Automation in claims is easy to do badly. The fastest system is the one that denies everything. The kindest is the one that pays everything. Neither is a claims operation. Claimwell exists to make the right decision quickly, explain it plainly, and keep a person accountable for it.
The same evidence produces the same recommendation, whoever the claimant is and whichever examiner is on shift. Every determination cites the plan provision and the pages it rests on. Appeals go to someone who was not in the room the first time.
A claim has three parties. The claimant needs a timely, plain-language answer and a payment that arrives. The employer needs their person back. The carrier needs duration managed to the guideline and no leakage. Claimwell is built to serve all three, not to minimize one.
Human time is the scarcest thing in a claims shop. It should be spent on judgment, not on chasing a fax, re-keying a date, or reading forty pages to find one. Agents do the chasing. People do the deciding.
These are not aspirations. Each one maps to something enforced in code, versioned in configuration, or measured in the audit store.
An agent can recommend a denial, a termination, or a reduction. It can write the letter. It cannot sign it. The workflow will not advance and the system of record will not accept the write without a recorded decision by an examiner.
This is not a policy or a prompt. It is a guard that the model cannot see as something it could change.
Everything the agents do up to the point a human signs is Claimwell's to get right: indexing, extraction, matching, chasing, calculation, clocks, and the recommendation with its package. If that work is wrong, that is on us.
That is also why we operate the agents rather than hand over the keys. A named Claimwell team runs them inside your cloud, replays every release against the scenario library, and answers for what they do. You cannot stand behind a process someone else is running. And we put money behind the four rules we can prove claim by claim: the warranted rules.
Every determination is the carrier's, because a person the carrier employs signs it, and so is the strategy that sets the dials. Claimwell never claims responsibility for the outcome of a signed decision, because a vendor that did would make the signature theater. Regulatory accountability stays where the law puts it. What we do is make the decision you own smaller, better evidenced, and easier to defend.
Buy the Mail Room or Claim File Intelligence and Claimwell is not in the decision loop at all. It indexes, it annotates, it cites. Your decisioning framework is untouched and your people decide exactly as they do today.
Add an agent and Claimwell is in the loop, so the rule changes shape without changing meaning: the agent recommends, a human signs. The step from one to the other is explicit, reviewed by your compliance team once, and recorded in the strategy your leadership owns.
Nothing lands on a claim without a reason and a reference: the plan section, the page of the physician statement, the HRIS record, the guideline row. A claimant, an auditor, or a regulator can follow the same trail the examiner did.
Benefit amounts, elimination periods, offsets, tax treatment, decision deadlines, extension windows, and appeal clocks come from rule engines that print their working. A language model is good at reading a messy fax. It is the wrong tool for arithmetic that a regulator will check.
Segmentation, straight-through rules, contact cadence, clinical triggers, and which handoffs are human are a declarative playbook your strategy owner edits and publishes. Every claim records the version it ran under. You can replay last quarter under a new version before you commit to it.
The target is the right duration for the diagnosis and the job, with return to work planned early and modified duty asked about. Extensions within the guideline are routine. Extensions beyond it get a nurse and an examiner. Nobody is rewarded for closing early.
Claimwell runs inside your account. Models are open-weight and served on your GPUs by default. If you choose a frontier model for a task, it is an explicit setting through your own provider contract, and the trace shows exactly what left the building.
The evaluation harness runs scripted scenarios with late providers, wrong dates of birth, duplicate faxes, and missing pages, and checks every outcome, dollar, deadline, and letter. Zero constraint violations is a hard pass condition. You run it in your own account on synthetic data first, and after go-live every release, strategy version, and model change is replayed against it before it reaches a claim, with the results written to your audit store.
Every metric comes from the audit store, never from agent self-report. The ones that matter for fairness are reported alongside cycle time and cost, and cut the same way.
Same evidence, same recommendation. Replays of the same claim under the same strategy must agree. Disagreement is a defect, not a feature.
How often examiners reject an agent recommendation, by handoff point and by reason. A rising rate is a signal to fix the agent, not to lean on the examiner.
Overturn rate by original decision type and segment. An independent reviewer and a clean file should make overturns rare and explainable.
Days from first notice to a claimant hearing something true. Regulatory deadlines are a floor, not a target.
Every row links to the claims behind it. Nothing here is computed from what an agent said about itself.
Some features sell well and make claims worse. We leave them out on purpose.
Bring a plan document and a scenario you find hard. We'll show you where the agent stops and the examiner starts.