Simplifying receipt filing and approvals for corporate travel
Summary
I led and designed the expense management system for Miraee, Mondee's AI-native corporate travel application. It lets corporate travellers submit and file their receipts quickly during travel, gives the finance team clear visibility into spend, and keeps approval times with managers short.
Product
Team
Impact
Why is expense management a big problem in travel?
Corporate expense management is notoriously slow. Travellers lose hours compiling receipts; finance teams lose hours cross-checking them against policy; and both sides trade emails over minor justifications. Neither side is unreasonable: the traveller wants to go home, and finance wants numbers it can defend.
Miraee was already in the market, in its early phases, with two customers who were also our investors: TCW and JP Morgan. Their pain points were where we started, and the product had to satisfy the rigidity of one side and the speed the other expects.
The three stakeholders, and the old process they were stuck in.
Traveller
1Compiles receipts
About 90 minutes a report, in the spreadsheet era
Traveller
2Submits the report
Sent off with the receipts attached
Manager
3Approves
Often without a close look
Finance
4Checks against policy
Line by line, by hand
Finance
5Rejects for something minor
Sent back over a small mismatch
Traveller
6Re-justifies over email
Then back to step 1. The loop this project removes
This is the current system, one prototype for each of the three stakeholders.
Home
A compliance score, a warning about overdue trips and a saved draft, all on one screen. Upload is the only way in.
The Gantt follows one expense from capture to reimbursement. Filing the receipts takes about twelve minutes, yet it waits about two weeks, the longest wait on the chart. The durations are averaged from usage log timings, so they are approximate but based on real user data.
The goal
Find out why travellers hold off on filing for so long, then make filing, review and finance checks faster for everyone involved.
Where we start
| Metric | Spreadsheet era | Our v1 (the baseline) |
|---|---|---|
| Time to create a report | ~90 min | ~12 min |
| Time to review | 19% of reports have errors | ~15–20 min per report, all manual |
| Turnaround, submit to approved | — | ~4–6 days, +1–2 days per rejection |
| Turnaround, trip end to approved | — | ~3 weeks |
| Auto-approval | 0% | 0% |
We analysed upload logs and interviewed a cohort of business travellers. Two capture behaviours came out of it.
Capture as you go
Batch uploaders
Whichever way they uploaded, both types hit the same wall at the sorting and verification stage.
In the interviews, a clear pattern emerged: travellers expected expense filing to be a quick, five-minute admin task. They were not avoiding it because the backend was slow, since our OCR and matching were fast. They were avoiding it because the v1 interface made the task look exhausting.
The structural flaw
Where the old flow hurt, and why
Capture
Sort & verify
Send
Manager review
Finance check
Reimbursed
Keeps receipts
Checks every receipt
Clears every issue
Waits
Waits
Gets paid
Waits, then reviews
Waits, then checks policy
Reimburses
Touch point
Traveller feels
Manager feels
Finance feels
Pain
Fear of losing one
12 min of re-checking
Locked, and unclear why
A queue, then fixes by email
Re-checks what was verified
Weeks after the trip
Needs
Capture is the whole job
Trust what the OCR read
Show where they stand
Review only the exceptions
Apply policy automatically
Say when it will land
We built
Capture with a saved draft
Checks run for you, in view
One summary, one next step
Only exceptions reach them
Policy applied automatically
Approved in-app, same day in policy
Lowest point
Sorting and verifying. Filing feels like petty bureaucracy, so they put it off for weeks. It is the longest wait on the Gantt.
Design focus
Make capture the whole job, and let them trust what the OCR already read.
Lowest point
Reviewing. A queue of everything arrives at once, and the fixes go back and forth by email.
Design focus
Show only the exceptions, and keep justifications in the app.
Lowest point
Checking policy. They re-check what was already verified, with the month-end close behind them.
Design focus
Apply policy automatically, and keep the paper trail unshakeable.
Two additional stakeholders, travel coordinators and HR, also get involved from time to time. Coordinators chase receipts over email, and HR wants policy applied consistently without feeling like a police state. They shaped the policy engine and the compliance score.
With the data in hand, we went back to the old screens and read them through three lenses: whether things can be found, whether they can be scanned, and whether the system earns trust. Under each lens, we looked at the behaviour the screens triggered.
The overdue banner helps, but has no deep links.
On: Home
The drop zone promises sorting, but nothing shows what needs attention.
On: Home
Policy is invisible until it fires.
