Built for disability and workers' comp carriers

Claims that run themselves.
Decisions that stay human.

Claimwell takes over claims work one rung at a time, inside your system of record and your cloud. Start by indexing every document the day it arrives. Grow to agents that run each stage with your examiners signing every decision that matters.

Runs inside your perimeter. Never makes an adverse decision.

CW-204118 Lumbar strain · Light duty · Northwind Traders
segment: moderate day 6 · approved 1 human touch
01day 0
Mail Room
Faxed claim form matched
02day 0–4
Intake
Eligibility verified from HRIS
03day 4
Evidence
APS extracted, cited by page
04day 5
Decision
Within guideline · examiner reviewed
05day 6
Benefit & pay
6 weekly payments scheduled
06wk 2–6
Ongoing
Recert · RTW outreach
07wk 6
Close
Return to work confirmed
Integrates with Guidewire ClaimCenter ClaimVantage FINEOS Duck Creek Workday Origami Risk Amazon Bedrock
Runs in your account AWS Google Cloud Azure Open-weight models on your GPUs No claimant data leaves your network Deploy in your cloud

Why carriers choose Claimwell

Most "AI for claims" is a copilot bolted onto an examiner's screen. Claimwell is the examiner's team: agents that own the file, chase the evidence, calculate the benefit, and hand you a decision package when a human is required.

Agents own the claim, humans own the decision

Every lifecycle stage has one owner. Agents arrive one at a time, in a fixed order, and run their stage. Examiners get decision packages, not alerts.

No adverse decision without a human

Standalone products never enter the loop. In the platform, denials, terminations, reductions, appeals, and banking changes are blocked in code until an examiner signs.

Your cloud, your weights

Deploys into your AWS, GCP, or Azure account with Terraform. Open-weight models run on your GPUs. Frontier models are an opt-in route, never the default.

Plugs into the system of record you already have

Agents are integrators. They work through one claims-admin port with adapters for Guidewire, ClaimVantage, FINEOS, and your HRIS, payment, and correspondence systems.

Every action cited, every step replayable

Each write carries a rationale and page-level evidence citations. Every model call is logged with its prompt version and can be replayed against a different model.

Clocks and money are deterministic

ERISA deadlines, statutory offsets, elimination periods, and benefit math come from rule engines, never from a model. The explanation is printed with every number.

Intake and evidence

A front door that never sleeps

Claims and documents arrive the way they really do: scanned mail, fax, email, portal uploads, employer feeds, call notes. The Mail Room agent classifies, matches, and routes each item. The Evidence agent keeps the file reconciled for the life of the claim.

Virtual Mail Room

Every channel, one queue. Duplicate faxes, cover sheets, and junk mail are discarded with an audit record. Ambiguous matches go to a human.

Extraction with citations

Layout-aware OCR with a vision model for handwriting and degraded scans. Every field carries a confidence and a page citation.

Evidence deltas, not raw documents

When an orthopedic report lands at week nine, the owning agent wakes to a summary of what changed, what conflicts, and what it implies.

Eligibility from the source

Identity, active-at-work status, hours, and coverage are verified against the HRIS census feed, with staleness visible.

See the Mail Room product
Mail Room · today's batch
Attending Physician Statement
fax · matched CW-204118 · confidence 0.97
routed
Claim form, 4 pages, skewed scan
scanned mail · new claim shell created
intake
Employer reply: last day worked
email · reference token matched
evidence
Attending Physician Statement
fax · duplicate of item 1 · discarded with audit
dup
APS, DOB mismatch on two candidates
fax · confidence 0.61 · below threshold
examiner
Items processed212 Auto-matched203 (95.8%) Sent to a human4
Human in the loop, by design

Three tiers of control. Your leadership holds the dial.

Cardinal rules are enforced in code and can't be relaxed by anyone. Your claims strategy is a versioned playbook your leadership edits without engineers. Agent judgment lives inside the room those two leave open.

Cardinal rules

No denial without a signature. No payment outside an approved period. No letter stating a decision that isn't on record. Checked at the workflow and at the write.

