Most underwriting automation stops at a score. It reads a submission, returns a number, and hands the case back to a person to do everything else. Agentic AI changes the unit of work. Instead of automating a single step, an agent takes ownership of the whole quotation journey and pulls in a human only when the case genuinely needs judgment. For insurers in Brazil's Seguros e Danos (P&C) market, where underwriters lose roughly 40% of their time to administrative tasks according to Deloitte, that shift is the difference between software that assists and software that actually clears the queue.
What agentic AI in insurance underwriting means
An agent is software that can perceive, decide, and act with enough authority to finish a task, not just advise on it. In underwriting, that means it reads the incoming submission, scores the risk against the insurer's policy, prices it, and then quotes, declines, or escalates the case, writing the outcome back to the system of record. A plain model only outputs a score and waits. A rules bot only moves data between screens. An agent reasons over the messy input a broker actually sends and carries the decision to a conclusion.
The part that makes underwriters comfortable is control. The agent is calibrated to your risk appetite and your underwriting manual, not to a generic template. It does not set your appetite. You do, and the agent enforces it, on every case, exactly as the manual defines it. Anything outside the rules it is given goes to a human with the reasoning attached.
How an agentic underwriting workflow runs
A working agent runs the same sequence a good underwriting desk runs, only faster and without the manual handoffs. It starts with multichannel intake and automatic validation, accepting the submission in whatever format the broker already uses, whether that is e-mail, an attachment, or an API call. Intelligent document reading then extracts the fields from unstructured paperwork, the part of the job where Gartner estimates companies lose 20% to 30% of their time. Next comes broker enrichment and prioritization, cross-referencing history and exposure so the best business surfaces first. This is the same discipline a dedicated submission triage layer brings to the front of the funnel.
From there the risk and fraud engine scores the case with machine learning calibrated to appetite, dynamic pricing produces a risk-adjusted premium, and the decision stage issues a quote, an automatic decline, or an escalation to a human. Every outcome is written back to the core and returned with a full audit trail. The underwriter stops being a data clerk and goes back to pricing complex risk and winning business.
Agentic AI versus rules engines and RPA
Rules engines and RPA were the last generation's answer to the same problem, and they are brittle. A rules engine breaks the moment a submission arrives in a new format, and an RPA bot fails when a screen it depends on moves. Neither can reason about a risk it has not seen before. An agent reads unstructured input, adapts, and applies judgment inside the boundaries you set, which is why it can handle the long tail of real submissions instead of only the clean ones.
Just as important is where the agent sits. It is designed to work as an external layer on top of the existing core rather than replace it. That is not a technicality. BCG has found that about 70% of insurers do not carry their innovation initiatives through because of IT limitations, so an approach that needs no core migration and puts no load on the insurer's IT team is often the only version of the project that actually ships.
Keeping agentic decisions auditable under SUSEP and LGPD
An agent that makes decisions on regulated business has to be able to explain each one. In Brazil that means two things at once. Under the LGPD, a data subject has the right to review decisions taken by automated processing, so the reasoning behind a decline cannot be a black box. Under SUSEP conduct and governance expectations, the insurer has to show that its acceptance decisions follow a consistent, documented policy. Both point to the same design. Every quote, decline, or escalation should leave an auditable decision record that a regulator, an internal auditor, or an ombudsman can query, with the data encrypted at every step.
The goal is not a model that is never wrong. No one in insurance should promise that. The goal is a decision you can always explain, calibrated to a manual you approved, with a human in the loop wherever the risk warrants it.
Where agentic underwriting is heading
The direction of travel is not subtle. McKinsey's widely cited Insurance 2030 scenario envisions manual underwriting ceasing to exist for most personal and small-business products by 2030. That is a scenario rather than a schedule, but the momentum behind it is real. Gartner named agentic AI its top strategic technology trend for 2025 and projects that by 2028 roughly a third of enterprise software applications will include agentic AI, up from less than 1% in 2024.
The competitive pressure is just as concrete on the distribution side. Capgemini reports that more than 60% of brokers choose an insurer based on response speed. In a Seguros e Danos market that grows double digits a year, the carriers that let an agent clear standard business in minutes build a service advantage the slower ones cannot easily close.
How WIR applies agentic AI to underwriting
WIR Innovation is built as exactly this kind of external AI layer for insurers and brokers. It runs the full quotation and underwriting journey, from multichannel intake through intelligent reading, enrichment, a risk and fraud engine calibrated to the insurer's appetite, dynamic pricing, and a final decision that is quoted, declined, or escalated to a human. Each decision is explainable, written back to the core, and returned with an audit trail, with data encrypted at every step and handled in line with the LGPD. Because the platform is 100% external, there is no core migration and no load on the insurer's IT team, and implementation runs as a defined setup of three to twelve months before continuous operation.
The approach is being proven in practice. WIR is currently running a proof of concept with a global insurer in the Transport line. The point of the platform is not to remove underwriters from the decision. It is to let the agent handle the standard, high-volume flow to policy, so the people keep their judgment for the risk that deserves it.
Perguntas frequentes
What is agentic AI in insurance underwriting?
It is software that runs the underwriting task end to end instead of only scoring it. An agent reads the submission, enriches it, scores the risk against the insurer's appetite, prices it, and then quotes, declines, or escalates to a human, writing a full audit trail back to the core system. It acts with authority inside the rules the insurer sets, rather than only advising an underwriter.
How is agentic AI different from RPA or a rules engine in underwriting?
RPA moves data between screens and a rules engine applies fixed if-then logic, so both break when a submission arrives in a new format or presents a risk they were not programmed for. An agent reasons over unstructured input, adapts to cases it has not seen, and carries the decision to a conclusion inside the boundaries the insurer defines.
Does agentic AI replace the underwriter?
No. The agent handles the high-volume standard flow so underwriters spend their time on complex risk and business development. Anything outside the rules it is given is escalated to a person, with the reasoning attached, and the insurer keeps the final say on appetite and authority.
Does an agentic underwriting system replace the core policy system?
No. A well-designed agent works as an external AI layer that sits on top of the existing core, reads submissions, and writes decisions back to it. There is no core migration and no load on the insurer's IT team, which is what lets these projects ship despite the IT constraints most carriers face.
How are agentic underwriting decisions kept compliant with SUSEP and LGPD?
Every decision is explainable and leaves a complete audit trail that a regulator, auditor, or ombudsman can query, with data encrypted at every step. That supports the LGPD right to review automated decisions and the SUSEP expectation that acceptance decisions follow a consistent, documented policy. The system escalates borderline cases to a human rather than promising a model that is never wrong.