Automate PolicyCenter, ClaimCenter and BillingCenter with an AI agent that adapts through UI churn. Every fix goes to a human for approval, so a heal never hides a regression. Then it proves the rating underneath is right. API to UI, self-hosted, your own LLM.
Demos run on a real Guidewire Cloud tenant · Self-hosted · BYO-LLM
The proof
Karate is so good at testing the hardest platform in insurance that Guidewire embedded Karate as the foundation of its own Testing Framework. Every Guidewire Cloud customer already runs Karate underneath.
“Guidewire decided to leverage Karate Labs for our Testing Framework, as it is a powerful, open-source, industry-standard technology designed to make testing automation simple and accessible.”
Will Murphy
VP of Global Technology Alliances, Guidewire
Why Guidewire is hard to test
Configuration-driven screens change with every LOB tweak, and Guidewire Cloud ships updates on the vendor’s cadence. Selector-based tests snap each time. Under the UI sits custom Gosu and rating logic a screen click cannot verify.
PolicyCenter, ClaimCenter, BillingCenter
They share a platform and a UI framework, so teams tend to test them the same way. The risk in each one is not the same.
Submission, quote, bind, issue, then the long tail that breaks in production: mid-term change, renewal, cancellation, reinstatement. Every line of business has its own product model, so screens are generated from configuration that differs insurer to insurer. Selector-based tests fail the moment a product analyst adjusts a coverage term.
The agent works from display text, following the flow the way an underwriter would rather than by DOM position. Underneath it drives the PolicyCenter APIs and checks premium against the rate book. Reaching the bind screen proves the workflow ran, not that the number is right.
FNOL intake, segmentation and assignment, reserving, payments, subrogation. Business rules concentrate here, and most resolve server-side. A test that confirms a claim was created verifies almost nothing: not which rule fired, not which adjuster was assigned, not whether the reserve is right.
Driving intake through the UI and asserting the resulting transactions and rule outcomes through the API, in one test, is the difference between a smoke test and evidence. The same run produces the traceability record.
Invoicing, payment allocation, disbursements, delinquency, commissions. BillingCenter moves money, which changes what a defect costs. A layout regression is embarrassing. A misallocated payment or a wrongly triggered delinquency is a remediation exercise, and in some jurisdictions a reportable one.
Verification here is arithmetic, not appearance: allocation across invoice items, delinquency thresholds, commission splits. These are assertions against expected ledger state, the shape of test Karate was built for.
One engine, from UI to rating
The agent adapts through every Guidewire update. Each fix goes to a human before the test changes, so a heal never hides a regression.
A purpose-built adapter drives custom LOB and Gosu code directly. That is why Guidewire embedded Karate in its Testing Framework.
REST, SOAP, messaging and contracts. Prove the integrations around PolicyCenter still agree, in one syntax.
A digital twin encodes the rate book as a live oracle actuaries can read and edit. The agent does not just complete the quote, it proves the premium is right, row by row.
Requirement traceability and a deterministic ship / no-ship verdict, computed by code. Audit-ready and reproducible.
Self-hosted in your VPC or air-gapped, BYO-LLM (incl. Azure OpenAI). Policyholder data and your model never leave your network.
Running Guidewire as part of a wider estate? See the full insurance estate →
On self-managed Guidewire, you chose when to upgrade and could freeze the suite around it. On Guidewire Cloud you do not. Updates arrive on Guidewire’s schedule, and the regression suite has to be green against a platform that moved without your involvement.
This is the specific reason selector-based automation struggles here. A CSS or XPath locator encodes the shape of a screen that Guidewire generates from metadata, so a change in configuration or a platform update invalidates locators in tests whose underlying business logic did not change at all. The failures are real failures, but they are not defects, and teams learn to ignore them. That is the worst possible outcome for a regression suite.
Working from display text removes most of that coupling. When something does move far enough to matter, the agent proposes the fix rather than applying it, so the suite adapts without quietly losing its meaning.
