5 → 2 min
Pre-auth claim submission time, with BenefSnap automating beneficiary search and PreAuthSnap filling the form
Designing Agentic AI at the Point of Friction Inside a Legacy Claims Platform
I design the agentic experience for five AI agents in pre-authorization. Together they form the Intelligent Pre-Auth Flow, the step before IHX MIRA.
Pre-auth time
Search failures
−40%
Beneficiary search failures, down from a 63% failure rate
Time lost
−20%
Extra processing time when beneficiary search fails, down from +37%
The Intelligent Pre-Auth Flow
Five agents inside one pre-auth flow, and two after submission.
PreAuthSnap.AIPilot · 3 hospitalsBenefSnap.AILive
A docket or pre-auth form runs PreAuthSnap: it finds the patient and fills the form. An insurance card alone runs BenefSnap for the beneficiary search. Submission drops from 5 to 2 min.
CodeRight.AIPilot · 3 hospitalsDocuSense.AIPilot · 3 hospitals
Get the claim right first time: fewer payer queries, faster TAT.
SDMS.AIIn development
Upload once, submit faster: no outside PDF tools, patient data stays inside IHX.
After claim submission
QueryLens.AISettleSense.AI
Business outcomes
Solving pre-auth for hospitals was integral to retaining enterprise hospitals and renewing payer contracts.
Enterprise hospital retention
PreAuthSnap fills the form, the biggest share of pre-auth time, and BenefSnap automates the beneficiary search. Together they cut submission from 5 to 2 min; CodeRight and DocuSense target fewer payer queries and faster TAT.
Payer contract renewal
Fewer queries and faster pre-auths support payer contract renewal.
New revenue
SettleSense targets revenue per claim from INR 120 to 150, sold as a paid add-on to Auto-Recon.
Data protection
SDMS keeps patient documents inside IHX instead of outside PDF sites, a selling point under India's data protection law (DPDP).
The long-term goal
IHX MIRA: a claim submitted as a conversation.
A concept I pitched. The hospital desk submits a claim by chatting. MIRA asks for the patient, the diagnosis and missing documents, and brings payer questions into the same chat. It needs kiosks and hardware in hospitals, so it is long-term. Every agent above is a part it needs.
IHX MIRA prototype: a claim taken from intimation to submission in one conversation.
TL;DR
The whole story, in five cards.
Problem
Staff retype what is already on their documents: member card, pre-auth form, docket.
Approach
An agent at each point of friction, using data already in IHX.
Trust
Show where every value came from, how sure the agent is, and what still needs checking. Staff can edit everything.
Status
BenefSnap is live. PreAuthSnap, CodeRight and DocuSense are in pilot at 3 hospitals; SDMS.AI is in development.
Where it leads
The five agents become the Intelligent Pre-Auth Flow, the predecessor to IHX MIRA.
01 · The Intelligent Pre-Auth Flow
One pre-auth, five points where staff used to get stuck.
Each step where staff used to type, search or use outside tools now has an agent. The agent reads and fills; staff check and edit.
- Step 1Find the patientBenefSnap.AI with a card, or PreAuthSnap.AI with a docket or pre-auth formLive
- Step 2Fill the pre-auth formPreAuthSnap.AIPilot
- Step 3Code the diagnosisCodeRight.AIPilot
- Step 4Check documentsDocuSense.AIPilot
- Step 5Package and submitSDMS.AIIn development
After submission
QueryLens.AIpayer queries
SettleSense.AIsettlement shortfalls
User document input: the system decides the path
User
Uploads a document
Docket or pre-auth form
System runsPreAuthSnap.AI Pilot
Automated beneficiary search
Pre-auth form filled
Insurance card
System runsBenefSnap.AI Live
Automated beneficiary search
User selects the policy
Uploads pre-auth form
System runsPreAuthSnap.AI Pilot
Or
Fills the form manually
User reviews
Submits
The actual capability, not the agents
Designing for AI starts with how people think and what makes them trust a result. So every agent shows where a value came from, how sure it is, and what is missing, and lets staff edit everything.