Versioned claims strategy

Segmentation, straight-through whitelist, contact cadence, clinical triggers, and sixteen named handoff points. Publish a version, replay last quarter, diff the result.

Decision packages

When a human point is reached, the examiner gets the recommendation, the rationale, the cited evidence, and the draft letter. Sign it or send it back.

Nurse and examiner queues

Clinical referrals, duration beyond guideline, and extensions route to the right persona with a prepared packet and a bounded ask.

See the Lifecycle Agents
Examiner queue · activity H1 locked by rule C1
Denial recommendation · CW-203877
Pre-existing condition exclusion · voluntary plan · complex segment
Treatment for the claimed condition appears in the 3-month lookback window visit note p.2 and the plan excludes it for 12 months after coverage effective certificate §5.1. Treating provider's statement does not dispute the onset date APS p.1. Recommend adverse determination with full ERISA content and 180-day appeal rights.
Plan provision and specific reason stated
Internal rules relied on disclosed
Right to records and appeal window included
Regulatory clock: 11 days remaining
Sign determination Send back with note
Ongoing management and outcomes

Managed to the guideline, closed on return to work

After approval the work isn't over. The Ongoing Management agent pays on schedule, requests recertification by segment, reaches out before the expected return-to-work date, asks the employer about modified duty, and flags long-term disability candidates early so nobody falls into a gap.

Cycle time to decision

Days from first notice to approval, pend, or recommendation. Cut by segment, cause, channel, and strategy version.

Duration against guideline

Paid duration versus optimal and maximum for the diagnosis and job class, as a distribution you can act on.

Human touches per claim

Counted by handoff point and persona, so you can see exactly where your examiners' time goes and move the dial.

Strategy replay

Change one setting, replay last month's claims under the new version, and compare the metrics before you publish.

How we measure fairness
Dashboard · STD · last 30 daysillustrative
Cycle time to decision
4.2d
median · all segments
Straight-through rate
38%
simple segment
Human touches / claim
0.9
by H-point
Within guidelineReturned to work earlyBeyond guideline, examiner reviewed
Built for three people

The examiner, the nurse, the strategy owner

Each has a queue, a workbench, and a bounded ask. Nobody reads a raw file. Nobody re-keys a date.

Opens the queue to find a denial package with the reason, the plan section, the cited pages, and the draft letter. Signs it, sends it back, or picks up the phone with a prepared script. Never writes a clock date by hand.

EX
Claims examiner
Adverse determinations, exceptions, identity mismatches

Receives a referral packet: structured clinical summary, restrictions against job demands, duration against guideline, and the specific question being asked. Answers it, recommends an IME if warranted, and the claim moves on.

RN
Nurse consultant
Clinical review, extensions beyond guideline

Widens the straight-through whitelist to include hernia repair, lowers the duration tolerance by five points, replays last quarter, and sees the change in touches and cycle time before publishing the next version. No ticket, no sprint.

CS
Claims strategy owner
Segmentation, cadences, handoff points, straight-through rules
Products

Start where the pain is. Grow to the whole lifecycle.

Claimwell is sold as a ladder, not a monolith. Two products stand alone and never touch a decision. The agents come after, one at a time, in a fixed order. The Harness proves all of it before go-live.

Deploy in your cloud

Inside your perimeter, on any cloud, in an afternoon

Claimwell ships as a deployable package: container images, a Terraform module per cloud, a Helm chart, a runbook, and an evaluation harness. Your platform team runs it in your account. We never hold your data.

Open-weight models on your GPUs

Served with vLLM in your VPC. Model traffic never leaves your network. Swap models with a config change.

Frontier models, opt-in

Route any task class to Amazon Bedrock or Vertex AI through the same gateway. Compare outputs side by side in the trace view.

No managed-service lock-in

Every managed service has an open equivalent in the compose stack. Cloud-specific code lives in one Terraform module and three thin adapters.

Audit store you can't edit

Append-only. Every constraint check, every model call, every human signature, with prompt and strategy versions attached.