Two insurers running the same Guidewire version do not have the same application. The difference lives in the product model, in configuration, and in custom Gosu, which is where the underwriting rules, the validation logic and the integrations to legacy policy administration actually sit.
A tool that can only operate the browser is testing the thin part. The deep adapter drives custom Gosu and LOB code directly, which is the capability Guidewire needed when it built its own Testing Framework on Karate. It is also the honest answer to why a generic AI browser agent pointed at PolicyCenter produces demos that look good and coverage that is not.
Most Guidewire test suites assert that a quote was produced. Very few assert that the quote was correct, because doing so means reimplementing the rate book, and a hand-maintained spreadsheet of expected premiums goes stale within a release.
The digital twin encodes the rate book as a live oracle that the test can query, so the assertion becomes a comparison against a maintained source of truth rather than a hard-coded figure. Actuaries can read and change it. For a rate change, that turns the regression question from “did anything visibly break” into “does every affected premium still compute correctly”.
A pass rate is not evidence. What an audit asks for is which requirement was covered, by which test, against which build, with what result, and who approved the exceptions.
Every run produces requirement traceability and a deterministic ship or no-ship verdict computed by code rather than inferred by a model, so the same inputs always produce the same answer and the record reproduces later. Because the whole stack is self-hosted, that evidence never leaves your environment. See security and compliance for the full posture.
The Guidewire Testing Framework (GT Framework) is Guidewire’s own testing tooling for Guidewire Cloud. Guidewire selected Karate as the technology underneath it. In Guidewire’s words, it decided to leverage Karate Labs for its Testing Framework because Karate is a powerful, open-source, industry-standard technology designed to make testing automation simple and accessible. Karate Labs was the first Solution partner to join the Guidewire Marketplace with testing technologies, and Karate’s technology is embedded as the foundation for Guidewire’s API testing for all Guidewire Cloud customers.
Both. The testing approach is the same either way: the agent drives PolicyCenter, ClaimCenter and BillingCenter through the browser using display text rather than brittle CSS or XPath selectors, and exercises the underlying REST and SOAP APIs directly. On Guidewire Cloud, where update timing is set by Guidewire rather than by you, the governed self-healing behaviour matters more, because screens shift underneath tests you did not change.
Yes. A purpose-built Guidewire adapter goes beyond the browser to drive custom Gosu and line-of-business code directly. This is the capability that separates Guidewire testing from generic web test automation: most of what makes one insurer’s PolicyCenter different from another’s lives in configuration and Gosu, not in the stock screens, and a tool that can only click through the UI cannot reach it.
Every proposed heal is presented for human approval before the test changes. The agent explains what moved and what it intends to update, and a person accepts or rejects it. Ungoverned self-healing is a genuine risk in a regulated environment: a tool that silently rewrites a failing assertion to make it pass will eventually paper over a real defect. Karate treats healing as a reviewed change, not an automatic one.
Yes. A digital twin of the backend encodes the rate book as a live oracle, so the test does not merely confirm that a quote reached the bind screen; it confirms the premium is arithmetically correct, row by row. Completing a submission proves the workflow ran. It does not prove the number at the end of it is right, and in insurance the number is the product.
No. Karate runs entirely self-hosted, inside your VPC or fully air-gapped, and you bring your own LLM, including Azure OpenAI or a local open-weights model. There is no Karate cloud service and no hosted control plane. Policyholder data, test evidence and your model weights stay inside your perimeter.
Generic tools treat Guidewire as an anonymous web application: they can click through screens, but they cannot reach Gosu or line-of-business code, they cannot verify rating, and they have no relationship to the platform’s release cadence. Karate is the technology Guidewire itself selected for its own Testing Framework. The difference shows up in the parts of Guidewire that are not the UI, which is where most insurer-specific risk actually sits.
Self-healing PolicyCenter, ClaimCenter and BillingCenter testing, on your own infrastructure with your own LLM.