Trust-Sensitive Patterns
02 · BenefSnap.AI · Beneficiary Search
LiveIndividual agentic workflow
A failed beneficiary search took 37% longer to resolve than a successful one.
BenefSnap reads a member or policy card and runs the beneficiary search automatically. It runs when staff upload only a card. A docket or pre-auth form runs PreAuthSnap instead.
Module walkthrough
+37%
Slower to resolve a failed beneficiary search (2h21m) than a successful one (1h40m median)
1
Failed searches route to a manual queue
Someone matches the beneficiary ID by hand, or the hospital submits by email instead.
2
Hospitals pushed back
They did not want to do beneficiary search by hand at all.
The decision
The data is already on the card the hospital uploads, so upload moved earlier in the flow and extraction replaced retyping.
Building trust
01
Works with what the hospital has
A docket or pre-auth form runs PreAuthSnap. A card alone runs BenefSnap; staff then select the policy and either upload the pre-auth form or fill it manually.
02
Inside the existing step
Upload sits where staff already pick the payer, not on a new screen. Nothing new to learn.
03
The document, shown with the data
The member card sits next to the extracted details: name, policy number, insurer and TPA. Staff can edit any field, and the agent learns from each correction.
Impact
Adoption
270hospitals live
Rolled out across the platform as a premium AI feature.
Automation
22% automated search adoption
Share of beneficiary searches now run automatically.
Success rate
60%
Beneficiary searches BenefSnap resolves successfully.
Completion rate
76%
Users who complete the BenefSnap flow once they start it.
03 · PreAuthSnap.AI · Pre-Auth Form
Pilot · 3 hospitalsIntelligent Pre-Auth Flow
Staff retyped pre-auth details that were already on the form they uploaded.
Part of the new Intelligent Pre-Auth Flow, the flow that cuts submission from 5 to 2 min. The parent agent: a docket or pre-auth form runs PreAuthSnap. It runs the beneficiary search, staff pick the card, then it fills the pre-auth form for staff to review.
The problem
Staff typed pre-auth details by hand, though the signed form they uploaded already had them. Double work, and data-entry errors.
Why it was hard
Uploads vary: a docket of up to about 100 pages, a single pre-auth form, or just an insurance card. The system has to identify the input. In a docket, it also has to find the pre-auth form among the other pages. No existing service could recognise a pre-auth form.
The decision
Staff upload whatever they have, and the system responds to it. A docket or pre-auth form runs PreAuthSnap: beneficiary search, card selection, then the Member, Treatment and Bill details are filled. The upload carries through to Document Upload, so nothing is uploaded twice.
Building trust
01
Every value shows its source
Each field shows where it came from: KYC, insurance card, pre-auth form or the hospital's HIS.
02
A priority order for sources
Payers read the signed pre-auth form, so when sources disagree, its values take priority.
03
Confidence from two models
OCR and Gemini both extract each field. Staff see a confidence score, choose the value, and can edit any field.
04
Up front about gaps
The form shows how many fields are filled and how many need checking. Skipped or missed steps get an orange alert.
Impact
Submission time
5→2min per pre-auth
PreAuthSnap fills the form, the biggest share of pre-auth time.
04 · CodeRight.AI · Diagnosis & ICD Coding
Pilot · 3 hospitalsIndividual agentic workflow
61% of diagnosis entries carried no ICD-10 code, and the person typing them is not a clinician.
CodeRight reads the diagnosis from the pre-auth form, or from plain words staff type, and suggests ranked ICD-10 codes and the probable treatment. Staff verify.
Module walkthrough
61%
Of 1.8–2 lakh monthly diagnosis entries carry no ICD-10 code
1
The user is not a clinician
Desk staff describe conditions in plain words: "cancer of bone" for sarcoma, "stone in kidney". Doctors use shorthand, like "TM perf".
2
A dropdown assumes what is missing
A searchable ICD dropdown only works if you already know the clinical term.
