# Endorsement Processing

Read a broker's policy change request, evaluate it against the policy on file, and either complete it straight through or hand a policy servicing rep a prepared evaluation with the specific issue already identified.

## The business problem

A broker's endorsement request looks routine — add a vehicle, add a location, raise a limit, add an additional insured — but it has to be checked against the real policy before it can be actioned: is it in force for the date requested, is it backdated, is the package complete, does it need underwriting's attention. Done by hand across a package of mixed documents, that check is exactly where a backdating issue or missing data element gets missed.

The **Endorsement Processing** template automates it: classify and extract every document, reconcile what's actually being requested, evaluate it against the available policy evidence, and either complete it straight through or route it to a policy servicing rep with the reason already called out.

> **Important: the evaluation is decision support, not a decision.** The agent lays out the reasoning and evidence; it doesn't make a binding underwriting decision. Clean requests complete straight through, even ones that also need underwriting's attention — a second, non-blocking review opens for underwriting instead. Everything else goes to a person to review, edit, and approve.

## Trigger and source data

The template ships with the **Manual File Upload** trigger. You can configure other triggers, including a connection to an Outlook email inbox. Inbound webhooks, scheduled/timer triggers, and SFTP are **not currently available**; drive runs from another system through the [Bevaya API](../api/flow-executions#run-a-flow). See [Configure triggers](../build-ai-agents/triggers.md).

Upload a broker change request package — typically an email with a few attachments — containing one or more of:

- A **broker request email** stating the requested change
- A **policy change request form**, if the broker uses one
- A **vehicle**, **property**, or **driver schedule**, for a fleet, location, or roster change
- A **contract or lease**, if an additional insured or certificate requirement is driving the change
- A **payroll or exposure update**, for a mid-term exposure change

The evaluation works from any policy evidence in the package — a declarations page, a prior endorsement schedule, a certificate. Connect the flow to your policy administration system (see below) to verify against the live record instead. Either way, if no policy evidence can be matched to the request, the item routes to review stating what's needed to proceed.

## Template starting point

Create an AI agent from the **Endorsement Processing** template for a working flow that classifies, extracts, evaluates, and routes out of the box. Open it in the canvas, select an environment, click **Start Editing**, and adjust field prompts, the evaluation prompt, or validation thresholds to your lines of business. See [Create an AI Agent](../build-ai-agents/create-agent.md).

## What the template preconfigures

| Step in the flow | Node it uses | Category |
| --- | --- | --- |
| Prepare the uploaded files | [Read Files](../build-ai-agents/utility-nodes.md) | Utility |
| Process each document in the package | [For Loop](../build-ai-agents/control-nodes.md) | Control |
| Classify each document and extract per-type fields | [InsurGPT: Custom](../build-ai-agents/insurgpt-nodes.md) | InsurGPT |
| Route each document to the right extractor | [Switch](../build-ai-agents/control-nodes.md) | Control |
| Reconcile the package into a single endorsement request | [InsurGPT: Custom](../build-ai-agents/insurgpt-nodes.md) | InsurGPT |
| Evaluate the request against the policy evidence available (in-force, backdating, completeness, premium impact) | [InsurGPT: Custom](../build-ai-agents/insurgpt-nodes.md) | InsurGPT |
| Check extracted and analyzed values against rules | [Field Validation](../build-ai-agents/utility-nodes.md) | Utility |
| Open a non-blocking review for underwriting, if flagged | [Switch](../build-ai-agents/control-nodes.md) → [Review](../build-ai-agents/utility-nodes.md) | Control / Utility |
| Route the prepared evaluation to a policy servicing rep, if flagged | [Switch](../build-ai-agents/control-nodes.md) → [Review](../build-ai-agents/utility-nodes.md) | Control / Utility |
| Close out the item | [Complete](../build-ai-agents/action-nodes.md) | Action |

The evaluation produces the findings shown to the reviewer — policy match, in-force check, backdating flag, data completeness, coverage impact, a premium change estimate, and an underwriting referral flag — plus the endorsement description and type code to post to your policy administration system.

