Cytora Autopilot Brings Self-Running Agentic Workflows to Insurers

Cytora launched Autopilot on March 17, 2026, an agentic platform that orchestrates underwriting and claims in one layer while holding persistent context across every channel a submission arrives on. Underwriters and claims handlers currently spend up to 50% of their time on manual assembly work. Within eight weeks Duck Creek and Guidewire had shipped competing agentic products on two different architectures, and the choice between the three is a governance choice.

Key Takeaways

• Persistent workflow context is the architectural claim: Autopilot links Monday's submission, Friday's loss run and the following week's phone call into one continuously updated risk record rather than three isolated events.
• Up to 50% of underwriting and claims time goes to manual review, gap identification and broker follow-up, which is the overhead the platform targets.
• More than 140 languages out of the box, across email, API, audio and video intake, which matters for portfolios where submissions arrive locally and guidelines are held in English.
• Applied Systems acquired Cytora in September 2025 above $300 million, giving a platform with roughly $43 million of prior funding access to more than 40,000 agencies and brokers.
• Straight-through processing rates move from 10-15% to 70-90% in AI-enabled small commercial operations, which is where the expense effect is realized or not.

What the Platform Actually Holds Onto

Commercial submissions do not arrive as documents. They arrive as a sequence.

A broker emails a submission on Monday. The underwriter requests data on Tuesday. A partial response lands Thursday, with the loss run in a separate email on Friday. A phone call the next Monday clarifies the property schedule. Every one of those carries information the underwriter has to reassemble manually before pricing anything.

Conventional automation treats each as an isolated event. It can parse an email or extract a table, but it has no memory linking Monday's submission to Friday's attachment. That fragmentation is why insurance AI has stayed at the level of point solutions.

Autopilot keeps the workflow state instead: which data elements are missing, what has arrived, what is still outstanding, and it advances when new information appears rather than stopping at incompleteness. Richard Hartley, Cytora's CEO, framed it in the launch announcement as systems that understand context and execute as the picture of a risk develops. The intake is multi-modal across email, API, audio and video, in more than 140 languages, and recorded broker calls become part of the persistent record rather than one person's memory.

The distribution changed at the same time. Applied Systems acquired Cytora in September 2025 in a deal above $300 million. A platform with roughly $43 million of total funding, serving Allianz, Beazley, Markel, Starr and HDI, now sits inside a network connecting more than 40,000 agencies and brokers.

Three Architectures, and Only One Checks Itself Before Production

The competing products differ on where the intelligence sits, and that determines what evidence a carrier ends up holding.

Factor	Cytora Autopilot	Duck Creek Agentic AI	Guidewire ProNavigator
Architecture	Core-system-agnostic overlay	Core-system-native (MCR + 5 layers)	Embedded assistant in InsuranceSuite
Scope	End-to-end workflow orchestration (UW + claims)	End-to-end with neuro-symbolic governance	Role-specific intelligence at point of decision
Best fit	Multi-vendor environments, legacy core systems	Carriers on Duck Creek or planning migration	Carriers on Guidewire Cloud
Persistent context	Yes, across all communication channels	Within core system transaction data	Within individual application sessions
Lock-in risk	Low (sits on top of any core system)	Moderate (deepest value requires Duck Creek core)	Moderate (tied to InsuranceSuite/InsuranceNow)
Governance	Chain-of-thought audit trails per workflow step	Built-in AI Assurance layer with neuro-symbolic checks	Role-based access controls, citation-grounded
Distribution	Applied Systems network (40,000+ agencies)	370+ carrier relationships	Guidewire Cloud installed base

Cytora is core-system-agnostic. It runs over Guidewire, Duck Creek, Majesco or a mainframe without a migration, which is the broadest addressable market and the lowest lock-in. It achieves context by intercepting communications rather than through native integration, so the price is API surface and maintenance.

Duck Creek, announced at Formation '26 on April 28, 2026, embeds intelligence in its own policy, billing and claims systems through a five-layer architecture anchored on a Model Context Repository. The distinguishing feature is not the models. Every AI output is checked against deterministic rules before it reaches production, producing a trail that maps a decision to specific policy language, rating tables and regulatory constraints. Duck Creek reports 370 or more customers including 33 of the top 50 North American insurers, on over $150 billion of annual premium.

Guidewire ProNavigator is narrower by design: an assistant inside InsuranceSuite and InsuranceNow that surfaces role-specific intelligence at the point of decision, with role-based access control and citations back to the insurer's own source material. It assists a user rather than coordinating a workflow.

For an actuary the distinction that matters is between a workflow log and a decision trail. A record of what the system did is not a record of why it was permitted to. Only the neuro-symbolic approach grounds an output against filed rules before it executes, which is the artifact a rate filing or a model validation actually needs.

Where the time saving lands is the other open variable. Straight-through processing rates in AI-enabled small commercial operations run 70-90% against 10-15% previously. If a carrier uses that to cut headcount, it captures an expense ratio improvement. If it redeploys underwriters onto complex risks and relationships, it captures the expense improvement and a loss ratio improvement from better selection on the accounts that were never going to process straight through. The same 50% figure produces two different P&L shapes, and only one of them shows up in the loss ratio.

A Vendor Audit Trail Is Necessary and Not Sufficient

Autopilot records what data was used, what reasoning was applied and what outcome was produced for every workflow step, embedded in the execution path rather than bolted alongside it. That is real, and it does not discharge the actuary's obligation.

The regulatory calendar makes the point concrete. The EU AI Act classifies life and health underwriting AI as high-risk under Annex III, with transparency, explainability and human oversight requirements in full effect from August 2, 2026, while claims processing sits in a grey zone that carriers operating across EU jurisdictions still have to plan around. In the US the NAIC's 12-state evaluation pilot runs January through September 2026 and looks specifically at accountability assignment for autonomous decisions, cascading error risk across multi-agent systems, and whether human-in-the-loop escalation is adequate for high-risk uses.

Under ASOP No. 56 the actuary must understand the models the work relies on. A vendor-supplied trail satisfies the documentation requirement and not the verification one: the actuary still has to establish independently that the trail accurately represents the decision logic, and that the automated outputs are consistent with filed rates, policy forms and jurisdictional requirements. Those are different questions, and the platform can only answer the first.

Graduated automation is the practical hedge. Routine submissions inside predefined eligibility criteria process straight through; anything above a complexity threshold requires human review before binding. That calibration is where the actuary's judgment actually enters, and it is the step the adoption data suggests carriers are finding hardest: 7% of insurers have reached full operational scale against 82% reporting some level of AI adoption. The gap is not model quality. It is the layer that decides which decisions the model is allowed to make alone.