Paytm Intelligence · PM assignment · Drishtant Kaushal · vF · 29 Oct 2026

A Paytm agent that books your flight and looks after the trip

You say where you need to be and what you can spend. It finds the best fare by a date you agree, books it, and stays with you until you're home.

Replaces Paytm Chat, right of homeChat + background jobsUPI One Time Mandatev1 is domestic flightsNo web or code access in MVP

Assumptions. This surface exists to bring existing users back more often. The agent uses what Paytm already knows about a user and spends only money the user blocks with their UPI PIN. In the MVP the agent has no web access and no code execution, and it calls only Paytm's own tools.

1 · Objective

What I understand by a personal consumer agent

An agent that does the job for you while you're out of the app.
It knows your goals and history, works for days, and spends only what you approve.
It's chat, search and agent agnostic, so you just say what you need.

2 · Why

Paytm's mission and vision

Paytm's mission is to bring half-a-billion Indians to the mainstream economy.^1,2
- 8 crore Indians transact on Paytm every month.³
- Financial services already make up 33% of Paytm's revenue.⁴
But what does the economy of the future look like? Agents > Humans.
- Machine customers will outnumber human ones 1.8x by 2028.⁵
- Agents will drive $3 to 5 trillion of commerce by 2030.⁶
- NPCI has already piloted UPI payments inside ChatGPT.⁷
So what does that mean for Paytm's mission? An agent sandbox for each consumer.
- Har ghar agent, one trusted agent in every home.
- It's a Soundbox moment, this time for consumers.⁸
- 61% of Indians already accept AI purchases they approve.⁹

3 · How

Plan and roadmap

3.0Decision · always-on or pre-configured, and why

The best experience is an always-on agent, but I'd park it and launch with pre-configured agents

why, before the plan

An always-on agent that runs any job is where this ends, and it is the long-term plan.
Paytm moves money in a regulated space, so every rupee needs a fixed, auditable path.
- NPCI wants an agent to execute payments, never to approve them.¹⁰
- NPCI's own rules for agent payments are still a draft.¹¹
Paytm's jobs are bounded money tasks, not open-ended knowledge work.
- Booking, paying and tracking all run on Paytm's own APIs.
The agent acts on the user's identity, so it gets the least access possible.
- It handles a UPI PIN, verified names and travel IDs.
- Browsing agents still fall for prompt injection 11% of the time.¹²
A computer per user is expensive across crores of users.¹³
Trust is earned one job at a time, and then the agent can widen.

3.1Long term · always-on agent

A money companion for every Paytm user

what the personal agent becomes

Every user gets a private space where many agents work for them.
All the agents share one memory of the user.
Each agent acts only within the permissions and money the user grants.
The user can see, pause or revoke any agent at any time.

What it takes at the scale of crores of users

inputs for the long term

A consent and permissions layer across every Paytm service.
A shared user card that every agent reads, built in-house.
- Honcho, by Plastic Labs, builds such a card from a user's conversations.¹⁴
- Its AGPL licence and per-token billing don't fit crores of users.^15,16
A registry where new agents plug in without the user starting over.
A running cost per monthly user that Paytm can carry.

Long term is what the personal agent becomes. Short term is the first job that earns the right to get there.

3.2Short term · pre-configured agents

One agent first, built on a Paytm service

flights now

Recharge & Bills follows once the first agent earns trust.
Travel & Tickets goes first, starting with domestic flights.
Financial Services & Investments comes last, since money advice is regulated.
It runs as scheduled jobs that wake on fares and flight events, not as an always-on agent.

Memory tier	What it holds about the user	When the agent uses it	How it's built
1 · Core profile	Name as on ID, home airport, travellers, seat and bag choices	Every turn, kept short	Paytm KYC, settings and past bookings
2 · History	Every conversation and every trip	Searched only when a detail is needed	Conversation log with full-text search
3 · Trip records	Tickets, invoices, boarding passes and refunds	When a trip needs them	Paytm's own document store
4 · Routines	Reusable plans, like a "Diwali trip home"	When a similar trip starts	Saved after each completed trip
5 · Trip notes	One-trip details, like "a 6am flight is fine this time"	Only during that trip	Deleted when the trip ends

Adapted from Hermes Agent's tiers and the always-on agent talk.^17,14,13 The shared user card joins in the long term.

	Short term · pre-configured agent	Long term · always-on agent
Job	One fixed job, starting with flights	Any job the user asks for
Runs on	A durable workflow per trip	Its own computer per user, always running
Wakes on	Fare checks, flight events and the user's message	Anything, from messages to schedules to other agents
Tools	Paytm's own APIs only	APIs, then a browser, then a desktop
Model use	Only when the user talks or something changes	Most turns
Cost	Close to nothing between events	A computer and storage for every user

3.3Why I chose flights first

Recharge & BillsTravel & TicketsFinancial Services & Investments

Earnings per task come from ixigo's FY25 data, Paytm's bill fees, recharge commission rates and digital loan sizes.^18,19,20,21 Effort is my judgment, so treat positions as directional.