## What builders must configure before deployment

- **Extraction, evaluation, and validation.** Tune the prompts, fields, and validation rules to your lines of business.
- **Review routing.** Set confidence thresholds and the reviewer queues — policy servicing for the main review, underwriting for the second, non-blocking one. See [Utility nodes](../build-ai-agents/utility-nodes.md).
- **A connection to your policy administration system.** We recommend an [HTTP node](../build-ai-agents/utility-nodes.md) before the evaluation step to look up the policy by policy number or insured name, followed by a [Custom Code Block](../build-ai-agents/utility-nodes.md) to shape the response for the evaluation prompt. Without it, the evaluation works only from policy documents in the package.
- **Straight-through completion conditions.** A change that also needs underwriting's attention still completes straight through, since a referral is a routing decision, not a defect. Confirm this matches your risk tolerance, and add any referral reasons that should instead block completion.
- **Posting the endorsement.** A straight-through item completes with the endorsement description, type code, evaluation, and source documents ready to post. To post automatically, add an [HTTP node](../build-ai-agents/utility-nodes.md) calling your system's API, with connection details stored in **Settings → Variables**. See [App integration nodes](../build-ai-agents/app-integration-nodes.md).

## How the AI agent behaves

The agent classifies and extracts each document (the request itself, schedule rows, contract or exposure detail), reconciles the package into a single statement of the change, and evaluates it against the policy evidence available — never assuming a mid-term change is fine just because a policy number matches. The evaluation is evidence-backed, not a verdict: it confirms the policy match and in-force position, flags backdating and incomplete data by name, estimates premium impact where the package supports it, and recommends whether underwriting needs to weigh in.

It works only from the supplied documents (and the matched policy record, if connected); it doesn't invent values, and low-confidence or missing fields are flagged rather than fabricated.

## Human review model

The reviewer works the item in the Bevaya Platform across two tabs:

- The **Insights tab** — an exception summary, an AI recommendation with confidence and suggested next steps, key-insights cards (requested change, policy match, in-force check, coverage impact, underwriting referral), a completeness and exceptions breakdown, a premium summary, the source documents, and an audit trail.
- The **Review tab** — field-by-field verification of the request, the schedule or contract behind it, and the policy evidence found, each linked to its source document and page.

Clean requests complete without stopping at this gate. Everything else pauses at **Review** until the reviewer submits. Reviewers can work items and reviews but cannot build or run flows. See [Human review](../monitor-review/human-review.md).

## Run status and reporting

Each request becomes an item moving through **In Progress, Review, Complete, Failed, Canceled**. Watch runs in [Run history](../monitor-review/run-history.md) and live counts by status — including the straight-through share — in [Item status reporting](../monitor-review/item-status-reporting.md).

## Example run

A broker emails to add an additional insured to a landlord's General Liability policy ahead of a lease commencement date, attaching the lease exhibit and the current declarations page.

1. The flow prepares the files, classifies the email, exhibit, and declarations page, and extracts the additional insured's details plus the policy's in-force dates and terms.
2. It reconciles the package into one statement of the requested change.
3. The evaluation confirms the policy is in force, the package is complete, there's no backdating, and no premium impact is expected.
4. Nothing needs underwriting's attention, so the item completes straight through with the endorsement ready to post — no reviewer touch needed.

## Common failure modes

- **No policy evidence available.** The item routes to review stating what's needed to proceed.
- **Backdating.** A requested effective date earlier than the request date carries adverse selection risk, so it's flagged rather than completed silently.
- **A change needing underwriting's attention.** Opens a second, non-blocking review — tune which referral reasons (if any) should instead block completion.
- **Low-confidence or missing fields.** Flagged by validation rather than fabricated.
- **A misclassified document.** The reviewer corrects it and resubmits; recurring patterns are best fixed in the classification prompt.

## Recommended rollout path

1. **Connect it to your policy administration system**, if you want live verification rather than package-only evidence.
2. **Add a write-back** to post completed endorsements automatically.
3. **Build and test** on sample change requests with **Run Draft**, inspecting each step. See [Test and debug a draft](../build-ai-agents/test-debug-draft.md).
4. **Validate the evaluation** against known changes with an experienced policy servicing rep.
5. **Pilot with the review gate on every request** to build trust before enabling straight-through completion.
6. **Turn on straight-through completion** once accuracy is proven out.
7. **Promote to production.** See [Drafts and publishing](../build-ai-agents/drafts-publishing.md).

## Where to go next

- [Utility nodes](../build-ai-agents/utility-nodes.md) — Field Validation and Review.
- [InsurGPT nodes](../build-ai-agents/insurgpt-nodes.md) — classification, extraction, and the endorsement evaluation.
- [App integration nodes](../build-ai-agents/app-integration-nodes.md) — connect a lookup or write-back to your own systems.
- [Human review](../monitor-review/human-review.md) — the Insights and Review tabs a reviewer works from.
- [Item status reporting](../monitor-review/item-status-reporting.md) — track requests by status.
- [Use cases overview](./overview.md) — the full catalog of supported patterns.