Three business workflows, one company model

Workflow Architecture

Auto claims, commercial underwriting, and subrogation — each exposing different seams where translation debt accumulates when AI inserts into the flow.

01

How the three workflows connect

Claims is the spine. Subrogation hangs off the far end of it. Underwriting runs in parallel, feeding the same agents and the same organizational context.

Primary value stream

Auto Claims

FNOL intake to closure, with two pipelines — AI for simple claims, human for moderate and complex. The most mature, most deeply modeled workflow. It is the largest surface for AI insertion and the source of the subrogation cases downstream.

Connections
Claims → Subrogation. Every paid claim with recovery potential becomes a subrogation case — the critical seam of the whole experiment.
Underwriting runs parallel. Separate value stream, but shares agents, metrics, and the same company context.
Downstream recovery

Subrogation Recovery

Post-claim recovery from at-fault parties. Originates from a closed claim, so it inherits every upstream failure. Highest seam risk in the model — recovery depends entirely on documentation it never produced itself.

Parallel value stream

Commercial Underwriting

Application to bind. Guidewire PolicyCenter-aligned. Models the sales cycle and the broker → underwriter → actuary translation chain. AI augments risk scoring but can’t do the actuarial sign-off or the quote.

The stakes of the connection: underwriting is a separate value stream, so its failures don’t contaminate claims. But subrogation is downstream of everything — a sloppy settlement decision in claims destroys recovery potential weeks later, and Tommy has no standing to fix it.

02

Auto Claims — the primary workflow

FNOL intake → triage → investigation → damage estimation → coverage verification → liability → settlement → approval → payment → subrogation → closed. Two pipelines split at triage.

01
FNOL IntakeGATE · classify
02
Triage — route by complexityGATE · classify
03
AI LANE · simple claimsDamage estimation (AI, Maria reviews) → Coverage verification (AI, Pat reviews) → Liability determination (AI, Maria reviews) → Settlement (AI, <$5K auto-approve)
04
HUMAN LANE · complex claimsInvestigation (Diana) → Damage estimation (Maria) → Coverage verification (Pat) → Liability + settlement (Diana)
05
Approval gateGATE · approve — <$5K auto · $5–100K Sanjay · >$100K Kathryn
06
PaymentAI · PAY — triggers subrogation review
07
SubrogationHUMAN · NEGOTIATE — Tommy, recovery
08
ClosedGATE · classify
Where AI inserts
AI owns the fast lane; humans own the judgment lane. Almost every step is AI-capable, but each has a confidence threshold below which it drops to a human exception path. Triage decides the lane: simple (single vehicle, no injury, clear liability, under $10K) goes AI; everything else goes human. AI runs FNOL intake, triage, photo damage estimation, coverage verification, liability determination, settlement under $10K, auto-approval under $5K, payment. Human only complex damage, disputed liability, negotiation, and everything over the confidence thresholds.
The seams AI exposes
Seam risk scores every handoff. The most expensive are where AI collides with institutional judgment.
COVERAGE 0.8 Policy-language interpretation. Pat’s expertise is a black box — AI reads exclusions literally that specialists know are never enforced.
DAMAGE ESTIMATION 0.7 AI uses database labor rates 6–18 months stale; adjusters mentally correct for it. When AI bypasses that, supplement requests spike.
APPROVAL 0.7 Sanjay trusts Diana’s informal review implicitly — rejection rate jumps 18% when AI skips her.
SUBROGATION 0.9 Downstream of everything. Incomplete files and missing liability documentation compound here.
03

Commercial Underwriting — the parallel stream

Application submission → data gathering → risk assessment → actuarial pricing → quote generation → bind/issue. Guidewire PolicyCenter-aligned. The translation chain is broker → underwriter → actuary → carrier.