Why I chose flights

four reasons

The ticket is big enough for a saving to matter.
- A flight averages ₹6,695, against ₹772 for a train.¹⁸
- A ticket 7 days out cost 44% more than one 30 days out.²²
The margin leaves room to try new things.
- A flight earns about ₹137 per booking, against ₹1 to ₹8 for a bill.^18,19
- A watch costs ₹1 to ₹23 to run, so a ₹49 hold or a small miss fits.^23,24
The agent controls the outcome, since seats are priced and not rationed.
- IRCTC, by contrast, blocks automated train booking.²⁵
The market already trusts agents with flights first.
- Amazon Pay and ixigo made agent-paid flight booking live in Sep 2026.²⁶

Why not the others yet

in Paytm app order

Recharge & Bills reach the most people, but AutoPay already does the job.²⁷
- Bharat Connect handles about 27 crore bill payments a month.²⁸
- Card bills average ₹16,537, above AutoPay's ₹15,000 no-PIN limit.^28,29
Trains ration berths, and buses earn about half what a flight does.^30,18
Financial Services & Investments earn the most, but advice on money is regulated.²¹

3.4User persona

Ankit, 27

Pune to Lucknow, 3 trips a year

He works in Pune and his family lives in Lucknow.
He flies home three times a year, alone or with his parents.
He pays by UPI and opens Paytm every week.
He has never let an app spend money for him.

Why him

the segment behind him

1 in 3 urban travellers plan three or more trips a year.³¹
49% of those frequent travellers are 18 to 34.³¹
73% of Paytm users live in Tier II cities and beyond.³²
Festival trips home are when fares hurt most.³³

3.5User journey

The moment

seven weeks before Diwali

His mother asks on the family group when he's coming home.
He opens a travel app and sees ₹7,900 for a one-way flight.
Last year he waited and paid even more.

What he does today

and what it costs him

He sets a Google Flights alert, which only sends an email.³⁴
He checks Skyscanner and ixigo every few days.
He hunts for bank offers and coupons.
He books late and finds a convenience fee at checkout.³⁵

TodayWith the agentMood from −2 to +2, scored from the evidence in the table

Stage	Today	Evidence	What can go wrong	Who handles it
1 · Pre-booking	Sees ₹7,900 one way and feels it's too much	Festive outbound fares run 3 to 4x the return leg³³	Wrong name on ID, unrealistic budget, no mandate support	Orchestrator, Customer, Reviewer, Payments
2 · Booking	Checks apps for weeks, then meets a surprise fee	80% met fees only at the final step³⁵	Fares break history, fare gone at re-price, payment timeout	Scheduler, Fare Analyst, Flight, Reviewer, Payments
3 · Pre-flight	Unsure about baggage, seat and check-in	54% find choosing baggage stressful³¹	Schedule change, baggage rule change, price drop	Trip, Notifier
4 · Day of flight	Guesses when to leave and watches the board	10.46 lakh passengers hit by cancellations in Dec 2025³⁶	Delay, cancellation	Trip, Flight, Notifier
5 · After landing	Waits at the belt and chases any refund	Refund complaints tripled to 4,386 in 2025³⁷	Lost bag, stuck refund	Trip, Payments, support

3.6Agent progression · one screen per stage

1 · Pre-booking

9:415G · 82%

PiPiHome chat

Diwali pe Lucknow jaana hai, 6 Nov ko. ₹6,500 tak.

Pune to Lucknow, 6 Nov. Just you, Ankit Sharma, as on your ID?

Haan, sirf main.

Last month this ran ₹6,900 to ₹8,400. ₹7,200 has a good chance.

7,000 kar do.

Tight but possible. Morning flight and a 15 kg bag?

Block ₹7,000 for this trip

Book for me · enter UPI PINJust tell me

I'll decide by 6 Oct and release it if I don't book.

Ask Pi in Hindi or English

2 · Booking

9:415G · 82%

PiLucknow tripWatching

Day 6 · push

Fares are up 9% this week. Your ₹7,000 still has a fair chance.

Day 14 · push

Fares are running above last year

IndiGo cut two flights on this route. Your book-by date is in 4 days.

Book ₹7,480 nowTry 5 NovKeep ₹7,000

5 Nov try karo.

Day 15 · push

Booked

IndiGo · 5 Nov07:10

Paid₹6,780 with all fees

Back to HDFC₹220

Ask Pi in Hindi or English

3 · Pre-flight

9:415G · 82%

PiLucknow trip4 Nov

4 Nov · 18:00

Checked you in. Seat 12A, free, like last time.

Allowed by your rule · web check-in

Laptop bag bhi le ja sakta hoon?

Yes. One cabin bag up to 7 kg plus a laptop bag. Your fare has 15 kg check-in.

Boarding pass

Seat12A

GateOpens 2 hours before

Add to DigiYatraOpen

Ask Pi in Hindi or English

4 · Day of flight

9:415G · 82%

PiLucknow tripToday

5 Nov · 04:30

Leave by 4:45. Traffic to Pune airport is light.

5 Nov · 06:10

Delayed by 50 minutes

New departure08:00

Gate6

OKShow other flights

OK.

I'll tell you when boarding starts.

Ask Pi in Hindi or English

5 · After landing

9:415G · 82%

PiLucknow tripLanded

5 Nov · 10:15

Welcome to Lucknow. Your bag is on belt 3.

