Connect reviewed AI workflows to core systems and CRM.

The short version

Workflow Studio offers one governed integration surface through two interfaces:

1. An Integration API with signed webhooks. Your integration team connects it to core systems and CRM through your API gateway or integration platform, such as MuleSoft, Boomi, Azure API Management or Kong. That is how most banks and insurers expect a vendor to connect.
2. An MCP server over the same operations, so your approved AI agents (for example Microsoft Copilot Studio, Salesforce Agentforce, Claude or ChatGPT Enterprise) can find and run governed workflows and read reviewed results.

Three rules hold for every interface.

• Results leave only after human review. No interface can accept, reject or correct a run. A named person decides in Workflow Studio.
• Meaning travels with the data. Inputs and outputs are keyed by approved term IDs from a pinned Ontology Studio release, so mapping to Salesforce or ClaimCenter fields means mapping to agreed meaning.
• We never hold your system credentials. Your systems call Workflow Studio, or receive its signed webhooks. Workflow Studio does not log in to yours.

Interface 1

Integration API (REST)

JSON over HTTPS under /api/v1, with an OpenAPI 3.1 specification your team can import into its gateway or iPaaS.

Deliberately absent: accept, reject, return and correct. Those stay with a named person in Workflow Studio.

Start a run

POST /api/v1/workflows/{workflowId}/runs
Authorization: Bearer <token>
Idempotency-Key: 6b0f2c9e-…
Content-Type: application/json

{
  "externalRef": { "system": "salesforce-fsc", "object": "Claim", "id": "a0X5e00000ABC" },
  "inputs": { "claim.id": "CL-9055", "policy.number": "GP-2201" },
  "documents": [ { "title": "claim-file.pdf", "contentType": "application/pdf", "base64": "…" } ]
}

Read the reviewed result

{
  "runId": "3f2a…", "status": "accepted",
  "workflow": { "version": 4, "ontologyRelease": { "version": "2.0.0", "fingerprint": "9ec69e11" } },
  "review": { "decision": "accepted", "reviewer": "j.alvarez@client.example",
              "at": "2026-11-12T15:04:11Z", "separateReviewer": true },
  "outputs": {
    "claimant.id":  { "value": "P-217", "citations": [ { "document": "claim-file.pdf", "quote": "member P-217" } ] },
    "prior_claims": { "value": ["CL-7012", "CL-8130"], "corrected": false, "citations": [ … ] }
  },
  "externalRef": { "system": "salesforce-fsc", "object": "Claim", "id": "a0X5e00000ABC" }
}

Conventions: problem-details errors with stable codes; idempotent starts (the same key and body return the original run); the major version in the path, with 12 months’ notice before retirement; per-integration rate limits; a request ID on every response, recorded in the audit log.

Interface 2

An MCP server for your AI agents

Your teams are deploying internal agents. An MCP server lets those agents use approved workflows and approved meaning instead of improvising their own. It is a second door onto the same governed operations, never a side door.

It is a remote server using the Streamable HTTP transport, with authorization following the MCP specification’s OAuth-based model. Tokens are issued for an integration or for a signed-in person, scoped to this server, and never passed through to other services.

Not available through MCP: review decisions, corrections, workflow or ontology editing, library administration, and anything that writes to your systems.

• Documents and tool arguments are treated as untrusted data. Instructions found inside a document are never followed, and exact-quotation checks stay in force.
• Tool descriptions are fixed and versioned with the server, so data can’t change them.
• Every call is scoped, rate-limited and written to the same audit log, with the agent’s client ID and, when present, the person.
• A read-only MCP server for Ontology Studio’s approved terms is planned alongside it.

Evidence for risk, audit and compliance teams

• Audit log for every call: time, request ID, integration, person if any, operation, workflow version, run, outcome and source address. Content is never logged. Retention is configurable and the log can be exported.
• Run record: frozen inputs, step results with exact citations, model and token counts, reviewer, decision, corrections with reasons and the ontology release fingerprint. Also available as the review packet.
• Change control: workflow versions are immutable, and test-case gates can stop an untested version from running on other documents.

The platform supplies access control, logging, data minimization, human oversight and model change records, designed to support obligations such as NYDFS Part 500, FFIEC guidance, GLBA safeguards and the NAIC model bulletin on AI. Certifications and agreements (SOC 2, penetration testing, a data processing agreement, data residency and SSO) are on the roadmap below. We will not describe the service as compliant with any regulation or standard before the evidence exists.

Roadmap

What the design will not do

• Hold credentials for your core systems or CRM.
• Let any machine interface approve, reject or correct a run.
• Push document content in webhooks.
• Let agents edit workflows or ontologies.
• Claim compliance or certifications before they exist.