The decision platform for AI
The judgmentlayer forsoftware
Refunds, disputes, approvals, agent actions. Every product makes calls that rules can’t settle. Show Moonbounce how your team decides them, and it makes the same call from one API request, in milliseconds.
“The heel came off my boots after a week of normal wear.”
- Does the customer describe a defect?0.91
- Has the item been used beyond trying it on?0.86
- Is it more than 30 days after delivery?0.08
Grant
Example evaluations. Each shows the case, how its questions were answered, and the outcome.
Software can execute rules.It can’t exercise judgment.
Every business makes decisions that depend on context, intent and exceptions.
Software struggles with these. So they pile up in queues, wait for a person, and get decided a little differently every time.
Refund
Technically late. Obviously fair.
The rule says no
Dispute
Both sides are half right.
No rule fits
Agent action
Almost within policy. Not quite.
The rule says yes
Always yesRules handle these.
It dependsThese wait for a person. This is where Moonbounce works.
Always noRules handle these.
What companies do today
Until now, there have been four bad options.
Hire people to decide
Accurate on a good day, slow on every day. Cost scales with volume. Consistency doesn’t.
Slow, costly, inconsistent
Write ever more rules
Every exception becomes another branch. The rules engine grows until nobody can safely change it, and it still misses the case nobody anticipated.
Brittle, unmaintainable
Train a bespoke model
Months to collect labels and train. A team to maintain. A full retrain every time the policy changes, which is exactly when you need it to work.
Costly to build, costlier to keep
Prompt an LLM to decide
It works in the demo. Then every exception becomes another paragraph, the paragraphs start to contradict each other, and nobody can test what a change will break.
Unpredictable, untestable
A prompt decides by feel.A decision shows its work.
The same refund request, decided two ways: “The sole cracked in normal wear. They arrived 41 days ago.”
One prompt5 rules, 2 exceptions
You are a refund assistant. Approve refunds for defective items within 30 days of delivery.
Deny refunds for used items.
Unless the item is defective, in which case it’s fine within a year.
IMPORTANT: never approve more than $250 without a manager.
If the customer has had two refunds this year, be stricter. But not with defects.
DenyTwo lines disagree. The model picked one.
A decisionb90c44e5
- Does the customer describe a defect?0.91
- Has the item been used beyond trying it on?0.86
- Learned from a correctionIs it more than 30 days after delivery, for an item that isn’t defective?0.09
- Two or more refunds in the last year?0.05
- Was the order above $250?0.12
GrantEach question answered on its own, then combined.
Some calls are a balance.Why not treat them that way?
How late is too late depends on what the item cost and why it’s coming back. A prompt can only say be more generous when the reason is serious.
A decision weighs the three against each other with a formula you can read, test and change.
refund-exceptionmove the sliders
Scored 0 to 4 by Moonbounce from the customer’s message.
pushes toward denypushes toward grant
Grantscore 0.90
score = logistic( −2 + 1.8 × seriousness − 0.04 × days late − 0.004 × price )Below 0.4 deny, from 0.4 escalate, from 0.7 grant. Example weights. In practice your agent fits them to your team’s decisions.
You write a policy once.It gets interpreted every time.
By a person, by a model, on cases you imagined and cases you didn’t.
Sooner or later, an interpretation lands somewhere you wouldn’t have. Moonbounce closes that gap by learning from the decisions you would have made.
We learn how your team decides.
Follow one refund from the wrong call to the right one.
Evaluate
A request comes in and Moonbounce decides it.
“The sole cracked in normal wear. They arrived 41 days ago.”
Deny4f2a91c0
Review
One of your people checks the call. Try both answers.
Moonbounce said Deny.
“A defect is covered for a year. Grant.”
Suggest
Moonbounce proposes a fix and tests it on past cases.
Before: Is it more than 30 days after delivery?
After: Is it more than 30 days after delivery, for an item that isn’t defective?
- cases fixed
- 6
- broken
- 0
Merge
You accept the fix. The same request now gets your answer.
Version
4f2a91c0b90c44e5Was: Deny
Grant
For developers
Your agent sets it up.Your code makes one call.
Connect the Moonbounce MCP server and describe the decision. Your coding agent drafts the questions, fits the weights to your team’s past decisions, tests the result and writes the call for you.
Prefer to write it by hand? The same decision is one API request.
- Versioned
- Every change to a decision is a new version you can diff and roll back.
- On record
- Every call is logged with its answers, its outcome and the version that made it.