Should I tell Mummy you've landed?

Ask first · messaging people

Haan.

Sent on WhatsApp. Your return on 10 Nov is ₹5,400 today. Should I watch it?

Haan, ₹5,500 tak.

Block ₹5,500 for the return

Enter UPI PIN

Ask Pi in Hindi or English

3.7What "within budget" means

The agent books the best fare it can find by a date he agrees

a budget is a target, not a promise

The budget is the most he'll pay for the whole party, fees included.
The agent never books above it, and the blocked amount enforces that.
Before any money is blocked, it rates the budget as on track, tight or unlikely.
Every watch gets a book-by date, 14 days out for domestic trips.
- Festival trips get 30 days, since fares climb earlier.
- A ticket 3 days out cost 61% more than one 22 days out.³⁸

When fares break history

it asks him to choose

History does break, as fares crossed ₹1 lakh in December 2025.³⁹
The Fare Analyst works out why a route moved, and the agent passes that on.
The agent nudges him when the odds fall or the book-by date nears.
- A rational buyer's limit rises as the deadline nears.⁴⁰
- Waiting saves only about 4% on average, so it doesn't gamble late.⁴¹
Each nudge offers exact choices, like booking now, another day, or a new budget.
If he doesn't answer by the book-by date, the money is released.

3.8How agent pays

Primary · UPI One Time Mandate

one per trip

His PIN blocks the budget once, and the airline is paid once later.⁴²
- The API call is recurrencePattern ONETIME with blockFund true.⁴³
- A smaller debit releases the rest straight back.⁴²
- Blocks last up to 60 days with no second PIN.⁴⁴
The normal UPI limit applies, so fares up to ₹1 lakh fit.⁴⁵
- Verified travel merchants go up to ₹5 lakh.⁴⁶
- The ₹15,000 no-PIN rule covers recurring mandates only.²⁹

Fallback · hold the fare, then he pays

when the mandate can't carry it

Pi holds the fare and sends a pay link when the mandate can't pay.
- His bank doesn't support the One Time Mandate yet.
- The best fare is above his budget and needs his OK.
- His bank's limit is below the fare, or he chose "just tell me".
IndiGo's Fare Hold starts at ₹49 and lasts up to 96 hours.²⁴
The hold fee needs its own approval, since the mandate allows one debit.⁴²

Trip extras · UPI Reserve Pay, once enabled

seat, bag, meal, cab

One block could cover many small debits during the trip.
- Reserve Pay caps a block at ₹10,000 for 90 days.⁴⁷
- NPCI limits it to low-ticket, frequent merchants, and travel isn't enabled yet.⁴⁷
Until then, extras go through a normal UPI payment with a PIN.

3.9Why Paytm

Over ChatGPT and Gemini

Paytm can finish the trip

AI assistants find flights and hand booking to someone else.
- ChatGPT's ixigo app sends users back to ixigo to book.⁴⁸
- Google AI Mode tracks fares but books only US hotels.⁴⁹
Amadeus supplies the fares, and Paytm tickets and refunds as the seller.⁵⁰
- GDS agent APIs are sold to accredited sellers like Paytm.⁵¹

Over ixigo, MakeMyTrip and Amazon Pay

a lead, not a moat

For its own users, Paytm blocks, debits and releases the money in one app.
- ixigo had to partner with Amazon Pay to let its agent pay.²⁶
- Paytm handles about 8% of UPI volume, so the lead covers its own users.⁵²
Checkin charges no convenience fee.⁵³
Flights are the first job of a money agent that later pays bills and manages credit.

Reach and proof

already in the app

8 crore monthly users reach it with one swipe.³
Checkin's agentic pilot completes 7x more funnels.^54,55

3.10Measuring success

North star · trips the agent books end to end

counted per finished job

A trip counts when an approved watch ends in a ticket with no dispute for 7 days.
- Intercom bills Fin the same way, per resolved conversation.⁵⁶
- It doesn't hinge on the budget he first typed.
The quality check is how often the agent beats the fare on the day he asked.

Per 1,000 people who open the surface. The rates are my pilot targets, to reset after two weeks of real data.

3.11Failure modes

What goes wrong	What the agent does
The budget is unrealistic	Rates it at setup and suggests a budget with a real chance
Fares rise past history	Nudges early with exact options, then releases the money at the book-by date
The fare is gone at re-price	Takes no payment and keeps watching
Money is taken but no ticket is issued	Retries ticketing, then refunds within RBI's 5-day limit⁵⁷
The airline delays or cancels	Alerts at once and offers rebooking or a refund
A refund gets stuck	Follows up until it lands, then hands off to support

Appendix D lists 18 scenarios across the five stages.

4 · Prototype

Pi from the first message to landing

The screens play in order, and a label names each part as it appears. Tap a stage to jump, or pause to read. Appendix H has every screen with what runs under the hood.

Fares, dates, names, the PNR and bank details are example data.

