SOLUTIONS · INSURANCE
AI claims triage, trained on how your team already handles claims
Sort claim documents, emails and new claims into your own queues with a private model trained on claims your team already handled. People decide every claim.
Every morning, claims mail and new claims arrive faster than anyone can sort them: medical bills, police reports, repair estimates, letters from attorneys, first-notice reports with photos. Someone reads each one and decides where it goes. StayCharted trains a private model on the claims and documents your team has already handled, so each new one arrives with a suggested category or queue and a confidence. Your integration can route confident suggestions; the unsure ones go to a person.
For third-party administrators (TPAs), independent adjusting firms, managing general agents (MGAs) and regional carriers. Start with one intake job and the categories your team already uses.
Insurance document classification, email routing and first-notice triage
| Job | What it reads | What it answers, in your categories |
|---|---|---|
| Insurance document classification | The document's extracted text and file name | Medical bill · Police report · Repair estimate · Invoice · Attorney letter · Proof of loss · Correspondence |
| Claims email routing | Subject and body of an email to a claims or agency inbox | New claim · Status question · Payment · Supplement · Complaint · Broker request |
| First-notice triage | The loss description, line of business and the fields on the first report, with photos if you have them | Fast-track · Standard · Field adjuster · Large loss · Specialist; and a cause-of-loss code |
| Damage photos | Photos, with the adjuster's note or the customer's message | Damage type and a severity band, as your adjusters grade them |
Each comes with a confidence, so you choose how sure it must be before anything is routed without a person.
Why trained on your own claims
A general AI can tell a police report from a medical bill. Without your examples or instructions, it doesn't automatically know:
- your queues and who works them;
- your cause-of-loss codes, and where your team draws the line between two of them;
- which first-notice reports your team fast-tracks, and which go straight to a senior adjuster.
Your claims system already holds those decisions, claim by claim. A model trained on them applies the same rules the same way, every day, including the ones nobody wrote down.
How to set it up
1. Export what your team already decided
Pick one job to start, usually documents or emails. Export a year of them, one row per item, with the category or queue your team gave it. For documents, include the text your claims system or OCR step already extracted, and the file name. Leave out anything filled in after the decision.
Leave out personal characteristics (age, gender, ethnicity and the like) and anything your compliance team wouldn't want routing a claim.
2. Upload and check
Upload the file (Excel, CSV or JSON). StayCharted checks it before training:
- the same item with two different answers, which is usually worth settling first;
- categories with too few examples;
- category names that look like the same category;
- personal information in the text, such as phone numbers, email addresses, card numbers and US Social Security numbers, which it can mask before training. Names and street addresses need separate review; masking does not catch every sensitive detail.
3. Train, and test it on newer claims
Train an AI Classifier Model, which reads meaning and many languages. Under How it's tested, choose Test on the newest rows with the date column, so the model learns from earlier claims and is tested on newer rows from each category. This does not select a fixed calendar month. If many rows belong to the same claim, choose Keep rows that share a value together with the claim number instead.
The report shows how often it matched your team, category by category, and how many items would go to a person at each confidence cutoff.
For photos, add Vision. On Business, a Dedicated AI Image Model can also read the adjuster's note or the customer's message with each photo.
4. Connect it to the intake
- As items arrive: a Zapier, Make or Power Automate step, or your own system through the API, sends each new document, email or report to the model and writes the suggestion back, so it lands in the right queue.
- In batches: fill a spreadsheet of the day's items.
5. Review the unsure ones, then retrain
On Essentials and Business, answers below your cutoff appear in the Review Queue for a person. API access is also available on Essentials and Business. Their corrections join the next training run. Retrain monthly, and whenever your queues or codes change.
Built for a regulated business
- People decide. It suggests a category, a queue or a code. Whether a claim is covered, paid or investigated stays with your adjusters.
- A record of how each model was tested, on which data and with what result, for every training run, to support your AI governance documentation.
- Workspace activity recorded in the workspace's Activity.
- Private to your workspace. Your claims don't train anyone else's model. See Security and our DPA.
What it doesn't do
- It doesn't read PDFs or scans by itself. Send it the text your claims system or OCR step extracts.
- It doesn't change your claims system. Your integration writes the suggestion; your team acts on it.
- It doesn't decide coverage, reserves or payment, and it isn't a fraud score. Use it to route work, not to decide outcomes.
- It doesn't learn categories it hasn't seen. A new queue or code needs examples first.
Common questions
What is AI claims triage?
Sorting new claims, claim documents and claims emails into the right queue or category as they arrive, so adjusters start on the right work. A model trained on past claims makes the first pass, with a confidence, and people check the unsure ones.
Can AI classify insurance documents?
Yes. Trained on documents your team has already filed, a model sorts new ones by type, such as medical bills, police reports and estimates, from their extracted text.
Can it triage first notice of loss?
It can suggest a route, such as fast-track, standard or field adjuster, and a cause-of-loss code, from the first report and its photos, learned from how your team routed past claims. A person makes the decision. Text with photos requires a Dedicated AI Image Model on Business with Vision.
Does it work with Guidewire, Duck Creek or our own claims system?
It doesn't connect to any claims system directly. It works through Zapier, Make, Power Automate or its API, or with files you export, so any system that can send a record or an export can use it.
How does this fit the NAIC AI bulletin?
The NAIC model bulletin describes expectations for a written AI systems program and supporting documentation. StayCharted's training and validation reports and Activity records can support that work. These features do not by themselves establish regulatory compliance; your organization remains responsible for its governance and applicable requirements.
Is our claims data used to train anyone else's model?
No. Your models and data stay in your workspace.