On: Review: policy warnings
13
2What the market does, and what it leaves open
After the spend
Automates an expense once it has been incurred.
Before the spend
Prevents non-compliant spending before it occurs.
Zero-touch filing. The expense is created and filed properly, with no manual input from the user.
Proactive gets close, but not there. A card swipe creates the expense, yet it carries little more than a merchant name. The receipt and the context still have to come from the user, and analysis suffers without them.
1
Built for: Modern growth companies, mid-market businesses and anyone who wants a unified financial stack.
The corporate card is the hub of the expense lifecycle. Each swipe creates the expense record automatically.
What works
What doesn’t
How it works
Company-issued card is used
Policy checked at swipe
Expense created automatically
Receipt requested if missing
Traveller replies with photo
2
Built for: Organisations whose people pay with their own or existing corporate cards.
Employees keep using their own cards. An aggregator like Astrada or Plaid streams the transactions, and the platform works from that feed.
What works
What doesn’t
How it works
Existing card is used
Aggregator pulls the transactions
Traveller uploads receipt
Receipt matched to transaction
Progress shown live
3
Built for: Small businesses, startups and SMBs that prioritise ease of use and rapid set-up.
Makes the first step of submitting an expense as effortless as possible. Expensify is the clearest example.
What works
What doesn’t
How it works
Traveller captures receipt
OCR reads it
Batched onto report
Traveller reviews
4
Built for: Large global enterprises with complex compliance and integration needs.
Typified by SAP Concur: control and global compliance ahead of usability, with deep ERP integration.
What works
What doesn’t
How it works
Add each entry
Attach documents
Fill in report details
Approval chain
Posted to the ERP
Understanding these four philosophies, the best path for an aggregator-based system is to borrow from the three modern approaches: the proactive mindset of card-first platforms, the frictionless capture of Expensify, and the elegant, exception-based workflows pioneered by Navan and Ramp. The enterprise model stays as the reminder of the baseline of functionality and compliance we still have to meet.
The gap
Benchmarks over receipt handling and policy enforcement mechanisms.
| Capability | Miraee | Expensify | Navan | Ramp / Brex |
|---|---|---|---|---|
| Receipt submission channels | Partly in place Camera, gallery and document upload. Scoped out for now: Email and SMS forwarding. Low return in corporate travel, next to capturing in the app. | Mobile app, drag-and-drop, email and SMS text. | Autofill from receipt, personal card import, manual entry and email. | SMS, email, web portal, mobile app, Slack and a Chrome extension. |
| Batch processing | Partly in place Several receipts can be uploaded at once, in a tab. Exploring: An AI chat layer, like Navan’s chat interface, for edge cases. | Yes. Drag in several files, or scan several receipts in one go. | Yes. Upload several receipts at once into a chat interface. | Yes. Group related expenses for batch approval. |
| Line-item policy enforcement | In place Each line item is checked against the policies that apply to it. | Yes. Flags prohibited items, such as alcohol, during the scan. | Yes. Auto-Itemization applies policy to each line item, at over 90% accuracy. | Yes. AI reviews transactions and can flag specific line items. |
| Lost receipt handling | In place A justification for the lost receipt, plus any other proof they have. | Manual entry with context, or a credit card statement attached. | A formal Missing Receipt Affidavit, with a digital signature. | A formal Missing Receipt Affidavit, with a digital signature. |
| Multi-currency support | In place Yes. We are a global company. | Yes. Global reimbursement in local currencies. | Yes. Local-currency spending in over 50 countries. | Yes. Local-currency spending globally. |
| Justification methods | In place Text-based justifications. | Text-based notes. | Text-based notes, plus video and audio explanations. | Text-based notes, plus contextual messages through Slack or SMS. |
Miraee
Partly in place
Camera, gallery and document upload.
Scoped out for now: Email and SMS forwarding. Low return in corporate travel, next to capturing in the app.
Miraee
Partly in place
Several receipts can be uploaded at once, in a tab.
Exploring: An AI chat layer, like Navan’s chat interface, for edge cases.
Miraee
In place
Each line item is checked against the policies that apply to it.
Miraee
In place
A justification for the lost receipt, plus any other proof they have.
Miraee
In place
Yes. We are a global company.
Miraee
In place
Text-based justifications.
The approval model decides how much work reaches a reviewer. We compared three, then looked at what the reviewer needs to see to trust the answer.
The traditional model, and Expensify’s full-review option
Every report lands in the reviewer’s queue and is inspected by hand.