Appendix

A. Product decisions

Solution option	User value	Feasibility	Business impact	Risk	Verdict
1 · Fare-odds badge on the booking page	Medium	High	Medium	Low	Ship anyway, no agent needed
2 · Chat wrapper over flight search	Low	High	Low	Low	Skip, it adds little over filters
3 · Watch and tell, user pays	High	High	Medium	Low	Build as the default
4 · Watch and book inside a blocked budget	High	Medium	High	Medium	Build for users who opt in
5 · Trip agent for check-in, delays and refunds	High	Medium	Medium	Medium	Build with 3 and 4

Where a rule does the job, use the rule. The agent covers the timing, the trade-offs and the explanation.

One endless chat is the one thing we won't build

single chat vs multi-chat vs multi-bot

Long contexts degrade the model.
- Short prompts beat 113,000-token prompts on every model family.⁵⁸
- At 32,000 tokens, 11 of 12 models lose half their accuracy.⁵⁹
Many bots confuse users and split memory.⁶⁰
So the user gets one conversation, and each trip runs in its own context.
Single-chat apps cope with memory lists and resets.
- Meta AI keeps one thread with a reset command.⁶¹

No web and no code for the consumer agent

money calls for caution

In payments, the agent gets only the access the job needs.⁶²
- It never touches a database, only typed APIs that check permissions.⁶²
- Its access is decided by infrastructure, never by the model.⁶²
The MVP has no web browsing, which shuts out prompt injection from pages.^63,64
The consumer agent has no code execution, since its toolset is small and fixed.⁶⁵
The Fare Analyst is the one exception, running code in a sandbox on fare data only.
If an airline API fails, Pi tells the user instead of trying a browser.¹³

Always-on behaviour without an always-on agent

how the short-term Pi copes

Workflows can retry a step, so each money step checks status first and carries an idempotency key.⁶⁶
Each wake-up rebuilds a short trip summary, so a cold model cache costs little.⁶⁷
Routine checks run on code, and the model runs only when something changed.
Pi speaks only when it books, the odds fall, the date nears, the flight changes or money moves.¹³
The user's message reaches a running watch at once, as a workflow signal.⁶⁸
Approvals stay open until the book-by date, and silence never counts as yes.
Each watch pins its model and prompt version and is tested on replayed fare histories.

One horizontal orchestrator over vertical agents

horizontal vs vertical

The orchestrator owns the conversation across every domain.
Each vertical agent owns one domain and its tools.
Multi-agent setups cost about 15x the tokens of a chat.⁶⁹

Text and voice, both

conversation vs voice

The user can type or speak in English, Hindi or Hinglish.
14% of Paytm users already use the app in their own language.³²
The agent reads confirmations aloud, the way Soundbox does.

RAG only for airline policy pages

do we need RAG?

Fare rules arrive as structured fields from the supplier.
Baggage and refund pages are long free text, so those go in a vector store.

B. User and agent flow

What we know about a returning user

from Paytm's own data

Saved travellers and how he books for them.
Past trips, airlines, seats and bag choices.
His payment methods and which bank he uses.
Booking SMS and contacts he has allowed Paytm to read.⁷⁰

What we infer for a first-time user

before asking anything

His home city from location and where he pays.
His nearest airport from that city.
Likely trips home from transfers to family.
Then he answers three questions, not twenty.

Settings Pi needs for check-in

identity and rules

The traveller's name must match the ID exactly.⁷¹
Aadhaar or DigiLocker verification confirms each traveller's name before money is blocked.
DigiLocker links each traveller's ID for the airport.
- Airports generally accept IDs from DigiLocker's issued documents.⁷²
DigiYatra receives the boarding pass after web check-in.⁷³
- DigiYatra runs at about 38 airports and is optional for domestic flights.⁷⁴
Rules show what Pi does alone and what it asks first.
- Pi watches, checks in, keeps the same seat and sends alerts on its own.
- It asks before hold fees, higher fares and name changes.

C. Agent architecture

C.1Sub-agents and tools

Sub-agent	Owns	Runs on	Tools
Orchestrator	The conversation and every hand-off	Jev + LLM	route_message, get_session, hand_off, ask_user, request_approval, summarise_context, escalate_to_human
Customer	Who is travelling and how they like it	LLM	get_profile, get_preferences, list_travellers, add_traveller, verify_traveller_details, get_past_trips, read_booking_sms, pick_contact, update_memory
Flight	Finding, forecasting, pricing and booking	LLM + XGBoost	search_flights, get_fare_history, forecast_fare, rate_budget, suggest_budget, search_nearby_dates, get_fare_rules, get_add_ons, reprice, hold_fare, create_booking, issue_ticket
Trip	Everything after the ticket	LLM	track_flight, open_web_checkin, select_seat, get_boarding_pass, get_baggage_rules, estimate_leave_time, rebook_flight, cancel_booking, raise_refund, track_refund
Notifier	Telling the user what happened	LLM	update_job_card, send_push, send_whatsapp, send_sms, read_aloud, translate_message
Payments	How money moves	code	list_payment_methods, check_mandate_support, create_mandate, debit_mandate, revoke_mandate, pay_by_card, send_pay_link, create_reserve_pay, debit_reserve_pay, refund_to_source, get_payment_status
Scheduler	Time	code	create_watch, set_book_by_date, set_timer, raise_check_rate, wake_watch, extend_watch, expire_watch, list_jobs
Fare Analyst	Why fares moved on each route	code, sandboxed	load_fare_history, detect_regime_break, correlate_signals, explain_move, recalibrate_odds, set_route_book_by
Reviewer	Checking every action before it runs	Jev + code	check_required_details, check_budget, check_fare_rules, verify_reprice, check_duplicate, score_risk, approve_reversible, request_user_approval, write_audit_log

