The Integration Blueprint
The number-one question before any AI deployment: how does this touch Guidewire, Duck Creek, Epic, or AMS360? This is the architecture answer — a layered blueprint that lets lab modules do real work without ever destabilizing the system of record.
Every module in the lab repository is placed on this map automatically. As new modules are added, they appear at the layer where they do their work — click any module to open its full page.
Untouched. Always.
Guidewire, Duck Creek, Applied Epic, AMS360, Salesforce. These remain the single source of truth. Nothing in the lab ever asks them to change how they work.
The translator in the middle.
A canonical data model sits between the lab and the core. Legacy formats — AL3, IVANS downloads, ACORD forms, SFTP drops — are normalized into clean, modern schemas before any module sees them. Writes go back in each system's native dialect.
Lab modules at this layer — live from the repository
Real-time where it matters.
Webhooks and event streams (Guidewire Integration Gateway, mailbox triggers, claim milestones) wake modules the moment something happens. Where no events exist, scheduled polling and nightly batch fill the gap — honestly labeled, never faked.
Lab modules at this layer — live from the repository
The lab, doing its work.
Each lab module reads from the canonical layer, reasons in an isolated zero-retention environment, and returns a recommendation, draft, or score — never an unapproved action on the system of record.
Lab modules at this layer — live from the repository
People lead. Always.
Every write-back lands as an activity, note, draft, or task assigned to a named human. Underwriters approve triage. Producers approve drafts. Nothing binds, sends, or closes on its own.
Lab modules at this layer — live from the repository
Every module-to-core connection in the lab uses one of these five patterns — chosen per system pair, never forced.
Modern cloud cores expose well-documented APIs — Guidewire Cloud APIs, Duck Creek Anywhere, Epic API, Vertafore Orange Partner. This is the default path: reads and writes in real time, authenticated with OAuth 2.0 against your existing identity provider.
Used with: Guidewire, Duck Creek, Applied Epic, AMS360, Salesforce
Where the core can publish events — a new submission, a claim milestone, a late payment — modules subscribe instead of polling. Latency drops from hours to seconds, and the core never feels the load.
Used with: Guidewire Integration Gateway, M365 / Google, claim events
Some modules live inside the tools people already have open: a churn score on the Salesforce account page, a draft in the producer's Outlook folder, a brief on the opportunity record. No new tab to remember.
Used with: Salesforce FSC, Microsoft 365, notification layers
The industry's existing plumbing still moves enormous volume. The bridge parses AL3 downloads, ACORD forms, and commission statements into the canonical model — so legacy feeds become machine-readable without changing the sender.
Used with: IVANS download, ACORD 125/126/130, commission statements
For systems with no API surface, scheduled SFTP exchange and warehouse extracts (like Guidewire Cloud Data Access) provide a reliable, auditable path. Slower, but honest — and often the only sanctioned route.
Used with: Legacy portals, Guidewire Data Platform, loss-run retrieval
Every tool page in the lab carries its own compatibility matrix — which systems it connects to, which pattern it uses, and how long a working connection takes.
Browse the lab →Every system is described the same way — the API layer it exposes, how identity works, what events it can publish, and which canonical entities flow in and out. Modules never talk to a system directly; they talk to the canonical model, and the adapter does the translation. That's what keeps the lab loosely coupled: swap the system, keep the module. Click a system to open its profile.
The bridge to agents
Every system above already has an API. That is not the hard part. The hard part is that those APIs were built for deterministic software, not for agents. An MCP microservice is the thin, governed layer in between — a small container that wraps the existing API and exposes a handful of business-shaped tools instead of hundreds of raw endpoints. Nothing changes in the core system. The API you already have becomes agent-ready.
High-intent actions, not endpoints
Instead of exposing the full REST surface, the microservice publishes a short list of business-shaped calls — look up a commercial account, summarise five years of losses, stage a renewal activity. Each one has a typed schema the agent cannot drift from.