16 to open
Navan, Ramp and Brex
Compliant spend is approved automatically. The reviewer gets a queue of flags, with the AI’s rationale and the policy section it broke.
2 to open
Exception-only, with the routine items out of sight
14 verified and matched, 2 that need attention. The 14 leave the reviewer’s view, the 2 can be grouped when they share context, and the justification stays in the app.
2 to open
Progressive disclosure
In the productInstead of a spinner, each check reports as it finishes, so waiting becomes proof that the system is working.
Confidence, in words
ProposedA raw 87% looks more precise than it is. Plain labels tell the user what to do next.
Vendor name read from the receipt
Accept, edit or dismiss, in one tap.
Policy flag
Send back, ask for a justification, or approve.
Trip assignment
Confirm, or assign by hand.
Company policies vary wildly: alcohol restrictions, seniority-based per-diems, international currency limits. A single rigid flow could not serve them.
Our answer: a configurable policy engine
When the virtual card fell through
Before sketching a single screen, we had to align on our core product philosophy. We evaluated the market and mapped our constraints against the three dominant expense models.
Too clunky, high anxiety and form-heavy.
True zero-touch filing. Tax liabilities, API limits and partnership timelines pushed a native card rollout to Phase 2.
Expensify’s frictionless batch capture, on our existing Astrada aggregator backend. As a B2B platform, onboarding can map straight to a company’s existing corporate cards and policies.
We benchmarked our feature set against industry leaders such as Navan and Ramp. We already had the heavy hitters in place, but had to make strict scoping decisions for the MVP.
Exception-only review
We already had this. Finance only sees what fails the pipeline.
SMS and email receipt forwarding
Low return for our corporate travel context, compared with capturing in the app.
AI chat interface
Can a conversational UI replace form-filling for edge cases?
How receipts meet transactions, and five checks on every one
Expense management is tied to the whole travel lifecycle, so I mapped an automated pipeline that fires the moment a receipt lands in the app. The aggregator pulls the card transactions for the dates of the trip, and each receipt is matched against them.
How it works
Card is used
The card the traveller already has.
Astrada pulls the feed
Live card transactions, for the dates of the trip.
A receipt lands
From the camera, the gallery or a document.
Five checks run
Fraud, OCR, trip, transaction match and policy.
Cleared or flagged
Cleared receipts approve themselves. Flags wait for a person.
Fraud detection
OCR validation
Trip matching
Transaction matching
Policy check
Engineering’s warning
Working memory holds only about four items at once. V1 showed every receipt, with every field, in one flat list, which pushed people into one-by-one processing. Progressive disclosure reveals the information in layers.
Layer 1: Glance
“5 checked. 2 cleared. 3 need you.”
Layer 2: Scan
The flagged rows, each with merchant, amount and state.
Layer 3: Resolve
The edit sheet, with tabs for Details, Trip, Payment and Policy.
The system never asks the user to hold more than about four pieces of information at once.
The backend was doing the work. The interface was making users redo it.
Since our feature set was already at industry benchmarks, the real problem was not what the system could do, but how it presented it. We found three flaws in our v1 design.
The system did the work. The UI made the user check it anyway.
The one-by-one bottleneck
Receipts were processed in linear tabs. A 20-receipt trip meant 20 separate screens.
Wasted automation
The system flagged receipts as AUTO-MATCHED, but the interface still made the user open the tab and check the OCR, the trip and the payment method by hand.
No macro view
There was no quick way to see the final report amount, or to clear the clean receipts in one go.
We explored three strategic approaches to the interface bottleneck, and weighed each one against what it cost the user and the roadmap.
Approach 1
Iterate the tabs
The idea
Keep the one-by-one tabbed flow, but make it easier to scan: better colour coding, confidence tags and clearer typography.
Why it failed
It doesn’t touch the interaction cost. The user still clicks through 20 tabs, however pretty they look, and the automation stays hidden.
Approach 2
Migrate to card-first
The idea
Pause the UI redesign and rebuild the backend to issue native virtual cards, like Brex, with instant push notifications for receipts.
Why it failed
Technical and regulatory roadblocks. Tax tagging at the point of sale and API partnerships would push the timeline out by 6+ months.
Approach 3
Keep Astrada, change the frontend
The idea
Why it won
It gives the zero-touch feeling of a card-first ecosystem on our existing aggregator stack, and solves the bottleneck now, without waiting on backend partnerships.