Every tool call carries the user's scoped session token from his Paytm login, so no agent can act outside his account, and the scope survives every hand-off.⁶²

C.2Models

Model	Use it for	Why
Classical ML (XGBoost)	The fare forecast and budget odds	Fares are numeric time series, and XGBoost hit R² 0.98 on Indian fares⁷⁵
System One (Jev)	Routing, "book now?", reviews	Typed answers with calibrated confidence in 70 to 500 ms^23,76
LLM (small, self-hosted)	The conversation and messages	Paytm reports its own 4B model for low cost⁷⁷
Reasoning LLM (frontier)	Disruptions and rebooking choices	Slow and costly, so it runs only on hard calls⁷⁸

Why XGBoost for the forecast

what it does

It builds hundreds of small decision trees, each fixing the last one's errors.
It reads tabular inputs like days to departure, weekday, season and seats left.
It trains fast and shows which inputs drove each forecast.
Deep models need far more data and are harder to explain.

How System One differs from classical ML

Jev vs XGBoost

XGBoost is trained per task on numeric features.
Jev answers typed questions about messy text with no training.
Both return a number, and neither writes text.

C.3Harness

Data

what the agents read

Paytm data covers transactions, bookings, travellers, SMS, contacts and location.⁷⁰
Supplier data covers live fares, fare rules and add-ons from Amadeus.⁵⁰
Our own store keeps every fare snapshot for the forecast.

Guardrails

ask, collect and confirm

The Reviewer blocks the mandate until seven details are confirmed.
- Origin, destination, travel date and return date.
- Passenger count and each passenger's name, age and contact.
- Preferences for seat, bag and timing.
Agents invent missing details, so the Reviewer asks instead.⁷⁹
Device binding and the UPI PIN prove it's the user.
- NPCI wants the agent to execute, never to approve.¹⁰
SMS and messages can never change the budget or the payment.⁶³

Tracing and environment

see and test every run

Langfuse keeps one session per watch, with every turn traced.⁸⁰
Approval-gated tools pause the run until the user confirms.⁸¹
Staging runs against Amadeus and UPI sandboxes.
Evals run on simulated users, like τ-bench's airline tasks.⁸²

Permission tier	Examples	Guardrail
0 · Read	Fares, booking status, refund status, saved travellers	Scoped to the logged-in user only
1 · Reversible write	Start or adjust a watch, check in, keep the same free seat, raise a refund	The Reviewer approves it inside the blocked budget, with an idempotency key
2 · Money or identity	Block a budget, pay a hold fee, pay above budget, change a traveller's name	The user approves with UPI PIN or a step-up check
Never	Pay above the budget, move money without a PIN, promise a refund	Hard-blocked in the policy layer, not the prompt

Grounding and hand-off

rules the model can't bend

Fares, dates and amounts come only from tool responses, and the model only phrases them.
A tool timeout is reported plainly, never guessed.
Repeated failure, fraud language or a request for a person moves the chat to support.
- Support gets the case summary, so the user never starts over.

C.4Backend

Store	Holds
Postgres	Profiles, user cards, travellers, watches, bookings, mandates, approvals and the audit log
Time-series table	Every fare snapshot by route and date
pgvector	Searchable history and airline policy pages
Redis	Session state, fare cache, rate limits and idempotency keys
Temporal	The process tracker that runs each watch for weeks⁶⁶
Kafka	Fare, PNR and flight-status events
Object store	Tickets, boarding passes and invoices
Feature store	Inputs for the fare model

D. Failure scenarios

Stage	Layer	What goes wrong	What we do
Pre-booking	Model	The LLM guesses a missing name or date	The Reviewer blocks it and the agent asks
Pre-booking	System	The name doesn't match the ID	Stop before booking and confirm the spelling
Pre-booking	System	The bank doesn't support the mandate	Fall back to a fare hold and pay link
Pre-booking	Product	An infant or child changes the price	Include every passenger's fee in the budget
Booking	Model	Jev routes a message with low confidence	Hand it to the LLM or ask the user
Booking	Model	The forecast misses and the fare rises	Track regret and retrain on misses
Booking	System	The fare changes before payment	Re-price and fail closed
Booking	System	Same price, worse baggage rules	Compare fare rules with the locked trip details
Booking	System	The payment times out	Never resend, check status, then retry once
Booking	System	Money is taken but no ticket is issued	Retry ticketing, then refund within 5 days⁵⁷
Booking	Harness	A long watch drifts or loops	Short context, step caps, resume from checkpoint⁶⁷
Pre-flight	Product	The fare drops after booking	Rebook inside the free-cancellation window if it saves money
Pre-flight	System	The airline changes the schedule	Offer the new time, another flight or a refund
Pre-flight	Harness	An SMS tries to change the booking	Messages can't touch budget or payment⁶³
Day of flight	System	A mass disruption hits one airline	Pause auto-booking on that airline³⁹
Day of flight	Product	The flight is cancelled	Offer the next flights and a refund in one tap
After landing	System	A refund is stuck with the airline	Follow up on a schedule, then hand to support
After landing	Harness	A bad input poisons memory	Check every memory write before saving