49% usable free text, nothing selected
12% garbage input
21% wrong code picked outright
The decision
No ICD dropdown. Staff can type in plain words, and the selected department helps narrow it down. In the Intelligent Pre-Auth Flow, CodeRight also reads the pre-auth form, so staff verify instead of typing.
Building trust
01
Multi-select, not auto-fill
The system suggests up to five ranked codes, and staff pick one. A wrong auto-picked code costs more than a slower right one.
02
Case volume as a trust signal
Each suggestion shows how many cases this hospital had with that code in the last three months, and how likely it is to be the right one. The probable treatment is suggested too.
03
Says when to check
The screen says suggestions may not match every insurer's rules and should be verified.
Targets
Coverage
65–70% ICD coverage
Diagnosis entries that carry a valid ICD-10 code, up from 39%.
Response speed
<2sfor top-5 suggestions
The most relevant ICD code appears in the top 3 results.
05 · DocuSense.AI · Document Intelligence
Pilot · 3 hospitalsIndividual agentic workflow
Hospitals found out which documents a payer needed only when the query came, often after discharge.
DocuSense builds a document checklist for each claim from its diagnosis and procedure, and shows how many mandatory documents are attached.
The problem
Payers asked for documents days later, often after discharge, when they were harder to get.
The decision
The checklist is built from the claim's ICD code and procedure, using the same mapping as CodeRight. No new input from the hospital.
Building trust
01
Document completion shown before submission
A bar shows how many mandatory documents are attached, while staff can still add the rest.
02
Documents grouped by reason
Stage-specific and ICD-specific documents are listed separately, so staff see why each one is needed.
03
Same disclaimer as CodeRight
The same "based on historical patterns, please verify" line, so AI suggestions read the same way across IHX.
Targets
QC load addressed
70–100/day, per flagship hospital chain
Manual QC cases tied to missing or mismatched documentation.
Classification accuracy
~80%
Document classification accuracy for the dynamic checklist.
06 · SDMS.AI · Smart Document Management
In developmentIndividual agentic workflow
Staff compressed patient records on outside PDF sites to meet payer file limits.
Staff upload documents once. IHX compresses, splits or merges them to fit each payer and channel.
The problem
Discharge dockets mix summaries and bills, and scan quality varies. Each payer and channel has its own file size and attachment limits. An API may accept a large file; the fallback email may not. Staff used outside tools like iLovePDF to fit them.
The decision
IHX packages documents per payer and channel. If a submission falls back from API to email, IHX repacks it for the smaller limit without asking staff.
Why it matters
01
Patient data stays inside IHX
No patient record goes to an outside compression site. This matters under India's data protection law (DPDP).
02
Upload once
Staff never build a payer-specific bundle by hand.
03
No silent failures
If the standard path fails, the system handles the fallback.
07 · After Claim Submission
Two agents for after the claim is submitted.
Outside the pre-auth flow. Both read what the payer sent and turn it into the next step.
QueryLens.AI
Payer queries sat 12 to 48 hours. No set categories, staff added teammates to each query chat by hand, and replies were spread across four tools. QueryLens categorizes each query, suggests an owner, and keeps the reply in IHX.
Targets
40–60%faster median query resolution
77→95%+queries the AI can categorize on its own
SettleSense.AI
Auto-Recon showed a short payment but not the reason. Recovery teams read two letters per claim to find it. SettleSense gives the reason with evidence, marks it contestable or not, and sends contestable cases to the dispute workflow.
Target
INR 120→150revenue per claim, as a paid add-on to Auto-Recon
Each agent fixes one point of friction. Together, they are the road to IHX MIRA.
IHX · AI Agents at Points of Human Friction, ongoing
Need a designer to bring clarity to your workflows?Talk to me.
Open to product design roles and contract work.
Download resume- {
- "name": "Shabnam Raghavan",
- "title": "Product Designer",
- "email": "shabnam.rs.work@gmail.com",
- "phone": "+91 70226 65338",
- "link": "linkedin.com/in/shabnamraghavan"
- }