Closing your company? Your data can still earn. License it to AI labs before you shut down.See what it's worth

Where models practice real work.

We build evaluations, expert data and RL environments from how professionals and companies actually work.

For AI labsLicense your data
20,000+finance professionals in our network
300accounting firms interviewed
40expert partner firms
10+AI labs in conversation
What we build

From finding the failure to fixing it.

Four products, one loop. Labs start at any step, or hand us the whole thing.

01 · EVALUATE

See exactly where an agent fails.

We write tasks the way professionals do the work, then grade every step.

  • Written by practitioners who do the job daily
  • Graded by expert rubrics and automatic checks
  • Re-run on every new model and checkpoint
How we evaluate
Eval run · agent under testGrading
TASK 0412 · FINANCE OPS

Reconcile March vendor payments for a 45-person logistics company in NetSuite. Flag duplicates and tie the AP sub-ledger to the general ledger.

  • ✓Opened the right periodstep 2
  • ✓Matched 212 of 214 paymentsstep 9
  • ✕Missed duplicate invoice INV-20931step 14
  • ✓Followed approval policystep 17
  • ✕Sub-ledger off by ₹48,200step 22
  • ✓Wrote a clear audit notestep 26
Score0 / 6
02 · CURATE

Capture how experts actually decide.

Verified professionals solve real tasks in our sandbox. Every step and judgment call is recorded.

  • Step-by-step trajectories with reasoning
  • Preference pairs and rubrics for RLHF
  • Reviewed by a senior expert before delivery
How we curate
Trajectory · task 0412Recording
AR
Arjun R.Chartered Accountant, 9 years
Verified
  1. Open AP ageing reportNetSuite · Reports
  2. Match payments to invoices212 of 214 matched
  3. Flag INV-20931 as a duplicateJudgment: same lane, date and amount
  4. Post credit note CN-118Journal entry with reason
  5. Tie sub-ledger to GLDifference 0.00
Steps31
Time9m 12s
ReviewPassed
03 · TRAIN

Give agents a place to practice.

Sandboxed copies of real business software with realistic data. Agents act thousands of times and earn rewards they can't game.

  • Built from licensed company data, not toy examples
  • Verifiers check the end state and the path
  • Delivered over API and MCP, refreshed quarterly
How environments work
Environment · kestrel-ops-erpTraining
TOOLS · 14
create_poapprove_invoicepost_journalsearch_vendorrun_reportemail_vendor+8
EPISODES0MEAN REWARD0.21
018,400 episodes
  • End state matches gold ledgerPass
  • Approval policy respectedPass
  • No duplicate paymentsPass
Example training run
04 · SOURCE

Learn from how companies really work.

Companies license years of tickets, docs and chat. We remove personal details and turn real workflows into training data.

  • Read-only connectors to the tools teams use
  • Names, emails, phones and customers removed
  • Every company approves a sample first
Data partnerships
Slack · #ops-escalationsDe-identifying

Ritika Menon: Shipment for Ashok Leyland is stuck at customs. They need the revised packing list by 4pm.

Dev Kapoor: Re-issue it from the ERP and copy dev.k@kestrelfreight.in. Call +91 98201 44718 if the broker pushes back.

Ritika Menon: Done. Moving the SLA to T+2 and tagging the delay as documentation.

Sources5
Records266,625
Removed0
Example from our demo company
Domains

Real tasks from real jobs.

We start where answers can be checked and our expert network is deepest.

Finance and accounting

Close, reconciliations, audit, tax.

EXAMPLE TASKTie the AP sub-ledger to the GL after a duplicate payment.
Software engineering

Repo-level bugs, migrations, review.

EXAMPLE TASKMove a payments service to gRPC without breaking 41 tests.
Law and contracts

Redlines, research, compliance.

EXAMPLE TASKRedline a vendor MSA against the company's playbook.
Healthcare

Coding, billing, chart review.

EXAMPLE TASKCode a discharge summary to ICD-10 and flag missing notes.
Customer support

Tickets, refunds, escalations.

EXAMPLE TASKResolve a double-billing ticket across Zendesk and Stripe.
Sales and CRM

Pipelines, forecasts, account plans.

EXAMPLE TASKClean a HubSpot pipeline and forecast the quarter.
Operations and logistics

Procurement, fulfilment, inventory.

EXAMPLE TASKRe-route 12 delayed shipments within carrier SLAs.
Indic languages

Hindi, Tamil, Bengali and more.

EXAMPLE TASKRead a Hindi insurance claim and extract the payable amount.
Provably human

Data you can trust was made by people.

Contractors are quietly passing model output off as human work. Every task we deliver is done in our sandbox, where human work leaves a signature.

  • Locked sandbox. External paste and outside tools are blocked and logged.
  • Verified people. ID, credentials and a paid test task.
  • Screened output. Compared with model answers before delivery.
Provenance check · two submissionsTyping rhythm
Arjun R.
Human
Contractor X
Flagged
2,400 characters arrived in 0.3 secondsPasted
91% match with a frontier model's answerModel-written
Illustration of our checks
Data partnerships

How much is your company's data worth?

Operating, winding down or already closed. Get a range in seconds.

Headquarters
Estimated payout range
$0to$0

Indicative only. The final amount depends on the tools you used, data quality and lab demand.

Get a detailed estimateHow it works