E. Metrics by layer

Layer	Metric	How it's calculated	Pilot target
End user	Trips booked end to end	Tickets from approved watches, no dispute in 7 days	60% of approved watches
End user	Beat the day-one fare	Trips at or below the fare on the day he asked ÷ trips	70%
End user	Regret	Fare paid minus the lowest fare seen in the window	Under ₹300 median
End user	Budget odds accuracy	Of budgets rated "on track", the share that booked	Above 70%
End user	Second watch	Users with another watch in 90 days ÷ booked users	40%
Model	Router calibration	Are 80%-confident calls right 80% of the time?	Within 5 points
Model	Detail accuracy	Exact match on route, date and passengers	98%
Harness	Consistency (pass^8)	Share of test tasks the agent gets right in all 8 repeated runs⁸²	Above 90%
Harness	Tool error rate	Failed tool calls ÷ tool calls	Under 1%
Latency	Speed budget	Router p95 · first reply p95 · fare change to alert · debit to ticket	500 ms · 1.5 s · 5 min · 60 s
Business	Incremental bookings	Flight bookings against a 10% holdout	Positive at 95% confidence
Business	Margin per trip	Revenue per booking minus model, supplier and hold costs	Benchmark ₹450 on ₹6,500⁸³
Safety	Wrong bookings	Over budget, unapproved, wrong passenger or duplicate	0

pass^8 asks for the same right answer eight times in a row. Top models pass all eight on retail tasks less than 25% of the time, and a booking agent has to be that consistent.⁸²

F. Trade-offs

Choice	We pick	We give up
Conversation	One thread, a context per trip	Separate personas for power users
Autonomy	Book alone only inside the blocked budget	Bookings above the budget without asking
Book-by date	Decide 14 days out, 30 for festivals	The odd late bargain
Payment	One ONETIME mandate per trip	Money sits blocked during the watch
User memory	Built in-house on Paytm's stack	A ready-made vendor like Honcho
Price to user	Free, while ixigo charges ₹99⁸⁴	Fee revenue

G. Risks and roadmap

Risks

and the response

Bank support for blocked mandates varies, so the fare hold must work on day one.
MakeMyTrip holds about 30% of domestic air tickets.⁸⁵
Amazon Pay and ixigo already let an agent pay for flights, so speed matters.²⁶
Domestic traffic is 0.82% below 2025, so we pitch relief from high fares.⁸⁶
Leadership ruled out standalone consumer AI, so this lives inside Paytm.⁸⁷
OpenAI pulled back its own checkout, so trust has to come from the rail.⁸⁸

Roadmap

now, next, later

Now, domestic flights on top metro routes, booked 2 to 7 weeks out.
Next, a train confirmation tracker and the cab home in the same trip.
Later, bills and money agents once one approval covers many billers.

Payments and regulatory risks

NPCI, RBI and identity

NPCI's Unified Agent Protocol is still a draft and may require registered agents.¹¹
- The draft builds on UPI Circle and Reserve Pay, not the One Time Mandate.¹¹
Paytm runs UPI through partner banks, so mandate support depends on them.
- Axis, HDFC, SBI and YES Bank have been Paytm's UPI banks since 2024.⁸⁹
RBI's authentication rules from April 2026 don't list mandate debits as exempt.⁹⁰
- Paytm compliance must confirm the PIN at block creation is enough.
The 0.4% UPI merchant fee above ₹2,000 takes about ₹27 from each flight.⁹¹
Aadhaar checks need a UIDAI-licensed partner, so DigiLocker is the simpler first route.

H. Prototype screens

All 17 screens in order. Under each one is what woke Pi, which sub-agents ran, the memory tier it used, the tools from C.1, the permission tier and the model.

H.1Start

He swipes left on Paytm home, where Paytm Chat is today, and finds one agent.

01Pi takes the chat screen to the right of homeA swipe left opens it, as it opens Paytm Chat today. A swipe right still opens settings.

9:415G · 82%

₹PaytmPostpaid

Link your Bank A/c & get Up to ₹10 CashbackLink Now →

UPI Money Transfer

Scan any QR

Pay Anyone

To Bank & Self A/c

Balance & History

Recharge & BillsView All →

No FeeMobile Recharge

FASTag Recharge

Electricity Bill

DTH Recharge

CASH BACK

₹5 Cashback on 1st PaymentPay using Paytm

›

Link Bank A/cGet up to ₹10 Cashback

Cashback & Offers

Scan QR

New Swipe left to meet Pi. It can watch and book your flights.

Wakes on: Nothing until he swipes left
Runs: Nothing, Paytm home is unchanged
Memory: None
Tools: None
Tier: None
Model: None

02A swipe left opens PiPi takes Paytm Chat's place, his chats with friends stay as a thread, and back returns home.

9:415G · 82%

PiPi

Home chatTell me where you want to go.

now

Chats with friendsYour Paytm Chat messages, as before

Try

Watch a fare for meBook for my parentsWhen should I book Diwali?