The decision
With the strategy settled, we took the finished review screen apart. Each piece got one verdict: keep it, iterate it, or build it new. Then we explored each one.
Review receipts
We checked 5 receipts against your trip, your card and Acme’s policy.
Total $371.45
Hilton Hotel
Feb 15–18 • Lodging
WeWork Day Pass
Feb 17 • Workspace
Uber
Feb 17 • Transport
Pret A Manger
Feb 16 • Meals
Wagamama
Feb 15 • Meals
Review 3 flagged receipts
2 cleared automatically · view
Segmented filter
Copy only
Batch summary card
Add a framing line
Receipt rows
Density and scale rules
Forward action
A state-aware label
Edit sheet
Keep the structure, iterate the defaults
Voice layer
A new atom, from the AI avatar screen
Hilton Hotel
Feb 15–18 • Lodging
$50 over the nightly lodging limit. Add a note or split.
Say why in one sentence…
Save changes
Hilton Hotel · $250.00
$50 over the lodging limit
Removed from v1
The words changed as much as the layout. Every message now says what is wrong and what to do about it.
| Previous | Newly designed |
|---|---|
| WARNING: Hotel expense exceeds daily limit of $200 (×2) | POLICY · “$50 over the nightly lodging limit. Add a note or split.” |
| “Please provide a justification for this expense…” | “Say why in one sentence. Type it or speak it.” |
| “2 issues remaining — resolve to confirm” | “Needs you 3” and a button: “Review 3 flagged receipts” |
| Unnamed thumbnail tabs | Named rows: merchant, then date • category |
Previous
WARNING: Hotel expense exceeds daily limit of $200 (×2)
Newly designed
POLICY · “$50 over the nightly lodging limit. Add a note or split.”
Previous
“Please provide a justification for this expense…”
Newly designed
“Say why in one sentence. Type it or speak it.”
Previous
“2 issues remaining — resolve to confirm”
Newly designed
“Needs you 3” and a button: “Review 3 flagged receipts”
Previous
Unnamed thumbnail tabs
Newly designed
Named rows: merchant, then date • category
The sheet’s tabs are the content categories:DetailsTripPaymentPolicy
Each atom was drawn as several real screens. Every variant is captioned with what it optimises and what it costs, and the chosen one is ringed in green.
Atom 1
The first thing users see decides whether this is a to-do list or a supervisor’s dashboard.
V1 · Receipts first
Hilton Hotel
Feb 15–18 • Lodging
WeWork Day Pass
Feb 17 • Workspace
Optimises: Vertical space.Costs: State has to be inferred row by row, so the mental model is “chores”.
V2 · Total only
Total $371.45
Hilton Hotel
Feb 15–18 • Lodging
Optimises: A money anchor.Costs: Hides the split, and gives no proof the system worked.
V3 · State-split card
ChosenWe checked 5 receipts against your trip, your card and Acme’s policy.
Total $371.45
Costs: About 120pt of height.
V4 · Progress ring
3 of 5need you
Hilton Hotel
Feb 15–18 • Lodging
Optimises: A glanceable ratio.Costs: Reads as loading, and finance thinks in money, not percentages.
Chosen for impact because it answers “where do I stand?” before anything else.
Atom 2
The structure is fixed. What we explored is what the tabs say.
V1 · Current
Pret A Manger
Feb 16 • Meals
Optimises: Familiar status words.Costs: “Ready” and “Not Ready” say nothing about who acts next.
V2 · Cleared / Needs you
ChosenHilton Hotel
Feb 15–18 • Lodging
Optimises: Verbs that match the pipeline and the voice script.
V3 · Taxonomy
Hilton Hotel
Feb 15–18 • Lodging
Optimises: Grouping by policy area.Costs: Users think in receipts, not taxonomies, and the counts fragment.
V4 · Question
Hilton Hotel
Feb 15–18 • Lodging
Optimises: Plain language.Costs: “Done” is ambiguous next to “submitted”.
44pt segmented control · white pill on #ECECE9 · counts in tabular-nums
Chosen for impact because the labels say who owns the next move.
Atom 3
V1 · Full row, 96pt
Hilton Hotel
Feb 15–18 • Lodging
New York Business Trip
WeWork Day Pass
Feb 17 • Workspace
New York Business Trip
Optimises: The trip name on every row.Costs: Four rows per screen, so 15 receipts is four scrolls.
