Automated insurance claims, three workflows to build
Automated insurance claims, shown through three workflows: first notice of loss, document review and subrogation. What to automate and what stays human.
Founder, Viithiisys

Which automated insurance claims projects pay back first?
The ones where claims arrive as messy documents and the decision after them is simple. First notice of loss, supporting-document review and subrogation follow-up fit that description.
This post sits in AI workflow automation, one of the three subjects we write about. Viithiisys has shipped software since 2007, over 500 projects, but nobody has handed us a confirmed claims delivery record for this article. So everything below is a design, labelled [Reference build], and none of it describes a client engagement.
Each workflow follows the same four questions. Who touches the claim today, where does it jam, what does the jam cost, and which steps should software take versus which stay with a person?

Why a scripted bot stalls on claims
A scripted bot follows fixed screen steps, and claims inputs never stay fixed. A repair estimate from one garage looks nothing like the next, and a bot that expects a field in one place fails the first time it moves.
The usual rescue is more rules. Within a year the rule set is larger than the process it replaced, and one person owns it.
Generic claims SaaS has a different problem. It assumes you will adopt its intake model, which rarely matches the mix of channels you actually receive: broker email, a call-centre note, a WhatsApp photo, a portal form.
The more workable build keeps your claims platform as the system of record and puts an extraction and routing layer in front of it. Platform vendors such as Guidewire document their integration APIs for this reason.
First notice of loss arrives as five different formats
Today a call-centre agent or shared-inbox clerk reads the report, retypes it into the claims platform, and attaches photos by hand. In a Guidewire ClaimCenter or Duck Creek Claims shop, that retyping is where policy numbers get mistyped.
It breaks on volume spikes. After a storm, the inbox backs up, and a claim sits unassigned while the customer calls to ask where it is.
[Illustrative example] Picture a hypothetical insurer taking 4,000 reports a month, with a clerk spending ten minutes retyping each. That is roughly 670 clerk hours a month spent on data entry. The numbers are invented to show the arithmetic, so replace them with your own.
Automate the reading: pull the policy number, loss date and loss type from the email or form, match the policy, and create a draft claim. Leave to a person any report where the policy match is uncertain or the loss description mentions injury. The draft goes to a clerk, not straight to settlement.
Supporting documents are where adjusters lose hours
An adjuster opens a claim and finds a PDF estimate, a police report, three photos and a medical invoice. They read each one, compare totals with the reserve and check dates against the policy period.
It breaks when a document is missing or wrong, and nobody notices until day nine. Then the claimant is chased, the clock restarts and the customer complains.
Automate the first pass. Document extraction services such as Azure AI Document Intelligence and Amazon Textract pull fields and tables from scanned files. An LLM can then compare extracted figures with the claim record and flag mismatches.
Keep three things human: the coverage decision, any estimate that disagrees with the photos, and any claim tripping a fraud indicator. Our document AI work covers this extraction layer.

Subrogation recovery dies in follow-up
After a claim is paid, someone must work out whether another insurer owes money, send a demand and keep chasing. In most shops that tracking lives in a spreadsheet beside the claims platform, with reminders in one person's calendar.
It breaks quietly. Demand letters go out late, replies sit in an inbox, and statutory deadlines pass without anyone noticing. The cost shows up as recoveries that were never made, which is hard to count because nobody records what did not happen.
Automate the clock and the correspondence. Detect recoverable claims from the loss description, assemble the demand pack from the claim file, send it, and log replies against the right claim. Ticket the exceptions in ServiceNow or Salesforce Financial Services Cloud, whichever you run.
Leave negotiation and arbitration decisions to a recovery specialist. Software should hand them a complete file and a clear deadline.
Where AI helps and where plain rules are enough
Use AI for reading and summarising unstructured content. Use plain rules for decisions you can write down and defend, such as payment thresholds and policy-period checks.
| Step | Plain rules | AI model | Human |
|---|---|---|---|
| Match policy number | Exact lookup | Fuzzy match on typos | Unresolved matches |
| Read estimate or invoice | Fails on layout change | Extracts fields | Totals that disagree |
| Check date inside policy period | Yes | Not needed | Disputed dates |
| Summarise loss narrative | No | Yes | Reads the summary |
| Coverage and liability | Threshold gates | Suggests only | Decides |
| Fraud referral | Indicator rules | Flags patterns | Decides |
Let software read the claim file, but let a named adjuster own every coverage decision.
Regulators are watching this boundary. The NAIC's work on artificial intelligence includes a model bulletin on insurers' use of AI systems, so keep a log of what the model suggested and who decided.
Three costs that get left out of the estimate
Teams budget for the model and forget the plumbing. Integration with the claims platform, policy admin system and document store usually takes more effort than the extraction itself.
The second cost is exceptions. Every unreadable scan and unusual claim needs a queue, an owner and a rule for how long it may wait. Skip that and the automation simply moves the backlog.
The third is evaluation. Before go-live, someone has to build a test set of real past claims with known outcomes, and re-run it whenever a model or prompt changes. Structured output from an LLM, as described in Anthropic's tool use documentation, makes results checkable, but the checking still needs people.
Data protection review also takes calendar time, particularly with claimants in the US, UK and Canada, where our teams work.
Price depends on three things we scope per workflow
We do not publish a price for this work. It is scoped per project after a discovery call, and a number written here would be invented.
The drivers are claim volume, the count of systems the layer must read and write, and how clean your historical data is. A single workflow such as first notice of loss can be a fixed-scope build, and our Moonship MVP approach ships a working version in 30 days against a fixed scope.
For anything wider, an AI workflow automation engagement runs on a scoped basis. If the hard part is connecting to Guidewire, Duck Creek or a legacy policy system, AI integration is the page that describes it.
Treat any vendor quote that skips exception handling and evaluation as incomplete.
Send us the claims queue that never empties
Every claims operation has one queue that grows faster than it clears. It might be unassigned first notices after a storm, files waiting on a missing invoice, or recoveries nobody has chased since spring.
Start by picking that one. Write down who touches it, which system each person is in, and what a stuck claim costs per week.
Our broken workflow assessment takes that single queue and tells you which steps can be automated, which stay with an adjuster and what the build would involve. Send us the workflow that gets stuck, and we will tell you honestly whether automation is the right fix.
For a conversation first, the contact form reaches our engineers in Mohali and Markham.
Sources
FAQ
- What is automated insurance claims processing?
- It means software handles the repeatable steps of a claim, such as capturing the loss details, reading attached documents, checking them against the policy and chasing missing items. A human adjuster still owns coverage decisions, disputed liability and anything above an agreed payment threshold.
- Which claims should be automated first?
- Start with high-volume, low-severity work where the inputs are messy but the decision is well defined: first notice of loss intake, document checks and recovery follow-up. Large bodily injury and contested liability claims should stay with experienced people.
- Do I need to replace Guidewire or Duck Creek to automate claims?
- Usually not. The practical route is an automation layer that reads and writes to your existing claims platform through its APIs, so adjusters keep working in the system they already use while intake and document handling move out of their queue.
- How much does insurance claim automation cost?
- It depends on claim volume, the number of systems involved and how clean your data is. We scope each build after a discovery call, so we do not publish a figure. Our assessment is the quickest way to get a scoped answer for one workflow.