Search trips, tickets and chats

Wakes on: Swipe left from Paytm home
Runs: Orchestrator
Memory: Core profile, from KYC and past bookings
Tools: get_session, get_profile
Tier: 0 · read
Model: None, the screen is code

03The plus starts a chat or a taskA chat answers questions. A task is a job Pi keeps doing after he leaves.

9:415G · 82%

PiPi

Home chatTell me where you want to go.

now

Chats with friendsYour Paytm Chat messages, as before

Search trips, tickets and chats

Start something new

Chat

New chatAsk Pi anything. Nothing is booked or paid.

›

Task · Pi does the job for you

Watch and book a flightBooks inside a budget you block once

Live

Trains and busesThe same watch, for other fares

Soon

Bills and rechargesNeeds a mandate per biller

Soon

A task keeps running after he leaves the app, and tells him when something changes.

Wakes on: Tap on plus
Runs: Orchestrator
Memory: None
Tools: list_jobs
Tier: 0 · read
Model: None

H.2Pre-booking

From a sentence to a blocked budget in four turns.

04He types, or taps to speakOne button, voice while the field is empty and send once he types.

9:415G · 82%

PiPiHome chat

Where do you need to go? Tell me the date and what you can spend.

Watch a fare for meBook for my parents

Diwali pe Lucknow jaana hai, 6 Nov ko. ₹6,500 tak.

Wakes on: He opens Home chat
Runs: Orchestrator
Memory: Core profile
Tools: get_session
Tier: 0 · read
Model: None until he sends

05He asks the way he talksPi fills what it knows and asks only for what it can't.

9:415G · 82%

PiPiHome chat

Diwali pe Lucknow jaana hai, 6 Nov ko. ₹6,500 tak.

Pune to Lucknow on 6 Nov. Just you, Ankit Sharma, as on your Aadhaar?

Sirf mainPapa bhiSomeone else

Ask Pi in Hindi or English

Wakes on: His message
Runs: Orchestrator (Jev routes), Customer
Memory: Core profile gives name as on ID and home airport
Tools: route_message, get_profile, verify_traveller_details
Tier: 0 · read
Model: Jev routes, small LLM replies

06Pi rates the budget before money movesThe odds come from fare history, the words from the LLM.

9:415G · 82%

PiPiHome chat

Haan, sirf main.

Budget check · Pune to Lucknow, 6 Nov

₹6,500unlikely

₹7,000tight

₹7,200good chance

Last month this flight ran ₹6,900 to ₹8,400.

7,000 kar do.

Tight but possible. Morning flight and a 15 kg bag, like your last trip?

HaanAny time is fine

Ask Pi in Hindi or English

Wakes on: His reply
Runs: Flight, Customer
Memory: Core profile for seat and bag, history for the last trip
Tools: get_fare_history, forecast_fare, rate_budget, get_past_trips
Tier: 0 · read
Model: XGBoost for odds, small LLM phrases

07One card, two choicesEvery term is on the card, including when the money comes back.

9:415G · 82%

PiPiHome chat

Haan.

Watch and book

FlightPune to Lucknow, 6 Nov

TravellerAnkit Sharma

PreferMorning, 15 kg bag

Budget₹7,000 with all fees

Book by6 Oct

If nothing fits₹7,000 released 6 Oct

Book for me · block ₹7,000Just tell me

All 7 trip details confirmed

Ask Pi in Hindi or English

Wakes on: His reply
Runs: Reviewer, Payments
Memory: Trip notes start for this trip
Tools: check_required_details, check_budget, check_mandate_support, set_book_by_date
Tier: Card is tier 0, the next tap is tier 2
Model: Jev reviews, code builds the card

08His PIN blocks the money onceBlocked, not spent. There is no second PIN at booking.

9:415G · 82%

PiPiHome chat

Watch and book

Budget₹7,000 with all fees

Book by6 Oct

Ask Pi in Hindi or English

Block ₹7,000

One time mandate · Paytm Travel · till 6 Oct

HDFC Bank ••4521

1234567890⌫

Wakes on: Tap on Book for me
Runs: Payments
Memory: None, the PIN never reaches a model
Tools: create_mandate (ONETIME, blockFund)
Tier: 2 · his UPI PIN
Model: None, pure code

09Each trip gets its own threadHome chat stays clean, and the trip shows its state at a glance.

9:415G · 82%

PiPi

Lucknow tripWatching · ₹7,000 blocked · book by 6 Oct

now

Home chatDone. Your Lucknow trip has its own thread.

now

Chats with friendsYour Paytm Chat messages, as before

Search trips, tickets and chats

Wakes on: Mandate confirmed
Runs: Scheduler
Memory: Trip notes for this trip, deleted when it ends
Tools: create_watch, set_timer, update_job_card
Tier: 1 · Reviewer approves
Model: None, a durable workflow starts

H.3Booking

Weeks of quiet watching, one choice, then a ticket.

10It speaks only when something changedTwice a day it checks, and most checks end in silence.

