An agent layer, not a replacement.
Agents need guardrails, human escalation, audit trails and grounded data. A process platform is the only responsible place to put them, and PSuite already has all four.
Agents need what BPM already built
The hard parts of putting an agent into a regulated workflow are not the model. They are everything around it, and each one is an existing PSuite construct.
Rules become guardrails
Underwriting, product and compliance logic stays deterministic and auditable. The agent proposes; the rule engine disposes. Nothing an agent suggests bypasses a rule.
Human tasks become escalation
Human-in-the-loop is the hardest thing to retrofit into an agent framework. PSuite already has a task inbox, RBAC and lifecycle management waiting for the handoff.
Events become the audit trail
Every agent action lands as a business event. Regulators do not accept “the model decided”. They accept a trace, and the Event Manager already produces one.
Service tasks become tools
A service task is a tool call by another name. Every business function already orchestrated and exposed over REST is a capability an agent can be granted, or denied.
Information models become grounding
Standardized, reusable information models give an agent a typed, validated view of the case. Grounded context is the single biggest defence against a confident wrong answer.
Process models become bounds
Unbounded autonomy is what makes agents unshippable in insurance. A process model is a pre-agreed boundary on what may happen, in what order, and by whose authority.
Agents make BPM adaptive; BPM makes agents accountable.
From flowchart to contract
Today a process model prescribes every path. An agentic model declares intent, boundaries and capabilities, then lets an agent choose the path inside them.
The runtime you already have keeps its job
The new layer sits above it and is only ever granted capabilities the process model allows.
Existing REST APIs as tools
Agents call tools; they never touch the database.
Every call is an event
The trace is a by-product of running, not an add-on.
Model-provider neutral
The gateway abstracts the provider, so the choice stays yours.
Life new business, re-imagined
The same platform, the same rules, the same workbench, with agents doing the reading, chasing and drafting that currently makes a case sit still.
Intake Agent
Reads unstructured submissions (PDF applications, agent emails, illustrations) and maps them onto the information model, resolving ambiguity at the door. Fixes not-in-good-order rather than rejecting it.
Requirements Agent
Determines which evidence is actually needed for this product, state and face amount, orders it, chases the vendor, and tracks aging. Pure latency work, removed.
Summarization Agent
Reads the APS, labs and prescription history and produces a structured, cited underwriting summary against the information model. It never decides; it proposes.
Underwriting Copilot
Proposes a class and decision with reasoning and citations. Deterministic underwriting rules validate it; disagreement or low confidence routes straight to an underwriter.
Correspondence Agent
Drafts agent and customer communication in the context of the case, gated by the same compliance and document adequacy rules that govern generated correspondence today.
Case Watch Agent
Subscribes to business events, spots cases that have stalled or are aging toward an SLA breach, and takes or recommends the next best action.
No agent issues a policy. The rule engine issues the policy. An agent’s job is to move a case from ambiguous to decision-ready.
Agents move the boundary, not just the cases
Classic STP tuning moves cases between auto-issue and manual. This is different.
The straight-through population grows
Today a case goes manual for two different reasons: genuine underwriting judgement, or missing and unstructured information. Only the first is real. Agents attack the second, so the set of cases that can be automated expands rather than the rules simply being loosened.
Underwriter time moves up the value chain
Reading hundreds of pages of medical records is not underwriting judgement; it is preparation for it. Handing preparation to an agent gives the underwriter a cited summary and a proposed decision to accept, amend or reject.
Copilot
Agents draft and summarize inside the existing workbench. Every output is reviewed by a human. Zero autonomy, immediate time savings, and a corpus of corrections to learn from.
Supervised autonomy
Agents act on low-risk, high-volume steps (requirements ordering, chasing, routine correspondence) with rules gating every action and events recording each one.
Bounded autonomy
Agent Tasks run inside modelled boundaries with budgets and acceptance tests. Confidence thresholds decide what escalates. The underwriter handles judgement, not preparation.
Start at Stage 01
A copilot pilot inside your existing workbench is a weeks-not-quarters engagement, and it produces the corrections corpus that everything after it depends on.