How Kickback uses AI
Kickback uses Claude in four places: the Designer, Guard explanations, Scribe and the Analyst. Each one proposes or explains. Deterministic code checks the result, and nothing an AI writes can send a refund on its own.
The short version
- Every number on a program proposal comes from the same reward engine that decides eligibility.
- Risk scores come from fixed signals; an explanation cannot change a score or approve a refund.
- Without AI access, presets, template copy and Guard keep working.
Designer
Why these numbers?
Claude proposes three sets of parameters from aggregate purchase history and your goal. The same deterministic reward engine used for eligibility backtests every proposal. It calculates cost, who could earn, first-reward timing, budget limits and break-even lift.
Backtests describe past purchases. They do not establish that a reward changes future behavior. Gross margin and the lift slider are assumptions you can edit.
Proposals exceeding 5% of qualifying revenue or a 15% refund count ratio are rejected. Revenue cost above 3% is highlighted for review. Per-reward and budget ceilings apply independently. Synthetic and replay history are labeled.
Every number in the explanation is checked against parameters and computed metrics. The figures displayed on each card come directly from the simulator. You can edit a proposed rule; it is checked and simulated again before activation.
The weekly idea on your Overview comes from the same Designer and passes the same checks.
Guard
Guard routes rewards using deterministic signals. An explanation cannot change a score or approve a refund. Merchant approval still passes customer, dispute, pause, budget and refundable-capture checks.
Scribe
Scribe proposes factual thank-you copy. Length, characters, links, prohibited promotional terms and unsupported numbers are checked in code. Fixed terms stay in place.
Analyst
The Analyst answers questions in the Cockpit. Relevant program rules, pseudonymous customer codes, metrics and chart instructions are sent to Anthropic to answer them. The Analyst cannot move money or change a program.
What AI can never do
- Send, approve or cancel a refund.
- Change a risk score.
- Activate a program. A proposed rule is checked and simulated again, and you activate it.
- Put a number on a proposal card. Those come directly from the simulator.
Without AI
Without AI access, manual presets, template copy and Guard continue working.