Operational Guardrails — for the three workflows that always run hot.
Pre-packaged bundles for the mid-market: Finance Guardrails (refunds, spend, vendor onboarding), AI Agent Guardrails (tool calls, API execution, workflow actions), Revenue Guardrails (discounts, contract changes, pricing overrides). Each bundle ships every surface Decionis already supports — pick one, install in shadow, watch the verdicts, then flip to enforce. No per-workflow integration project.
Pick the bundle whose audit you'd be asked about first.
Each bundle bakes in the policy pack, the surfaces you'll install on, and the four workflows it gates on day one — so "decide what to govern" isn't a three-week project before you can produce the first dossier.
Finance Guardrails
Finance and ops teams — refunds, spend approvals, vendor onboarding
Stop the 👍-in-Slack approval. Every refund, spend release, and vendor onboarding leaves a signed Decision Dossier — the audit your CFO and your auditor both ask for.
Gated on day one
Refund approvals above threshold
Discretionary spend / PO releases
Vendor onboarding & supplier risk review
Wire / payout authorization
Flagship metric: % of refunds / vendor onboardings that resolve through a signed dossier
Agent / AI-product teams — every tool call, API execution, workflow action
The LLM picks the tool. Decionis decides if the call fires. Every tool invocation becomes a signed dossier; jailbreaks and hallucinated calls short-circuit before they move state.
Gated on day one
Agent tool calls that move money or write to ERP
RAG / external API calls on regulated data
Multi-step LangGraph workflows with branching execution
Outbound webhooks / GitHub Actions triggered by agents
Flagship metric: Hold rate on agent-initiated high-blast-radius actions (target: < 2% surprise rate)
Stale coupon codes, aggressive auto-discounts, and AI-driven price overrides quietly destroy margin. Decionis evaluates every change against margin floor + channel consistency before it ships.
Three real scenarios per bundle. The dossier you'd hand to the reviewer is the same artifact every time.
Finance Guardrails
three real scenarios
Refund above $10k
Stripe webhook → Decionis gates the release against policy version + reviewer authority + duplicate-claim signals. Dossier ID stamped on the Stripe metadata.
Net-new vendor onboarding
ERP creates a vendor record → Decionis verifies KYC evidence + sanctions check + spend cap → ALLOW only if policy is satisfied; ESCALATE to ops otherwise.
Discretionary spend > $5k
Slack approval ping carries the policy snapshot, the requester's authority, and the public verify URL. The reviewer's 👍 becomes a signed verdict.
AI Agent Guardrails
three real scenarios
send_refund tool call
Wrap any LangChain BaseTool with DecionisGateTool.wrap — the inner tool only fires on APPROVE; REJECT / REVIEW / ESCALATE returns a structured refusal carrying the dossier id.
Multi-step LangGraph plan → execute
Drop decionisGateNode between plan and execute; the graph's conditional edge branches on `state.decionis.outcome` (allowed / blocked / errored).
Agent-triggered deploy
Decionis GitHub Action gates the workflow step — agent can author the change but the deploy only runs on an ALLOW verdict; the PR carries the verdict + verify URL.
Revenue Guardrails
three real scenarios
Stacked-promo checkout
Shopify Functions hand the cart to Decionis pre-payment; margin floor + promo-eligibility verdict holds checkout when the combined discount would breach the policy.
AI pricing engine override
Pricing service calls the Decionis Action Gate before publishing a new price; verdict + dossier ID is the audit trail finance reviews weekly.
Cross-channel consistency
Same SKU, different channels. Decionis evaluates the parity contract on every update and escalates drift to a reviewer with a signed comparison.