Innovation Lab · AI Business Use Case

Drop & Done

From a pile of receipts to "ready to submit" in under 2 minutes

An AI-native reimbursement assistant in the chat apps employees already use. It reads every bill, groups the spend, checks policy, catches duplicates and fake receipts, and asks only the questions it can't answer itself.

Presented by Supriya Manjrekar01

The problem

Reimbursement is low-value work that everyone pays for

Employees collect bills, type in details, pick categories, check what's allowed and build a claim. Finance then re-checks all of it by hand.

20 min

to prepare a single expense report

GBTA Foundation / HRS

19%

of reports have errors or missing info, so they get sent back

GBTA Foundation / HRS

$58

fully loaded cost to process one report, plus $52 to correct one

GBTA Foundation / HRS

14%

of fraudulent expense documents were AI-generated fakes (Sep 2025, up from 0%)

AppZen via Skift, Nov 2025

India makes it harder: handwritten auto/taxi slips, faded thermal prints, UPI screenshots in place of invoices, bills in regional languages, and lost GST input credit when hotels don't bill to the company GSTIN.

Drop & Done02

Who feels the pain

Three users, one broken loop

👤 Employee

Hunts for bills in email, WhatsApp and their wallet
Types every field by hand into a form
Has to guess the policy rules
Finds out about a rejection days later

Feels: "chore, unfair delays"

✅ Approver

Gets 20 claim lines with no context
Either rubber-stamps them or digs through each one
Has no signal for what's risky

Feels: "I can't really check this"

🔎 Finance / Audit

Checks 100% of claims by hand
Chases missing GSTINs and invoices
Catches fraud only by sampling
Has no data on policy leakage

Feels: "drowning in rework"

Today: Collect → Type → Categorise → Guess policy → Submit → Rejected → Fix → Resubmit → Approve → Audit → Pay 10–15 days

Drop & Done03

Market research

A large, fast-moving market that still feels form-first

$9.1B → $21.4B

Global expense management, 2026 → 2035 (~10% CAGR)

Global Market Insights

43% CAGR

AI-powered receipt automation segment, 2024–29

Technavio

Consolidation: Capital One bought Brex ($5.15B), Ramp passed $1B ARR, Sage bought Fyle. SAP Concur holds about 19% share.

Where incumbents leave room

Gap	Concur / Zoho / Happay	Drop & Done
Interface	App + form, one receipt at a time	Chat, whole pile at once
Policy check	After submission	Before submission, explained
Missing info	Mandatory fields	≤3 targeted questions
Indian bills	Tuned for clean receipts	Handwritten, UPI, Hindi
Fake receipts	Emerging add-on	Built in, every claim
GST credit	Capture only	Prevent leakage at source

Drop & Done04

The experience

Drop the pile. Answer two taps. Submit.

📎 12 files: WhatsApp Image…jpg, Hotel_Saffron_Invoice.png, Screenshot_UPI.png, auto_receipt.jpg …

Read 12 bills ✅ I found 2 trips and ₹29,247 in spend.
🧩 Delhi trip · 21–23 Sep · 9 bills
🧩 Pune day trip · 29 Sep · 3 bills

🚩 IMG_1004.jpg is the same cab ride as QuickRide_21Sep.jpg, just a different photo, so I left it out.
⚠️ Hotel invoice has no company GSTIN, so we lose ₹1,392 of GST credit. Ask the hotel for a revised invoice?
Yes, draft emailSkip

Question 1 of 3: dinner on 22 Sep for ₹4,200 is over the solo meal limit. Was it a client meal?
Yes, client mealJust me, claim ₹1,500

Rahul Mehta (Zenith Pharma), Priya Nair (Zenith Pharma)

Claim ready ✅ 11 lines · ₹27,847 · 1 flagged for finance
Pile to ready in 0m 35s · 3 questions asked
Submit claimReview lines

Design principles

Zero forms. The chat is the whole interface.
Batch first. Handle the whole pile, not one receipt at a time.
Ask by exception. Only gaps that block submission; tap to answer.
Explain every flag. Say what's wrong, why, and how to fix it.
Always show status. ✅ complete · ⚠️ needs info · 🚩 questionable

North-star UX metric

Time-to-ready-to-submit
Target under 2 min for a 10-receipt trip, against about 20 min today.

Drop & Done05

AI intervention

Where AI does the work, and where rules keep it honest

1

Read every documentClaude vision turns photos, PDFs and screenshots into strict JSON: vendor, date, amount, tax, GSTIN, city and payment mode, with a confidence score per field.

2

Validate deterministicallyChecks that items + tax = total, the GSTIN checksum, that dates make sense, and converts foreign currency. These are rules, not guesses.

3

OrganiseGroups expenses into trips or events by date window, city and travel bookings.

4

Check policyRules-as-code by grade and city tier. The LLM handles fuzzy cases and writes the explanation.

5

Protect integrityDuplicates at 3 levels (image hash, same transaction, across employees), signals of AI-generated receipts, and anomalies against peers.

6

Recover GSTFlags missing company GSTIN while it can still be fixed, and tags eligible input credit.

7

Ask only what's missingLow-confidence or missing fields become at most 3 batched, tap-to-answer questions.