- Tested first
- Each fix is run against past cases before it ships, so you see what it changes.
your coding agentMoonbounce MCP server
$ claude mcp add --transport http moonbounce https://api.moonbounce.ai/mcp
> /moonbounce:build Decide refund exceptions the way our support leads do. Past tickets and their outcomes are in data/refunds.csv.
- Imported 212 past tickets with your team’s outcomesimport_examplesdata/refunds.csv
- Drafted what to weighvalidate_decisionseriousness of the reason · days late · item price
- Fitted the weights to your team’s decisionsdecision_fitseriousness +1.8 · days late −0.04 · price −0.004
- Tested on 40 tickets it had not seencreate_run38 match your team · 2 sent to a reviewer
- Set the cut pointssweep_cutgrant from 0.7 · deny below 0.4
- Registered refund-exception, version 1register_decisionand added the call to your code
Illustrative session. Counts and weights are examples.
POST /v1/evaluaterequest
curl -X POST https://api.moonbounce.ai/v1/evaluate \
-H "Authorization: Bearer $MOONBOUNCE_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"moonbounce": { "decision_id": "refund-exception", "item_id": "t-001" },
"state": { "message": "The sole cracked in normal wear." },
"attribute_values": { "days_late": 11, "price": 180 },
"attributes": {
"days_late": { "type": "number", "minimum": 0 },
"price": { "type": "number", "minimum": 0 }
},
"questions": {
"seriousness": {
"type": "score",
"instructions": "How serious is the reason for the return?",
"criteria": [
"Changed mind",
"Didn't suit",
"Not as described",
"Defective",
"Dangerous, or serious hardship"
],
"reads": ["state.message"]
}
},
"aggregations": {
"grant_score": {
"type": "logistic",
"bias": -2,
"inputs": [
{ "name": "questions.seriousness", "weight": 1.8 },
{ "name": "attributes.days_late", "weight": -0.04 },
{ "name": "attributes.price", "weight": -0.004 }
]
},
"handling": {
"type": "rules",
"rules": [
{ "when": { "name": "aggregations.grant_score", "gte": 0.7 },
"outcome": "grant" },
{ "when": { "name": "aggregations.grant_score", "lt": 0.4 },
"outcome": "deny" }
],
"else": "escalate"
}
},
"result": "aggregations.handling"
}'200response, trimmed
{
"id": "evl_0199f2c4…",
"status": "decided",
"outcome": "grant",
"answers": {
"seriousness": { "type": "score", "score": 2.96 }
},
"aggregations": {
"grant_score": { "type": "logistic", "value": 0.9 },
"handling": { "type": "rules", "outcome": "grant", "via": "0" }
},
"version": { "structure_hash": "sha256:4f2a91c0…", "number": 1 }
}your codeyou take the action
if (evaluation.outcome === "grant") grantRefund(order)Where it applies
Every industry has decisions rules can’t handle.
Commerce
Refunds and returns
Decide who gets the exception without a support agent reading every ticket.
“Late return, but the item was defective on arrival and the customer has never asked before.”
Marketplaces
Fraud and disputes
Weigh evidence from both sides and the account’s history in one pass.
“Buyer claims non-delivery. Tracking says delivered. Fourth such claim in sixty days.”
AI agents
Action authorization
Check every tool call before it executes. Allow, block or escalate inline.
“Credit is within the agent’s limit. Changing the billing email without verification is not.”
Platforms
Conversations and content
Enforce your own community policy across text, images, audio and video.
“Allowed in a fiction context. Not allowed when the target appears to be a real person.”
Financial services
Compliance review
Check every client communication against the rules before it goes out.
“A performance claim in a client email, with no disclosure attached.”
Support
Routing and escalation
Send each conversation to the right place by what the person needs.
“Caller says cancel, but the problem is a double charge. Billing, not retention.”
In production today
- decisions a day.
- 500M
- AI image generation
- Video analysis
- Dating chat messages
- AI agent context
- Financial compliance
0decisions made since you opened this page.
250Mpeople a month use the platforms those decisions run on.
If Match Group trusts Moonbounce with its hardest calls, every decision easier than those is within reach.
What two people may say to each other in a private conversation is about as hard as a judgment call gets: high stakes, ambiguous context, constant policy change. Refunds, disputes and agent actions are easier.
Bring us yourhardest decision.
Pick the call your team is tired of making by hand. Show us a few examples of how you’d decide it, and watch Moonbounce make the same call.