Read the deploy and security page

AWS Google Cloud Azure On-premises Kubernetes
$ claimwell deploy --cloud aws --region us-east-1
✓ VPC, private subnets, TLS ingress for claims.yourcarrier.com
✓ GPU node · vLLM · open-weight 70B, quantized
✓ Workflow runtime · Postgres · S3-compatible object store
✓ Adapters: guidewire-claimcenter, workday-hris, payments
✓ Constraint checker armed · C1–C10 enforced
✓ Strategy 2026-11-01.1 published
✓ Evaluation harness: 11 scenarios · 0 constraint violations
$ claimwell deploy --cloud gcp --region us-central1
same package. same stack. different module.

Disability and workers' comp, one platform

The lifecycle, the control plane, and the evidence layer are shared. Each line brings its own stages, rules, and handoff points. Who it's for

Short-term disability

First notice to return to work

Fully insured STD with the complete money loop: eligibility, elimination period, benefit, statutory offsets, payments, recertification, extensions, closure.

  • Straight-through for whitelisted simple claims
  • ERISA clocks, extension notices, tolling
  • Appeals with an independent reviewer
Long-term disability transition

No gap, no surprise

A Transition agent assesses candidacy at the strategy's midpoint, gathers the LTD claim form and fresh physician statement, packages the file, and lines up the last STD payment against the LTD elimination period.

  • Candidacy flagged from duration trend
  • Examiner approves the recommendation
  • Package acknowledged by LTD intake
Workers' compensation

Lost time and medical-only

Same Mail Room, same evidence layer, same examiner workbench. Jurisdiction-specific indemnity rules, compensability, medical bill routing, and return-to-work coordination with the employer.

  • Compensability recommendation, human signed
  • State-specific waiting periods and rates
  • Modified-duty outreach built in

Frequently asked questions

Do we have to buy the whole platform?
No. The Mail Room and Claim File Intelligence stand alone, need one adapter to your claims system, and never touch a decision. The lifecycle agents are added one at a time in a fixed order, and the first one brings the control plane with it. Most carriers start with the Mail Room because it replaces a manual indexing operation with a two-week integration.
Does Claimwell deny claims?
No. The standalone products are not in the decision loop at all; your decisioning framework stays exactly where it is. In the platform, agents can recommend a denial, termination, or reduction and assemble the full package, but the determination requires a recorded signature from an examiner. That is enforced in code at the workflow and at the write to your system of record, and no configuration can relax it.
Where does claimant data go?
Nowhere. Claimwell runs inside your cloud account and your network. Models are open-weight and served on your GPUs. If you choose to route a task to a frontier model, that is an explicit, per-task-class setting through your own provider account, and the trace shows exactly what was sent.
Which models does it use?
Open-weight instruction-tuned and vision-language models, selected by the evaluation harness on your document mix. A gateway with one interface lets you compare any task against Amazon Bedrock or Vertex AI side by side, and switch with a config change.
What systems does it integrate with?
Claimwell's agents work through one typed claims-administration port with adapters for Guidewire ClaimCenter, ClaimVantage, FINEOS, and Duck Creek, plus HRIS feeds, plan repositories, payment, and correspondence systems. Intake channels include scanned mail, fax, email, portal uploads, employer feeds, and call notes. New adapters are built against a contract test suite. See all partners and integrations.
Does it handle workers' compensation?
Yes. Short-term disability is the reference implementation. The lifecycle, control plane, Mail Room, evidence layer, and examiner workbench are shared across lines, and the workers' comp stage set adds compensability, jurisdiction-specific indemnity rules, medical bill routing, and employer return-to-work coordination.
How long does deployment take?
The package deploys into a sandbox account from Terraform in an afternoon. Running your first real claims depends on adapter configuration for your system of record and HRIS, and on your strategy owner's first playbook version. Most of that is configuration, not code.
How do we know it works before it touches a real claim?
The evaluation harness runs scripted scenarios with simulated claimants, employers, and providers under accelerated time, and checks every outcome, benefit amount, clock date, and letter against expectations. You can run it in your account on synthetic data before any real claim is routed.

See Claimwell run on the claim your examiners dread

Bring a plan document and a scenario you find hard. We'll run it under accelerated time and show you where the agent stops and the examiner starts.

Synthetic-data sandbox available on request. No claimant data required to evaluate.