V2 · Compact row, 72pt
ChosenHilton Hotel
Feb 15–18 • Lodging
WeWork Day Pass
Feb 17 • Workspace
Uber
Feb 17 • Transport
Optimises: About 5.5 rows per screen.
V3 · Two-column grid
Hilton Hotel
$250.00
WeWork Day Pass
$38.75
Uber
$25.00
Pret A Manger
$18.75
Optimises: More receipts per screen.Costs: Amounts don’t align, and the scan path zig-zags.
V4 · State-grouped sections
Needs you (3)
Hilton Hotel
Feb 15–18 • Lodging
WeWork Day Pass
Feb 17 • Workspace
Cleared (2)
Pret A Manger
Feb 16 • Meals
Optimises: Separates the work from the proof.Costs: Duplicates the tabs and kills chronology. We kept the idea for the AI-native queue.
Scale rules at 10 to 15 receipts
3 need you · $313.75Jump
2 cleared · $57.70
Compact row: 72pt · 40pt thumbnail · merchant 15/20 semibold, truncated at 18 characters · “Feb 16 • Meals” 13/18 · amount 15/20 tabular, right · 16px status glyph · 4pt rail
Chosen for impact because one glance gives who, what kind, how much and what state.
Atom 4
V1 · Tabbed sheet (the base)
Hilton Hotel
Feb 15–18 • Lodging
$50 over the nightly lodging limit. Add a note or split.
Say why in one sentence…
Save changes
Optimises: One surface, with issues marked by dots.Costs: It opens on the first tab, even if the problem is elsewhere.
V2 · Single-scroll sheet
Details
Trip
Payment
Policy
…and it keeps going
Optimises: Everything in one place, issue sections highlighted.Costs: A long scroll, and the issues aren’t quantified.
V3 · Wizard
Step 1 of 3 · Policy
$50 over the nightly lodging limit.
Next
Optimises: Issues come first.Costs: A linear path is the old tab nightmare in new clothes.
V4 · Inline expand
Hilton Hotel
Feb 15–18 • Lodging
$50 over the nightly lodging limit.
Say why…
Keyboard covers the list
Optimises: No sheet to open.Costs: The list jumps, the keyboard covers it, and the pinned context is lost.
Final · V1 with three tweaks
ChosenHilton Hotel
Feb 15–18 • Lodging
$50 over the nightly lodging limit. Add a note or split.
Say why in one sentence…
Save & next flag · 1 of 3
Top radius 20 · grabber 36×4 · header 64 · tab bar 44 · fields 48 with 12 radius
Chosen for impact because it lands on the problem and chains one fix into the next.
Atom 5
V1 · Static
Review 3 flagged receipts
2 cleared automatically · view
Optimises: Simplicity.Costs: It never changes, so it never rewards.
V2 · Dual call to action
Proceed with 2 cleared
Resolve 3
Optimises: Both outcomes are on screen.Costs: Two primaries compete for the thumb.
V3 · Gated submit
Submit report
Resolve 3 flagged receipts to continue
Optimises: Nothing goes in until it’s clean.Costs: The locked button, and its anxiety, come back.
V4 · State-aware morph
ChosenReview 3 flagged receipts
2 cleared automatically · view
Submit report · $371.45
2 cleared automatically · view
56pt full-width pill · ghost line 13/18, centred, with a link
Chosen for impact because the button always says the one thing to do next.
Atom 6
A screen can show many things at once. A voice can say only one, so the card under the avatar is the violation, not a flight.
V1 · Avatar, card and chips
SpeakerMuteMinimiseClose
Hilton Hotel · $250.00
$50 over the nightly lodging limit
Optimises: Everything on screen at once.Costs: The transcript is hidden, and the large avatar dominates.
V2 · Avatar and live transcript
First: Hilton, $50 over the lodging limit.
Client dinner, Acme deal.
Hilton Hotel · $250.00
$50 over the nightly lodging limit
Optimises: Nothing is missed.Costs: Reading and listening compete, so it isn’t eyes-free.
V3 · Audio-first, the card as an echo
ChosenHilton Hotel · $250.00
$50 over the lodging limit
Idle: the avatar fades to a small orb and waits.
The voice script
Greeting. A short hello.
The batch. “5 scanned, 2 cleared, 3 need you.”
For each flag. State the violation, listen, confirm, then “next”.
Closure. “All set. 5 submitted, $371.45.”
Chosen for impact because everything is actionable by voice, with the card only echoing it.
Before the atoms went together, we tried four basic structures for the whole screen, each as a real screen at the same width.