01
Application intakeCLASSIFY — Alicia, broker submission
02
Risk assessmentASSESS — Nick, AI scores, Nick smells
03
Actuarial pricingASSESS — Greg, technical price
04
Quote generationNEGOTIATE — Nick, market price
05
Bind / issueAPPROVE — Alicia, policy language
Where AI augments
vs replaces
AI scores the data; the smell test stays human. AI parses applications and scores risk factors, but the tacit knowledge that keeps a book profitable lives in Nick’s head. AI pricing models produce a technical price; only a human bridges it to a competitive quote. AI augments application intake and classification, risk-factor scoring, pricing models, policy documents at bind. AI cannot replace quote generation and the actuarial statement of opinion — Greg won’t certify a model he hasn’t validated, and state DOI won’t accept uncertified rates.
The seams
Every translation strips context from the risk narrative.
RISK ASSESSMENT 0.8 The Type 2 absence: AI doesn’t know to check adjacent-building risk, prior-carrier non-renewal reasons, or application fill time. AI-written risks run 4–8% worse loss ratios after 18 months.
APPLICATION INTAKE 0.7 Broker language → underwriter language. AI prices “well-managed restaurant” the same as any restaurant.
BIND 0.7 “Full coverage” to the broker is six exclusions in the policy. AI-generated language may not match negotiated terms.
PRICING 0.6 Technical price → quoted price. Greg won’t sign off on a rate that isn’t actuarially justified, but the market may need a different number.
04

Subrogation Recovery — the downstream seam

Post-claim recovery from at-fault parties, originating from a closed claim: recovery assessment → liability review → demand package → negotiation → collection → closed.

01
Recovery assessmentASSESS — Tommy, screens closed claims
02
Liability reviewVERIFY — Rachel, legal basis
03
Demand packageNEGOTIATE — AI assembles from structured data
04
NegotiationNEGOTIATE — Rachel, up to $100K
05
CollectionPAY — Tommy, offsets paid claim
06
ClosedCLASSIFY
Why it’s the highest
seam risk
It’s downstream of everything. Subrogation starts with a paid claim it didn’t handle. Recovery potential exists only if the original adjuster documented liability evidence — and AI pipeline decisions often don’t include a “should we check for recovery?” flag. Tommy screens what he can see, not what should be there.
ASSESSMENT 0.9 The core Type 2 absence. 12–18% of claims with recovery potential close with $0 because the system never asked who else is liable.
UPSTREAM DEBT Incomplete files and missing police reports kill demand strength — recovery rate drops 30% when files lack evidence.
AI inserts screening and demand-package assembly from structured data — but can’t do liability review or negotiation.
Carrier tactics Rachel’s team knows which at-fault carriers pay fast and which fight everything. AI treating all carriers the same drops the recovery rate.
Jurisdiction
complexity
State laws vary. Liability review and negotiation depend on state-specific subrogation law — comparative-negligence rules, anti-subrogation doctrines, differing recovery rights. The demand package must reflect the governing state; disputes that can’t settle move to a legal referral.
Handoff to legal (0.6): claims → legal is a domain boundary with different priorities — the slowest, least flexible seam in the model.
05

How they interact

The handoffs that define the whole experiment.

Claims → Subrogation is the critical seam. Payment triggers a subrogation review. Whether recovery is even possible depends entirely on what the claims pipeline documented and whether anyone flagged third-party liability. The single most expensive translation in the model.
Underwriting is parallel — but not isolated. A separate value stream that shares the same agents, metrics, and organizational context. Its failures stay in underwriting; claims failures bleed into subrogation. Different exposure to downstream contamination.
WorkflowPositionAI insertionsPeak seamHeadline metric
Auto ClaimsPrimary value streamBoth pipelines, nearly every stepSubrogation 0.9
Coverage 0.8
Cycle time, cost per claim, loss ratio
UnderwritingParallel value streamAugments scoring; cannot replace quote or actuarial sign-offRisk assessment 0.8Written-book loss ratio, adverse selection
SubrogationDownstream of claimsScreens recovery, assembles demand; liability and negotiation stay humanAssessment 0.9Recovery amount, recovery rate

The thesis in one line: AI transforms each workflow at its seams — and the seams that matter most are where institutional judgment, informal controls, and downstream dependencies live. Subrogation is where that thesis bites hardest.