Answers, not hallucinations
Citation-gated answers mean a desk operator can verify before acting. The trace link is the audit answer.
Capability · AI desk assistant
A desk assistant that reads your data — and cites everything.
A conversational AI grounded on your lock, hedge, pricing, and exception data. Asks of the desk get cited answers with trace links, suggested resolutions for exceptions, and an audit-chained reply log. Tenant-scoped, rate-limited, never used to train models across customers.
Overview
What it is, in one paragraph
The AI Desk Assistant is a conversational interface to your platform data — locks, exceptions, hedge view, pricing decisions, audit chain. Questions get grounded answers with citations into the source events. Suggested resolutions on exceptions land as a draft you can accept, edit, or reject. Every prompt and reply writes to an audit-chained AI log tagged with the operator's identity and the correlationIds of the cited records. Rate-limited at 30 rpm per tenant; never used to train models across customers.
Grounded on your data
Tenant-scoped retrieval over locks, exceptions, ratesheets, hedge view, and audit chain. The assistant cannot answer about another tenant's data.
Cited answers
Every claim cites correlationIds. Click to land on the source event in the audit chain — the assistant cannot hide its sources.
Exception auto-suggest
On the Exception Inbox, the assistant proposes a resolution as a draft. Resolve, dismiss, embed for context, or recompute — every action is audit-chained.
30 rpm rate-limit cap
Token usage and call rate capped per tenant to keep cost predictable and prevent runaway prompts.
Tenant-isolated training
Operational tuning happens within your tenant data only. We never aggregate prompts or replies across customers.
Audit-chained replies
Every prompt and reply lands in an audit-chained AI log with operator id, scope, citations, and tokens consumed.
How it works
The pipeline, end to end.
Numbered steps from input to output. Each step maps to a specific subsystem you can inspect via OpenTelemetry.
- 1
Operator asks (or the platform asks on their behalf)
Operator types a question in the assistant pane; or the Exception Inbox calls /v1/ai/assistant/suggest with the exception payload.
- 2
Retrieval over tenant-scoped data
Retrieval queries the lock store, ratesheet versions, hedge view, and audit chain for relevant records — only within the operator's capability scope.
- 3
Model generates a grounded answer
The model is prompted with the retrieved facts and instructed to cite every claim by correlationId. Hallucination guard rejects answers without citations.
- 4
Citations resolve to traces
The UI renders each citation as a link to the audit-chain trace. Clicking the link lands on the source event.
- 5
Audit row writes
Prompt, retrieved record ids, model output, citations, and tokens consumed all land in the AI log. Hash-chained into common_audit_log.
- 6
Operator acts
On an exception: accept the suggested resolution, edit it, dismiss, or recompute. On a free-form question: drill into a citation or ask a follow-up in the same thread.
Hands on
Ask the assistant via API
Live cURL sample — copy, paste, ship.
# Ask the assistant a question over your desk's data
curl -X POST https://api.ratestack.com/v1/ai/assistant/ask \
-H "X-API-Key: $RATESTACK_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"scope": "desk:primary",
"thread": null,
"message": "Why did pullthrough drop on the 7yr cohort this morning?"
}'
# Returns a cited answer:
# {
# "answer": "Pullthrough confidence on the 7yr cohort dropped 0.18 ...",
# "citations": [
# { "type": "lock", "id": "lck_8a7c", "correlationId": "c0c0…" },
# { "type": "ratesheet_version", "id": "rs_9b1a", "correlationId": "c0c0…" }
# ],
# "auditId": "ai_log_4f2c"
# }Why this matters
The pain it removes.
Exceptions resolve faster
AI-suggested resolutions cut triage time on the Inbox. The operator stays in control — accept, edit, or override.
Compliance keeps the receipts
Every prompt and reply is logged with citations. Auditors get a complete picture of what AI was asked and what it claimed.
Frequently asked
Direct answers, no marketing spin.
Is the assistant trained on our data?
No across-tenant training, ever. Operational tuning — feature embeddings, learned-template patterns — stays inside your tenant. We do not aggregate prompts or replies for shared model improvement.
What model is behind it?
Frontier large language model accessed via a vendor-neutral abstraction. The model can be pinned per environment; Enterprise customers can request a specific model floor for change-management reasons.
Can the assistant take actions on its own?
Today it suggests; operators act. Auto-apply is on the roadmap for narrow, audit-chained workflows — and will be opt-in, capability-gated.
What if it hallucinates a citation?
The citation must resolve to a real record in your tenant. The UI verifies citations server-side before render; an unresolved citation degrades the reply and writes an audit warning.
How is cost controlled?
Per-tenant rate cap (30 rpm default), per-call token cap, and a monthly cost ceiling on every tier. Exceeding the ceiling routes new prompts to a degraded mode (template-based response) rather than billing surprise.
Related capabilities
Pieces that work better together.
Ready to see it on your data?
Wire a desk assistant that reads your data — and cites everything. up to your real workflow.
We'll spin you a sandbox, load your actual ratesheets, and walk you through this capability against your top scenarios.