Structure 1 · The Excel table
Review receipts
| Receipt | Date | Category | Trip | Payment | Policy | Amount |
|---|---|---|---|---|---|---|
| Hilton Ho | Feb 15 | Lodging | NYC | Card | Over | $250.00 |
| WeWork Da | Feb 17 | Workspa | NYC | Card | Over | $38.75 |
| Uber | Feb 17 | Transpo | NYC | Card | Over | $25.00 |
| Pret A Ma | Feb 16 | Meals | NYC | Card | OK | $18.75 |
| Wagamama | Feb 15 | Meals | NYC | Card | OK | $38.95 |
Optimises: Everything visible at once.Costs: A 390px screen can’t hold seven columns, and you can’t act from a cell.
Structure 2 · The accordion
Review receipts
Hilton Hotel
Feb 15–18 • Lodging
WeWork Day Pass
Feb 17 • Workspace
Uber · $25.00
Trip
Payment
Policy
Note
Save
Pret A Manger
Feb 16 • Meals
Wagamama
Feb 15 • Meals
Optimises: Depth without leaving the list.Costs: Every peek costs two taps, and the batch disappears off screen.
Structure 3 · Quick-action rows
Review receipts
Hilton Hotel
$250.00
WeWork Day Pass
$38.75
Uber
$25.00
Pret A Manger
$18.75
Wagamama
$38.95
Optimises: Speed for the clean majority.Costs: It approves sight unseen, which trains blind trust, risks mis-taps, and flags still need depth.
Structure 4 · Problem-first inline
Needs you
2 cleared · hidden
Hilton Hotel$250.00
Say why…
WeWork Day Pass$38.75
Say why…
Uber$25.00
Say why…
Optimises: Focus on the exceptions.Costs: It erases proof the system worked, and users ask where the rest went. We reused it for the AI-native queue.
Review receipts
We checked 5 receipts against your trip, your card and Acme’s policy.
Total $371.45
Hilton Hotel
Feb 15–18 • Lodging
WeWork Day Pass
Feb 17 • Workspace
Uber
Feb 17 • Transport
Pret A Manger
Feb 16 • Meals
Wagamama
Feb 15 • Meals
Review 3 flagged receipts
2 cleared automatically · view
Chosen
How they compare
The Excel table
Scan time at 15 receiptsWeak
Taps to resolve 3 flagsWeak
Proof the system worked is visibleStrong
Voice parityWeak
The accordion
Scan time at 15 receiptsPartial
Taps to resolve 3 flagsWeak
Proof the system worked is visiblePartial
Voice parityPartial
Quick-action rows
Scan time at 15 receiptsStrong
Taps to resolve 3 flagsPartial
Proof the system worked is visiblePartial
Voice parityPartial
Problem-first inline
Scan time at 15 receiptsStrong
Taps to resolve 3 flagsStrong
Proof the system worked is visibleWeak
Voice parityStrong
Final: summary, list and sheet
Scan time at 15 receiptsStrong
Taps to resolve 3 flagsStrong
Proof the system worked is visibleStrong
Voice parityStrong
A design judgement from walking each structure through the same fifteen receipts, not measured data. Problem-first wins on focus, which is why we reused it as the AI-native queue.
The decided design
The decided model: an exception-first batch review where the pipeline’s work is visible, cleared receipts approve themselves, and humans touch only flags, entered through a list, a table, or a conversation.
Review receipts
We checked 5 receipts against your trip, your card and Acme’s policy.
Total $371.45
Hilton Hotel
Feb 15–18 • Lodging
WeWork Day Pass
Feb 17 • Workspace
Uber
Feb 17 • Transport
Pret A Manger
Feb 16 • Meals
Wagamama
Feb 15 • Meals
Review 3 flagged receipts
2 cleared automatically · view
Segmented filter. All, Cleared and Needs you: attention without hiding the batch.
Batch summary. Where do I stand, in under two seconds.
Receipt rows. Who, what kind, how much and what state.
Forward action. One primary action, always the next one.
Edit sheet. One sheet, four rooms, opening on the problem.
Voice layer. The same queue, spoken.
A clean list. Tapping a receipt opens a bottom sheet with the OCR data, the trip match and the policy status, so a justification takes seconds.
Review receipts
We checked 5 receipts against your trip, your card and Acme’s policy.
