Carry the context
Connect the case, account history and current policy across the systems your team uses.
Human experience.
Autonomous
operators.
Your best people know how your business works. Refinery turns that institutional memory into an autonomous workforce that carries the work forward.
Built for mid-market services. Starts with a defined responsibility and progresses toward the connected work of entire departments.
A new unit of capacity
AI takes charge
of outcomes.
Refinery builds AI operators that own the work: understand the case, act according to judgment, follow through to the best outcome.
Exercise judgment
Apply the decisions your experts approve, with explicit limits and a route for exceptions.
Close the loop
Our direction: operators that track unfinished work, coordinate handoffs and return when new information arrives.
Explore Refinery
From a workflow
to a working operator.
Walk through a customer-care example: define the work, connect systems and see how expert decisions become a reviewed policy.
Open full demo ↗Interactive prototype with sample data. Runs locally in your browser; entries are not sent anywhere.
Inside Refinery
Capture the judgment.
Build the operator.
Institutional knowledge lives in how people work. Refinery makes that expertise explicit, testable, and actionable.
IBuild the training signal
Learn how your
best people decide
01 / Reconstruct the case
Join recordings, tickets and system records. Rebuild what the expert knew at the time.
02 / Extract actions & reasons
Map the sequence of actions. Ask experts about exceptions, conflicts and missing context.
OUTPUT Grounded cases + approved action sets
IILearn the policy
Refine evidence
into a rulebook
03 / Learn and refine the rules
Draft rules→Practice cases↓Expert review←Trace errors↑
Generate new rules supported by cases. Prune for accuracy and human readability. Measure how revision cycles improve accuracy.
BOUNDARY Temporally consistent generation and validation
IIITest and deploy
Earn the right
to act
04 / Test on unseen work
Check actions, policy compliance and outcomes against held-out cases and a strong baseline.
05 / Review, then widen scope
Start in draft mode. People approve actions. Expand autonomy one action type at a time as evidence supports it.
CONTROL Permissions enforced outside the model
Reviewer corrections → labelled examples → next tested policy version
Explore the technical controls +
Decision-time grounding
Use the records and policy version available at the original decision. Split cases by time to prevent hindsight from inflating results.
Evidence-backed policies
Each rule cites supporting cases. The main rulebook is kept short enough for a person to read and approve, with evidence in appendices.
Independent checks
Pair deterministic action and permission checks with model judges calibrated to human grades. Score what happened, not just how a reply reads.
Versioned improvement
Trace edits and rejections to their rule and policy version. Re-test corrections before release and return regressing action types to review.
The direction
The knowledge of workers.
The capacity of an
autonomous workforce.
We are building toward entire departments of connected operators: carrying context, coordinating decisions and taking responsibility for outcomes. A business’s expertise becomes capacity it can keep building on.
Start with one responsibility.
Prove it in practice.
Expand what the workforce can own.
Put your expertise to work
What could your team
hand over?
Bring us a recurring workflow, the systems it touches, and the person who knows it best.