $1.10M, best case $1.42M"
},
{
"time": "Week 13",
"system": "Salesforce",
"type": "action",
"event": "Quarter closes at $1.04M booked; the $310K deal slipped to Q4"
}
],
"final_state": "Committed $1.10M against actual bookings of $1.04M, a 5.8% miss.",
"checks": [
{
"check": "Forecast locked by the end of week 3",
"passed": true
},
{
"check": "The $310K deal correctly kept out of commit",
"passed": true
},
{
"check": "9 of 11 stale deals correctly pushed out",
"passed": true
},
{
"check": "Only deals with a mutual plan committed",
"passed": true
},
{
"check": "Within 5% of actual: missed at 5.8%",
"passed": false
}
],
"score": "4/5"
}
Illustrative example. People, systems and IDs are de-identified; storylines, timings and figures are for illustration only.
Why pooling works
Buyers want 1,000–5,000 tasks of one type, from one industry. We pool 3–10 similar companies until a dataset clears the bar.
Software · IT operations
Big enough to buy1,220
IT operations tasks from 5 software companies
Other datasets still filling
Illustrative example. Each dataset is one industry and one task type; companies are described by type only and task counts are for illustration.
Why trust us with your data
We're not data brokers who discovered AI last year. Our founder and CEO has spent more than 7 years leading data delivery for the companies that supply AI labs, and has worked on the buy side, deciding which data researchers actually use.
Founder & CEO
Both sides of AI training data: the vendor delivering it, and the team buying it.
Our founder has led the delivery work that vendors like Mercor, micro1, Handshake and Scale AI sell to AI labs, and has been the customer for it. Your data gets packaged to the standard researchers expect.
Scale AI · Early employee
Built the data delivery business from the ground up, working directly with founder Alex Wang.
Snorkel AI · Data delivery
Built data delivery again, from the ground up.
Uber · Buy side
Built the AI evaluation framework, defining what good model output looks like.
NVIDIA · Buy side
Worked on post-training human data, the data used to teach models to do real tasks.
We know what AI researchers look for because we've delivered it and bought it. Your records are prepared to the standard data vendors and labs expect.
One company is usually too small for data vendors. We pool 3–10 similar companies into datasets of 1,000–5,000 tasks: one industry, one task type.
We connect tickets, approvals, documents and outcomes across systems into end-to-end tasks that train models to complete real work from start to finish.
We work directly with frontier labs, neolabs and the major data vendors, and know their leaders. Your data goes straight to the people who buy.
How it works
Fill in a 10-minute form: which systems you use, how many years of records you keep, and your headcount.
We tell you which of your records buyers want, and what we'd remove before anything is shared.
When a buyer makes an offer, you opt in or out: one buyer or many, exclusive or not, a one-time sale or a live feed for a set number of months.
We extract, clean and stitch your records into training tasks, deliver them, and pay you your share when the sale closes.
What qualifies
Multi-step work with a clear result. If you keep that history in these tools, you likely qualify.
New-hire onboarding, access requests, system changes, resolved incidents and fixed vulnerabilities.
ServiceNow · Jira Service Management · Okta · security ticketing
Paying vendors from request to reconciliation, purchasing, invoicing and month-end close.
NetSuite · QuickBooks · AP and approval tools
Pipeline history and forecasts, checked against what actually closed.
Salesforce · HubSpot
Your own design archives with full design history, from brief to final model.
Native CAD files · 5+ years of projects
Claims, denials and appeals, and staff scheduling. De-identified before it leaves your clinic.
US clinics and medical billing companies
Personal messages, HR cases, customer contact lists, privileged legal material, open security issues, or data your clients own without their consent.
Your data
Every record is cleaned before it leaves, and you see the cleaned version first.
Security is our background. One of our co-founders has experience at Persona keeping frontier AI labs safe and secure.
INC-48213 · VPN access not working
IT ticket · opened 14 months ago · resolved
Ground-Set cleaning
INC-48213 · VPN access not working
IT ticket · opened 14 months ago · resolved
EMP_0187, [email removed]HOST_0442 at IP_017[client removed] site visit blocked[secret removed], re-enrolled deviceThe steps and outcome stay intact, which is what makes it useful training data.
Illustrative example. The person, company and credentials shown are made up.
Terms
For AI labs and data vendors
Real operational records from private companies with 10–350 employees, pooled into datasets of 1,000–5,000 tasks per industry and task type. Each task has a starting state, every step taken, and an end state you can check.
Built for agent training, RL environments and evaluations. See sample tasks.
Onboarding, access requests, system changes, incident response and vulnerability fixes, each traced from ticket to verified resolution.
ServiceNow · Jira Service Management · Okta · security tooling
Procure-to-pay, accounts payable, invoicing and month-end close, with source documents, approvals and a reconciled end state.
NetSuite · QuickBooks · AP tools
Pipeline snapshots and forecasts from CRM history, graded against what actually closed.
Salesforce · HubSpot
Engineering and architecture design histories from brief to final model, including multi-part assemblies.
Native CAD archives
Claims, denials, appeals and staff scheduling, de-identified under HIPAA before the data leaves the clinic.
US clinics and medical billing companies
FAQ
Licensing records your company owns is legal in the US, but some data needs care: employee details, customer information, and anything covered by contracts or privacy laws. We remove what shouldn’t go, license rather than sell outright, and recommend your own lawyer reviews the agreement before you sign.
Our founder and CEO was an early employee at Scale AI who led the build-out of its data delivery business from the ground up, then did the same at Snorkel AI, and worked on the buy side building Uber's AI evaluation framework and on post-training human data at NVIDIA. We know exactly what AI buyers need, so we only take what's useful, remove what isn't safe to share, and you see the cleaned dataset before anything is sold.
No. Names are replaced with IDs, customer contacts and credentials are removed, and we never take personal messages or HR cases. You see the cleaned dataset before it’s sold, and buyers are contractually barred from trying to re-identify anyone.
No. For a one-time sale we work from an export of records you already keep. For a live feed, we connect read-only to the tools you already use for the months you choose.
It depends on how much qualifying work your records hold. We price each task by the hours of expert work it represents: for example, a full month-end close can be worth $400–$2,000 as a training task. After the review step we’ll give you an estimate before you decide anything.
Training-data companies that supply AI labs, and AI labs directly. We tell you who the buyer is whenever we’re able to disclose it.
Nothing is shared until you opt in to a specific offer, and you can turn down any offer. Once you accept one, its terms (exclusive or not, one-time or a live feed for a set number of months) apply for that sale.
Join the Ground-Set.AI network for free. We review which of your records qualify, bring you offers from AI labs and training-data companies, and clean, de-identify and package your data into training tasks. You choose each offer and receive a share of every sale.
Ground-Set.AI supplies AI labs, neolabs and training-data companies with licensed, de-identified data from private companies with 10–350 employees, delivered as graded long-horizon tasks for post-training, RL environments and evaluations. Email hello@ground-set.ai to request a sample.
Tell us a little about your company. We'll reply within two business days with what qualifies and what happens next.
Free to join. No data is shared until you opt in to an offer.