Total $371.45
Hilton Hotel
Feb 15–18 • Lodging
WeWork Day Pass
Feb 17 • Workspace
Uber
Feb 17 • Transport
Pret A Manger
Feb 16 • Meals
Wagamama
Feb 15 • Meals
Review 3 flagged receipts
2 cleared automatically · view
Hilton Hotel
Hilton Hotel
Feb 15–18 • Lodging
$50 over the nightly lodging limit. Add a note or split.
Say why in one sentence…
Save & next flag · 1 of 3
The summary sets the mental model: the system has done the work, and you supervise.
Dots on the tabs show where the issues are, without opening each one.
“Save & next flag” chains one fix into the next.
A dense table where out-of-policy flags are resolved inline with dropdowns.
3 flagged · 2 cleared
Approve 2 cleared · $57.70
| Receipt | Traveller | Amount | Resolve |
|---|---|---|---|
| Hilton Hotel | A. Rao | $250.00 | Choose |
| WeWork Day Pass | A. Rao | $38.75 | Choose |
| Uber | M. Chen | $25.00 | Choose |
| Pret A Manger | M. Chen | $18.75 | Auto-approved |
| Wagamama | J. Okafor | $38.95 | Auto-approved |
Ten flags, zero sheets.
Finance resolves each flag in its row, with a dropdown, and approves the cleared receipts in one action.
To test the hypothesis that filing should feel like talking to an assistant, we built a headless conversational interface. Travellers drop receipts and let the AI sort them. Voice-to-text (through Wispr) and a digital avatar let them literally speak a justification — "this dinner was with the client, John Doe, which is why it's over the $50 limit."
One entry point for camera, gallery or a document.
Receipt capture, with the amount picked out live.
Batch-select what you already photographed.
Hilton Hotel · $250.00
$50 over the lodging limit
1 of 3
Hilton Hotel$250.00
$50 over the lodging limit
2 cleared · hidden
Idle: the avatar fades to a small orb and waits.
The script, in one line: “5 scanned, 2 cleared, 3 need you. First: Hilton, $50 over the lodging limit.”
The assistant proposes the whole job: scan, match, policy check, in one go.
Each stage surfaces as it runs — OCR, transaction matching, policy checks.
The full picture: 14 verified and matched, 2 that need attention.
The auto-approval rule
Managers are pulled in only for exceptions, which is where the hours of review time go.
Proceed with the 14 that cleared, or resolve the two that did not.
One receipt, one problem at a time: assign it to a trip.
Then match it to the card transaction the aggregator already pulled.
Scope note: because of the tax and legal liabilities involved, the payout and reimbursement module was deliberately left out of this phase. The work covers capture, validation and approval routing.
Context promised two things: speed for travellers, and defensible numbers for finance. Capture and approval deliver the first. These three flows deliver the second.
Approved expenses map automatically to GL codes, ready to export to NetSuite, Xero or Sage.
Why: Finance’s numbers land where finance lives.
Every user carries a score. A consistently low one flags their future reports to managers and can reduce their future per-diem limits.
Why: It is a behaviour feedback loop, not a punishment. The score is visible to the user, which is what removed the anxiety.
Automated 30-day submission reminders, with escalation from manager to travel coordinator to finance.
Why: The email loop dies when deadlines have owners.
The pilot cohort, in the coded prototype
We took the decided design to the pilot cohort. Four things worked as intended. Four did not, and each became a change.
Who
The pilot cohort: travellers from the internal pilot, our finance stakeholders and a manager.
Where
In the coded prototype, not Figma. Latency states and voice input can’t be tested in static frames, so testing happened in the coded prototype with real receipt data.
Tasks
Measured
Review receipts
We checked 5 receipts against your trip, your card and Acme’s policy.
Total $371.45
5 receipts checked
Hilton Hotel
Feb 15–18 • Lodging
Pret A Manger
Feb 16 • Meals
Say why: “Hilton Hotel, $250.00, $50 over the limit.”
What worked
“Dropped them and walked away. The summary told me it was handled.”
“The checklist filling step by step proved it wasn’t just spinning.”
“Speaking the justification beat typing it on the metro.”
“I finally saw what passed, so I stopped double-checking.”
What we changed
“The red rail felt like failure, even for warnings.”
Tweak: A warning gets an amber rail. An error gets a red one.
“Let me peek at what “Ready” contains before I trust it.”
Tweak: The Cleared tab lists the receipts that passed.
“The avatar staring at me was uncanny.”
Tweak: The avatar stays dormant until it is listening.
