AI Automation

AI Automation

LLM-driven flows, agentic conversations, scheduled jobs. Bring your own DAG, plug in Quantum AI for reasoning + sibling services as tools.

Branded as Quantum AI. Internal model stack is not exposed via this docs surface.

What it does

Capability	API
Flows (LLM DAGs)	Visual builder + REST API. Supports `llm_step` nodes alongside `service_call`.
Triggers	API call, public HMAC-signed webhook, or cron schedule.
Agents + conversations	Reusable system-prompt + tool sets; multi-turn chat with tool use.
Service connections	Same shape as Orchestration — register sibling products + custom HTTP.
Approval gates	LLM proposes action, human approves before execution.
Watchers + incidents	Conditional rules over events/checks; stateful, dedup'd incidents with ack/resolve/escalate.
Digests	Cron-scheduled, LLM-summarized reports to email · Slack · PDF.
Evals	Run a prompt against a labeled input set; pass/fail metrics for prompt CI.
Outbound webhooks	Subscribe to flow runs, conversations, incidents, digests, evals.
Audit + immutability	Postgres-trigger-enforced.

When to use it

AI Automation is right when you want:

Summarisation of a customer's history before an analyst makes a decision.
Classification of a free-text complaint into a routing bucket.
Structured extraction from a conversation (e.g. dispute claims).
Agentic workflows — LLM iteratively uses tools until it reaches a goal.
LLM-in-the-loop with human approval — model drafts, human signs off.

Use Orchestration instead when the flow is fully deterministic and you don't need an LLM step. Use both when you want a deterministic spine with LLM-augmented leaves.

Core concepts

Concept	What it is
Flow	A versioned DAG. Like Orchestration's workflow, with extra node types: `llm_step`, `http_request`, `http_sink`, `webhook_emit`.
Run	One execution of a flow. Has trigger source (API/webhook/cron), inputs, steps[], status.
Agent	Reusable LLM character: system prompt + tool set + model choice.
Conversation	Multi-turn chat session with an agent. Supports streaming.
Tool	A callable function exposed to an agent: a sibling endpoint, your HTTP service, or built-in.
Trigger	How a run starts: API call, public webhook (HMAC-verified), or cron schedule.

Node types

In addition to all of Orchestration's node types (service_call, decision, transform, wait, webhook_emit, human_approval), AI Automation adds:

Type	What it does
`llm_step`	LLM completion. Templated prompt, structured output via JSON schema, optional tool use.
`http_request`	Outbound HTTP to any URL (not just registered services). For ad-hoc integrations.
`http_sink`	Posts the step's input to a configured URL. Convenience wrapper for "send to my service".

Triggers

Trigger	Use for
`api`	Your service explicitly calls `POST /api/flows/{id}/execute`.
`webhook`	A public URL (`/api/triggers/wh/<token>`) accepts HMAC-signed POSTs. Use for partner-triggered flows.
`cron`	Standard 5-field cron expression. Runs on schedule with 60-second tick resolution.

Common integration shape

Customer complaint triage (cron-triggered every 15 min)

1
Trigger
Cron tick.
Every 15 minutes.
2
`service_call` (your CRM).
GET `/api/tickets?status=new&since=last_seen`.
3
`llm_step` (Quantum AI) — classify the ticket.
Output schema: `{ category: enum, urgency: 1–5, summary: string }`.
4
`decision` — branch on urgency.
If `urgency >= 4` → page on-call. Else → tag ticket and continue.
5
`service_call` (your CRM).
PATCH `/api/tickets/{id}` with classification fields.

Scope naming

AI Automation scopes use a colon separator (flows:write, incidents:read) — different from AML / Identity / Anti-Fraud, which use a dot (sar.write). Both formats are accepted suite-wide for backwards compatibility, but new keys should use the form documented per-product. Mixed-format keys are fine; the gateway normalizes on parse.

Endpoints at a glance

Group	Endpoints
Flows	`GET/POST /api/flows` · `GET/PATCH/DELETE /api/flows/{id}` · `POST /api/flows/{id}/execute` · `POST /api/flows/{id}/rotate-webhook-secret` · `GET /api/flows/{id}/runs`
Triggers	`POST /api/triggers/wh/{token}` (public, HMAC-signed)
Runs	`GET /api/runs/{id}` · `POST /api/runs/{id}/approve` · `POST /api/runs/{id}/reject`
Agents	`GET/POST /api/agents`
Conversations	`GET/POST /api/conversations` · `GET /api/conversations/{id}` · `POST /api/conversations/{id}/messages`
Watchers	`GET/POST /api/watchers` · `GET/PATCH/DELETE /api/watchers/{id}` · `POST /api/watchers/{id}/test`
Incidents	`GET /api/incidents` · `GET /api/incidents/{id}` · `POST /api/incidents/{id}/ack` · `/resolve` · `/escalate` · `/comments`
Digests	`GET/POST /api/digests` · `GET/PATCH/DELETE /api/digests/{id}` · `POST /api/digests/{id}/run-now`
Evals	`GET/POST /api/evals` · `GET /api/evals/{id}`
Service connections	`GET/POST /api/service-connections`
Cron preview	`GET /api/cron-preview?expression=...`
Webhooks	`GET/POST /api/webhooks`
API keys	`GET/POST /api/api-keys`

Production considerations

A short list of things every regulated-finance client asks before going live.

Concern	Answer
Data residency for LLM	All inference runs inside our Indonesia VPC (`id-jkt-1`). Customer payloads sent to the LLM never leave the region and are never used to train models. Full posture: Data residency →.
Model tier	One tier today: `quantum-ai:default`. ~8K-token context, schema-constrained JSON output, p95 inference ~2s. A `:fast` (cheaper, smaller) and `:large` (longer context) tier are on the 2026 Q4 roadmap.
Prompt + body limits	Request body capped at 1 MB. Per-`llm_step` prompt + completion cannot exceed the model's context window — `outputSchema` enforcement may retry up to 3× on parse failure.
Rate limits	Per-org default: 1,000 flow executions/min, 200 LLM inferences/min, 50 webhook trigger requests/sec. Inference quota is also enforced daily by tier — see Rate limits →.
Idempotency	`POST /api/flows/{id}/execute` accepts an `idempotencyKey` (any string, ≤ 256 chars). Repeated calls within 24h return the original `runId` instead of starting a new run.
Audit	Every flow execute, run approve/reject, watcher fire, digest send, eval start, conversation message is audit-logged with actor + timestamp + before/after. Retention 7 years.
Webhook signatures	Incoming trigger webhooks AND outgoing event webhooks both use `X-Quantum-Signature: sha256=...`. They use different secrets: the trigger's `webhookSecret` (per-flow) for inbound, the subscription's bearer secret (`qe_whsec_...`) for outbound. Don't reuse secret material between them.