The reward layer for customer-facing AI
Graders and judges score how a reply looked. Churney scores what it caused: the renewal, the repeat order, the churn that never happened. Predicted from your data, early enough to train on.
Trusted by 30+ teams to set the reward for Meta, Google and TikTok bidding
Why outcomes
Train against a judge and you get the replies the judge likes: more apologies, more discounts, more escalations. Train against the outcome and the model learns what keeps the customer.
Simulated training run. An accurate judge is still the wrong objective.
The platform
Connect your warehouse, score every response by the value it creates, and train against it. On the platform you already use, with a research engineer on your team.
Talk to a research engineerThe vision
A reward is only useful while it stays true. We are building the loop that keeps it true as your product and customers change.
Predicted long-term value as the reward for any training run or agent decision.
Every reward is checked against the outcome it predicted and retrained, so it holds up under optimization.
Train on the interactions and outcomes you already hold, corrected for how past decisions were made.
Proof
Meta, Google and TikTok run black-box bidding models that learn from the conversion signal you send. Since 2022, for 30+ brands, that signal has been ours.
“By focusing on predicted long-term value instead of early conversions, we doubled new customer volume and improved ROAS by 190%.”Kaveh Motamed, Sr. Director of Marketing, Zapier
ROAS uplift across 30+ brands
When our predicted value replaced the conversion signal on Meta, Google and TikTok.
Move accuracy in our chess experiment, where a model tuned on a predictive reward beat a frontier model five times its price.
Gap between a tuned open 27B model and a tuned frontier model on BIRD text-to-SQL, in our RFT study.
Training platforms we have run the same RFT tasks on: Vertex AI, Together, Fireworks, Tinker and SageMaker.
Predictions run in your tenant or warehouse. Training sees prompts and scores, never raw customer records.
Start with approved and synthetic data. Outcomes come in only where you choose.
Every candidate passes your evals and your approval before it sees traffic.