“Finance: ten flags shouldn’t mean ten sheets.”
Tweak: On the web table, flags resolve inline.
Web, finance
| What they said | What we changed | Atom touched |
|---|---|---|
| “The red rail felt like failure, even for warnings.” | Warnings went amber. Errors stay red. | The receipt row |
| “Let me peek at what “Ready” contains before I trust it.” | The Cleared tab lists the receipts that passed. | The filter tabs |
| “The avatar staring at me was uncanny.” | The avatar stays a dormant orb until it is listening. | The voice layer |
| “Finance: ten flags shouldn’t mean ten sheets.” | On the web, flags resolve inline in the table. | The edit surface |
“The red rail felt like failure, even for warnings.”
Warnings went amber. Errors stay red.
The receipt row
“Let me peek at what “Ready” contains before I trust it.”
The Cleared tab lists the receipts that passed.
The filter tabs
“The avatar staring at me was uncanny.”
The avatar stays a dormant orb until it is listening.
The voice layer
“Finance: ten flags shouldn’t mean ten sheets.”
On the web, flags resolve inline in the table.
The edit surface
Situations a single rigid flow could not serve
Real expenses are untidy: foreign currency, a hotel bill with a minibar on it, a receipt left on a train. Each needed its own small piece of interface.
Izakaya Tanuki
Receipt · Tokyo
¥6,200
OCR: ≈ $41.80 @ trip-date rate 148.33
Card feed
Matched to the receipt
$41.80 Matched
Meals limit $50: in policy
A receipt in yen is converted at the trip-date rate: ¥6,200 becomes $41.80. The rate shows on the receipt side and on the transaction side, so a reviewer can see that both agree.
Hilton Hotel
$250
Approve $230
Justify $20
The $250 Hilton bill splits into a $230 room, in policy, and a $20 minibar charge, flagged. Policy runs per line, so the traveller approves the room and justifies only the wine.
No receipt? Declare it.
Amount$41.80
MerchantIzakaya Tanuki
ReasonLost it on the train
Sign here
Affidavit #A-2291 attached
A guided affidavit with an e-signature replaces the free-text note. It leaves an auditable record that finance can stand behind.
Acme Corp: no receipts under $75. Matched from the card feed.
Black Sheep Coffee
No receipt needed
$6.40
Uber
No receipt needed
$25.00
Pret A Manger
No receipt needed
$18.75
Where a company waives receipts under a threshold, matching runs on the Astrada feed alone. Rows carry a “No receipt needed” chip instead of a missing-receipt flag.
AssistantI heard “client dinner with John Doe”. Yes or no?
Yes.
Justification attached
A mishear is the voice flow’s worst case. The assistant reads back what it heard and asks yes or no before anything is attached to the expense.
The decided design, in four frames
We checked 5 receipts against your trip, your card and Acme’s policy.
Total $371.45
Hilton Hotel
Feb 15–18 • Lodging
WeWork Day Pass
Feb 17 • Workspace
Pret A Manger
Feb 16 • Meals
Review 3 flagged receipts
2 cleared automatically · view
Hilton Hotel
Feb 15–18 • Lodging
$50 over the nightly lodging limit. Add a note or split.
Say why in one sentence…
Save & next flag · 1 of 3
3 flagged
Approve 2 cleared · $57.70
| Receipt | Amount | Resolve |
|---|---|---|
| Hilton Hotel | $250.00 | Choose |
| WeWork Day Pass | $38.75 | Choose |
| Uber | $25.00 | Choose |
| Pret A Manger | $18.75 | Auto-approved |
Hilton Hotel · $250.00
$50 over the lodging limit
Idle: the avatar fades to a small orb and waits.
The prototype
Measured against the baseline, and the pilot
This was a one-week sprint to an MVP followed by an internal pilot, so impact is measured through baseline comparisons and adoption signals rather than a long-run production dataset.
~90 min → under 4
Average time to create an expense report, from the spreadsheet era to the AI-aggregator flow
80%
Estimated reduction in the time managers spend reviewing standard travel receipts
Creating a report
Review
Turnaround
Auto-approval
We replaced the tab-by-tab flow with a scannable batch table and an exception-only model. That restored the five-minute perception without changing backend performance. A second round of testing with the new structure showed:
Time to sort and verify
A behavioural shift
The reality check
Enterprise validation
Less anxiety
What leading this actually involved
Leading through a technical blocker
Changing how we prototype
Listening past the demo
Translating between two audiences
In conclusion