Claims Agent Console
Five agents move every claim. One screen keeps the people who answer for them in charge.
Under NDA, so names and screens here are re-created.
-
5agents
Intake, coding, eligibility, submission and denial follow-up
-
1screen
Where a person sees all of it and steps in
The problem
Behind the product, a chain of five agents takes each claim from the visit to the payer and back. They work fast and around the clock. The billing team is still the one that answers for every claim.
The risk is an agent system that works and can't be read. When something goes wrong, nobody can see where, why, or whose move it is.
Constraints
- Every design decision had to hold up against a detailed technical design of the agent pipeline.
- Payer rules, prior authorisation windows and appeal deadlines are fixed.
- People stay accountable for every claim the agents touch.
Decisions
Three decisions shaped the console.
01 Show the pipeline, not a chat
The agents' work is laid out as stages a claim moves through, with counts at each stage. Anyone can see where work is piling up without asking a bot.
02 Rank by what's at stake
The priority queue orders claims by value and time left, so the most expensive, most urgent problem is always the first thing on screen.
03 Stop the loop and hand it over
When a claim keeps bouncing between the agents and the payer, the console raises it as a denial loop, shows what was tried and why, and puts the decision in front of a person, with a suggested fix they can accept or override.
Try it
A claim is stuck in a denial loop. Open it, read what the agents tried, and decide: approve the suggested fix or take it over yourself.
Re-created for this page. It shows the pattern, with sample data.
Beyond the happy path
The states I check on work like this, because nobody asks for them until they break.
- Agent stuckA stage stops moving and says why
- Denial loopRaised to a person with what was tried
- Low confidenceThe agent's certainty is shown
- Human overrideLogged under the person's name
- Window closingAuthorisations near their deadline rise to the top
Evidence
- Each component traces back to a part of the pipeline's technical design.
- The denial loop gets its own alert instead of hiding in a list.
- The copilot rail stays on screen, so getting help never means leaving the queue.
How I'd measure it
These results weren't tracked on this project. With access, these are the numbers I'd watch.
- Time to resolve a denied claim
- Share of claims that needed a person
- First-pass approval rate