Delivery refunds & disputes
"Where's my order?", missing items, never-arrived, late delivery, repeat-refund abuse.
01The problem
Plain LLMs can't be trusted with real decisions.
Point a raw model at your data and ask it to decide, and you inherit three deal-breakers.
-
01
They hallucinate
Models confidently invent refund amounts, policies, and facts that were never in your data.
The cost: every made-up number is real money out the door — or a promise you can't keep.
-
02
You can't audit them
When a decision is wrong, there's no trail back to the exact rule and record that produced it.
The cost: a chargeback, complaint, or regulator asks "why?" — and you have no answer.
-
03
You can't test them
No way to prove the system behaves — or to catch it silently regressing after a prompt change.
The cost: one prompt tweak, and production quietly breaks with no alarm.
The fix
A decision engine that shows its work.
Grounded
Answers are built only from the retrieved record and policy. Nothing invented.
Verified
A deterministic gate checks the decision against the record before it ships.
Auditable
Every decision carries a receipt: the exact rule and record behind it.
Tested
A golden-case scorer turns "trust me" into a reproducible number.
02How it works
One request. A decision you can defend.
A multi-stage engine — not a single model call. High-risk requests never reach the AI, answers are built only from your data, and every decision is checked before it ships. It always fails safe to a human.
1ScreenBlocks fraud, legal & abuse before any AI call.
2PlanAI picks the intent and what data it needs.
3RetrieveFetches the exact policies & records — nothing more.
4Escalation checkDeterministic rules route risky cases to a human.
5ValidateGate 1 — fails safe if required data is missing.
6GenerateAI drafts the answer from retrieved data only.
7VerifyGate 2 — re-derives the decision, rejects mismatches.
8Output checkFinal scan blocks any internal detail from leaking.
Deterministic where it counts — the model where it's smart. The whole pipeline is wrapped in a full audit trail: any error escalates safely, never a bad auto-decision.
03Capabilities
Engineered for trust, not demos.
Beyond the core guarantees — the production capabilities that turn a prototype into something you can rely on.
Fail-safe by design
Any error, missing data, or answer it can't verify routes to a human — never a confident-but-wrong auto-decision.
error → humanunverifiable → humanunsure → human
Provider-agnostic
OpenAI, Anthropic, or a local model — the engine never hard-codes a vendor.
Domain-agnostic core
Swap the policies + data; the engine is reused untouched across domains.
Scales with your rulebook
Built to handle hundreds of policy documents.
04The difference
A plain LLM app vs. Grounded.
Capability
Plain LLM app
Grounded
Answers grounded in your data
Sometimes
Always — or it escalates
Decision verified against the rules
No
Deterministic gate
Traceable to a rule + record
No
Full audit trail
Behaviour is testable
Hard
Golden-case scorer
Safe on failure
Guesses
Escalates to a human
Swap the AI provider freely
Vendor lock-in
Provider-agnostic
05Scope & potential
One engine. Any decision that pairs a record with a rule.
Shipped today for delivery refunds — and built to extend to any domain where a decision must follow policy.
Insurance claims
Adjudicate a claim against a coverage policy — with the citation regulators expect.
Finance & invoice approvals
Approve or hold an invoice against approval rules, every call traceable.
HR policy eligibility
Leave, benefits, and eligibility decisions grounded in the record + HR policy.
Customer support
Policy-grounded answers and actions, with escalation for anything high-risk.
Compliance checks
Deterministic, auditable checks where "the model said so" will never fly.
06Questions
Straight answers, before you ask.
The things engineers and operators actually want to know before trusting AI with a decision.
Is this just a chatbot or a RAG wrapper?
No. Grounded is a decision engine, not a conversation. It runs each request through deterministic stages, builds the answer only from data it actually retrieved, and re-derives the decision in code before it ships. A chatbot talks; Grounded decides — and proves it.
What happens when the engine isn't sure?
It escalates to a human. Any error, missing data, or answer it can't verify routes to a person instead of guessing. It fails safe by design — you never get a confident-but-wrong auto-decision.
Which AI models does it work with?
Any of them. Every AI call goes through one shared interface, so you can run OpenAI, Anthropic, or a local open-weights model — and switch without touching the engine. No vendor is hard-coded.
How is a decision actually "proven"?
A deterministic check re-derives the correct decision from the real record and rule, and rejects the AI's answer if it doesn't match. Every decision also carries a receipt — the exact policy and record behind it — logged in a full audit trail you can replay after the fact.
Can I use it beyond delivery refunds?
Yes. The core engine is domain-agnostic. Swap in your policies and data and the same pipeline handles insurance claims, invoice approvals, HR eligibility, or compliance checks — any decision that pairs a record with a rule.
Let's talk
See Grounded make decisions on your data.
Exploring grounded, auditable AI for refunds, claims, approvals, or compliance? Run the live demo, or get in touch.