Sign in Book demo →
Platform Research Use cases Docs About
katyar/Use cases
Use cases

The tasks your agents repeat every day. Own each one.

Most agent work is the same few tasks, run thousands of times. Those are exactly the tasks an open model can learn from your own verified data — one at a time, with the frontier model as the safety net.

Industries7
Example use cases8
Best fithigh volume · checkable outcome

Pick an industry. See what moves over.

01 · by industry
8 use cases · parity illustrative
Support owned

Refund triage

Decide whether a refund is due, issue it, and tell the customer.

What gets checked
  • Refund tool called with valid argumentslibrary
  • Money actually moved in the ledgerlibrary
  • Inside your refund policy windowyours
Your model takes over

Routine refunds inside policy.

Frontier keeps

Disputes, fraud flags, exceptions.

past the parity linehandoff once it matches
Support shadow

Ticket reply drafting

Draft the first reply to a customer ticket from your help center and order data.

What gets checked
  • Answer matches a help-center sourcelibrary
  • No personal data in the replylibrary
  • Customer didn't reopen within 7 daysyours
Your model takes over

Order status, returns, account how-tos.

Frontier keeps

Escalations and multi-issue threads.

closing on the parity linehandoff once it matches
Engineering shadow

Code review & test fixing

Fix failing tests and apply review feedback on routine pull requests.

What gets checked
  • Test suite passes in a sandboxlibrary
  • Diff builds and lints cleanlylibrary
  • Follows your repo conventionsyours
Your model takes over

Flaky tests, lint fixes, dependency bumps.

Frontier keeps

Architecture and cross-service changes.

closing on the parity linehandoff once it matches
Data owned

SQL reporting

Turn a plain-language question into a query and a number people can trust.

What gets checked
  • Query runs against the warehouselibrary
  • Result matches the referencelibrary
  • Uses your approved metric definitionsyours
Your model takes over

Weekly KPI pulls and standard breakdowns.

Frontier keeps

New analyses with no reference yet.

past the parity linehandoff once it matches
Finance training

Invoice reconciliation

Match invoices to purchase orders and payments, and flag what doesn't add up.

What gets checked
  • Totals reconcile to the ledgerlibrary
  • Vendor exists in your master datayours
  • Tax lines follow your rulesyours
Your model takes over

Clean matches with standard vendors.

Frontier keeps

Partial payments and disputed invoices.

still traininghandoff once it matches
Life sciences training

Assay protocol planning

Draft a lab protocol from an approved template, ready for a scientist to sign off.

What gets checked
  • Reagents are in stock in your LIMSyours
  • Temperatures stay within your SOPyours
  • Samples exist and are releasedyours
Your model takes over

Routine plans from approved templates.

Frontier keeps

Novel protocols — sent to a scientist.

still traininghandoff once it matches
Insurance shadow

Claims decisions

Recommend a decision and payout on a claim, with the reasoning an adjuster would give.

What gets checked
  • Required documents are presentlibrary
  • Coverage applied the way your adjusters apply ityours
  • Payout within policy limitsyours
Your model takes over

Straightforward auto and property claims.

Frontier keeps

Large losses and suspected fraud.

closing on the parity linehandoff once it matches
Legal training

Contract clause review

First-pass review of incoming contracts against your negotiation playbook.

What gets checked
  • Every cited clause exists in the documentlibrary
  • Flags match your playbook positionsyours
  • No privileged data leaves the reviewlibrary
Your model takes over

NDAs and standard MSA first passes.

Frontier keeps

Bespoke deals and sensitive language.

still traininghandoff once it matches

From frontier to owned, one task at a time.

02 · the path
01

Collect

Your agents keep running on the frontier model. Every call, tool call and outcome is captured and redacted in your cloud.

CC · gateway connector
02

Verify

Library and custom checks decide which answers were actually right. Only those become golden examples.

CT · verifiers
03

Train

An open model learns the task from the golden set, rewarded by the same checks.

CT · RL on your compute
04

Hand off

Once it matches the frontier on that task, traffic moves over. If it slips, the task moves back.

Parity gate

Is your task a good fit?

03 · fit check

A good fit

  • ✓
    High volumeHundreds or thousands of runs a week, so there's plenty to learn from.
  • ✓
    An outcome you can checkA ledger entry, a passing test, a number that matches.
  • ✓
    Stable rulesPolicies change, but not every day.
  • ✓
    A safe fallbackWhen the model isn't sure, the frontier or a person takes it.

Not yet

  • –
    Open-ended creative workNo single right answer to check against.
  • –
    No ground truthNothing in your systems confirms the outcome.
  • –
    Rare, one-off tasksToo few examples to learn from, for now.

Not sure where your task lands? Send it to us and we'll tell you honestly.

Check my task →
Start with one agent

Point us at one agent. We'll show you its parity curve.

You get a first golden set, the verifiers that built it, and an honest read on which tasks your own model can take over.

01One agent, one task family, fixed length.
02Connector in your VPC. Redaction before anything moves.
03You keep the golden set, the verifiers and the weights.
04Your frontier model stays in place until parity is proven.