Read-only context the agent can pull
Documents, dec pages, loss runs, and claim timelines are exposed as addressable URIs. The agent pulls only what it needs, when it needs it, instead of being pre-loaded with the whole record.
Governed workflows, not free improvisation
Reviewed workflow templates ship with the server — submission triage, renewal drafting, pre-bind compliance. Every write tool is implemented as a task assigned to a named human, so the human gate is enforced in code, not policy.
Same container, different adapter. Each one wraps the API that already exists, speaks the canonical model on the inside, and keeps credentials out of the agent runtime entirely. Click a system to see the tools it exposes.
Every lab module works the same way. Each one consumes these adapter tools and publishes its own MCP interface, so it can be called by your assistant, your orchestrator, or another module. Open any module page to see the exact tools it consumes and provides.
Press Run and look at the request tab. It has exactly four meaningful parts:
Flip to the response tab. The answer lives in three places:
Epic's raw API returns deep, 100+-field account objects. Watch what the MCP layer does with that noise: Switch the Guidewire tab above to see its version.
Common agency tasks, mapped to the exact tool calls an agent would make. Run a step and watch the request and response fill in below.
A CSR needs the full policy schedule for Harbor Logistics before a renewal meeting.
Find the account and its client_id from a name — no navigating Epic's account screens.
Pull every in-force policy, premium and limit as 7 clean fields instead of a 143-field account object.
The Ghostwriter module spotted a 14% renewal increase and needs a human to own the conversation.
Stage an activity assigned to the account's CSR, due in 3 days. Nothing in Epic changes until that person approves it.
Run one task and compare the exact request and response formats Applied Epic and Guidewire produce through their MCP adapters.
Find an account by name, FEIN, or policy number.
// Pick a tool, choose arguments, and press Run.
All data is fictional. Write tools only ever stage a task for a named person — notice they never change a policy or advance a job.
No rip-and-replace, no core system modification, no downtime. Modules attach to the edges and prove value in days, not quarters.
Every engagement starts read-only. Write-back is earned — and even then it lands as a draft, note, or task, never an unapproved transaction.
Zero-retention inference. No customer data trains any model. Field-level audit trails on every value that moves in either direction.
ACORD, IVANS, OpenAPI, OAuth 2.0 / mTLS. Boring, proven, and supportable by the team you already have.
The layers aren't just technical — they're commercial. Each one is a different way to engage, priced for the risk you're taking on.
Every engagement opens with a read-only module — Retention Radar, Producer Copilot, Loss Run Analyzer — connected in days with zero write access. The agency proves value on its own book before spending a dollar, and the core system never feels a thing.
Best fit: Agencies & brokers testing the water
Once a pilot proves out, modules convert to a flat monthly subscription per active module — priced so a single saved account or avoided E&O event covers the year. No per-seat pricing: the whole team benefits from one deployment.
Best fit: Agencies scaling what works
Carrier API Bridge and Book Intelligence Warehouse are infrastructure, sold as an annual foundation layer. Once the canonical model and warehouse exist, every other module plugs in at marginal cost — this is where carriers and large brokerages start.
Best fit: Carriers, MGAs & enterprise brokers
Technology partners build their own modules on the canonical layer and list them in the lab. BionicAgent provides the integration fabric and distribution audience; the partner keeps the IP. New modules appear on this map automatically.
Best fit: Technology partners & insurtechs
Executive download
The BionicAgent Blueprint condenses this architecture — the layers, adapters, MCP bridge, governance, and the six-week rollout — into a 12-page document you can share with your leadership team.
The people-first operating model: architecture, adapters, MCP, governance, and a six-week rollout.
Free download
Connecting Agentic Workflows to Guidewire PolicyCenter — the zero-disruption implementation guide. Five integration patterns, eight security guardrails, three production-tested workflows, and a six-week rollout plan.
Describe your distribution challenge and the Matchmaker will recommend the lab modules that fit — with the integration pattern for your stack.