8

Prepare & routeBuilds the submission-ready claim, gives approvers a one-line summary, and sends finance a risk-ranked audit queue.

Human in the loop by design: AI prepares and flags, people approve payouts. Every AI decision is logged with the image region and policy clause it relied on.

Drop & Done06

Solution architecture

Drop & Done07

The backend that makes it reliable

A claim is a state machine, not a form

Each extracted field carries evidence

"amount":  { "value": 4200,
             "confidence": 0.97,
             "source": "bottom-right total" },
"gstin":   { "value": null,
             "confidence": 0.0,
             "status": "MISSING" },
"category":{ "value": "Meals",
             "confidence": 0.88 }

Confidence ≥ 0.85 is auto-filled. Below that, or missing, it becomes a question. Anything contradictory is flagged.

Line states

COMPLETE NEEDS_INFO QUESTIONABLE EXCLUDED

Claim lifecycle

DRAFT → READY → SUBMITTED → APPROVED → PAID

A claim is READY only when every line is COMPLETE or EXCLUDED.

Category checklists

🏨 Hotel: invoice, stay dates, GSTIN, city tier
🍽 Meal > cap: attendees + purpose
🚕 Cab: from/to or trip link
✈️ Flight: boarding pass or e-ticket

Drop & Done08

Business impact

Measurable value across time, money and risk

Illustrative model: 10,000 claim-filing employees × 2 claims/month = 240,000 claims/year. Assumptions are listed so they can be replaced with real figures.

Lever	Today	With Drop & Done	Annual value
Employee prep time	20 min	3 min	68,000 hrs
Rework rate	19%	<5%	10,000 hrs
Finance check / claim	10 min, all claims	3 min, risk-based	28,000 hrs
Reimbursement cycle	10–15 days	2–3 days	Employee experience
GST input credit	Partly captured	Leakage prevented	₹ direct recovery
Duplicates / fakes	Sampled	100% screened	Loss avoided

PRODUCTIVITY RELEASED

~1,06,000 hrs

per year, about ₹6.4 Cr at an assumed ₹600/hr loaded cost

Plus hard savings: GST credit recovered on hotel and travel invoices, and duplicate or fake claims stopped before payout. At ~1% of claims flagged as suspect, even a fraction prevented covers the run cost.

Benchmarks: GBTA (time, error rate) · AppZen / SAP Concur (fake receipt prevalence)

Drop & Done09

How we'll prove it

KPIs measured from day one

Experience

Time-to-ready-to-submit (median)
Questions asked per claim (target ≤ 2)
Employee CSAT / NPS
Days from spend to payout

AI quality

Field extraction accuracy (labelled eval set)
% fields auto-filled without edits
Duplicate precision and recall
Confidence calibration

Business

First-pass approval rate
Finance hours per 1,000 claims
GST credit recovered (₹)
Suspect ₹ blocked before payout

Pilot success criteria (8 weeks, one business unit)

< 3 min

median time-to-ready

≥ 95%

field accuracy on key fields

≥ 90%

first-pass approval

≥ 60%

less finance review effort

Drop & Done10

Roadmap & risks

Start narrow, prove value, then scale

Phase 1 · Prototype

Now

Web chat with drag-and-drop pile
Claude extraction + confidence
Grouping, policy, duplicates
Mock reimbursement API

Phase 2 · Pilot

+8 weeks

Teams / WhatsApp channel
HRMS + SSO integration
Approver copilot
GST recovery + fake-receipt signals

Phase 3 · Scale

+6 months

SAP / finance system integration
Card + UPI feed, zero-touch capture
Cross-employee duplicate detection
Spend analytics for finance

Risk	Mitigation
Extraction errors on poor images	Confidence thresholds; ask instead of guess; deterministic arithmetic checks
Data privacy (PII, bank data)	PII masking before LLM, India data residency, encryption, RBAC, audit log
Over-trusting AI decisions	AI never approves payouts; every flag cites evidence; humans override, and overrides feed the eval set

Drop & Done11

The ask

Let employees drop their bills,
and let AI do the paperwork.

1 · Live demo

12 messy bills to a ready claim, timed live

2 · Pilot BU

One business unit, 8 weeks, measured against KPIs

3 · Sponsors

Finance policy owner + IT security partner

Drop & Done · Thank you12

Drop & Done

Reimbursement is low-value work that everyone pays for

Three users, one broken loop

👤 Employee

✅ Approver

🔎 Finance / Audit

A large, fast-moving market that still feels form-first

Where incumbents leave room

Drop the pile. Answer two taps. Submit.

Design principles

North-star UX metric

Where AI does the work, and where rules keep it honest

A claim is a state machine, not a form

Each extracted field carries evidence

Line states

Claim lifecycle

Category checklists

Measurable value across time, money and risk

KPIs measured from day one

Experience

AI quality

Business

Pilot success criteria (8 weeks, one business unit)

Start narrow, prove value, then scale

Phase 1 · Prototype

Phase 2 · Pilot

Phase 3 · Scale

Let employees drop their bills,and let AI do the paperwork.

1 · Live demo

2 · Pilot BU

3 · Sponsors

Let employees drop their bills,
and let AI do the paperwork.