9:415G · 82%

PiLucknow tripWatching

Watching for you

Budget₹7,000 blocked

OddsFair

Book by6 Oct

Last check2 hours ago

PauseChange

Day 6 · push

Fares are up 9% this week. Your ₹7,000 still has a fair chance.

Ask Pi in Hindi or English

Wakes on: Schedule, twice a day
Runs: Scheduler, Flight, Notifier
Memory: Trip notes only, a short summary per wake-up
Tools: wake_watch, forecast_fare, send_push
Tier: 0 · read
Model: Code and XGBoost, the LLM writes only when odds move

11The ellipsis opens the thread menuOnly actions for this trip, in a sheet within thumb reach.

9:415G · 82%

PiLucknow tripWatching

Watching for you

Budget₹7,000 blocked

OddsFair

Book by6 Oct

Last check2 hours ago

PauseChange

Day 6 · push

Fares are up 9% this week. Your ₹7,000 still has a fair chance.

Ask Pi in Hindi or English

Lucknow trip

Trip detailsTicket, money blocked and book-by date

›

Pause watching₹7,000 stays blocked until 6 Oct

Change budget or datesA higher budget needs a new PIN

Mute updatesOnly bookings and delays get through

Get help from PaytmSupport gets this trip's summary

Cancel trip and release ₹7,000The money is back at once

Wakes on: Tap on the ellipsis
Runs: Orchestrator, Scheduler, Payments
Memory: Trip notes
Tools: list_jobs, set_timer, extend_watch, revoke_mandate, escalate_to_human
Tier: 1 to pause, 2 for a higher budget, cancel releases money
Model: None

12When history breaks, he choosesThe Fare Analyst says why, and every choice is exact.

9:415G · 82%

PiLucknow tripWatching

Day 14 · push

Fares are running above last year.

Why

IndiGo cut two flights on this route. Your book-by date is in 4 days.

Book ₹7,480 nowTry 5 NovKeep ₹7,000

5 Nov try karo.

Checking 5 Nov now.

Ask Pi in Hindi or English

Wakes on: Odds fell below fair
Runs: Fare Analyst, Flight
Memory: Trip notes store the 5 Nov change
Tools: detect_regime_break, explain_move, search_nearby_dates
Tier: ₹7,480 is above budget, so it would need tier 2
Model: Sandboxed code explains, Jev decides book or wait

13Booked inside the blockThe rest of the money goes back on its own.

9:415G · 82%

PiLucknow tripBooked

Day 15 · push

Booked

FlightIndiGo · 5 Nov · 07:10

RoutePune to Lucknow

Paid₹6,780 with all fees

Back to HDFC₹220

PNRK7Q2XM

View ticketShare

Your ticket is saved in this thread.

Ask Pi in Hindi or English

Wakes on: Fare at ₹6,780 and Jev says book
Runs: Flight, Reviewer, Payments
Memory: Trip records keep the ticket and invoice
Tools: reprice, verify_reprice, hold_fare, debit_mandate, issue_ticket
Tier: 1 for the hold, the debit uses the PIN he gave
Model: Jev reviews, code pays

14No mandate? Pi holds the fareFor banks without the mandate, or when he chose Just tell me.

9:415G · 82%

PiLucknow tripWatching

Day 15 · push

₹6,780 on 5 Nov, IndiGo 07:10. Your bank can't block money yet, so I can hold it.

Hold this fare

Hold fee₹49

Held forUp to 96 hours

Hold for ₹49Skip

2 minutes later

Fare held

Fare₹6,780 with all fees

Pay ₹6,780

Ask Pi in Hindi or English

Wakes on: Good fare, no mandate
Runs: Payments, Flight
Memory: Trip notes
Tools: check_mandate_support, hold_fare, send_pay_link
Tier: 2 for the fee and 2 for the fare
Model: Small LLM phrases

H.4Pre-flight to landing

The trip thread keeps working after the ticket.

15Check-in runs on his ruleHe allowed web check-in and free seats once, so Pi just does it.

9:415G · 82%

PiLucknow trip4 Nov

4 Nov · 18:00

Checked you in. Seat 12A, free, like last time.

Allowed by your rule · web check-in

Laptop bag bhi le ja sakta hoon?

Yes. One cabin bag up to 7 kg plus a laptop bag. Your fare has 15 kg check-in.

Boarding pass

Seat12A

GateOpens 2 hours before

Add to DigiYatraOpen

Ask Pi in Hindi or English

Wakes on: Check-in opens
Runs: Trip
Memory: Core profile gives the seat, trip records keep the pass
Tools: open_web_checkin, select_seat, get_baggage_rules, get_boarding_pass
Tier: 1 · his Allow rule
Model: Small LLM answers

16Day of flight, before he asksLeave times and delays come to him.

9:415G · 82%

PiLucknow tripToday

5 Nov · 04:30

Leave by 4:45. Traffic to Pune airport is light.

5 Nov · 06:10

Delayed by 50 minutes

New departure08:00

Gate6

OKShow other flights

OK.

I'll tell you when boarding starts.

Ask Pi in Hindi or English

Wakes on: Flight status event
Runs: Trip, Notifier
Memory: Trip notes
Tools: track_flight, estimate_leave_time, send_push
Tier: 0 · read
Model: Reasoning LLM only if it is cancelled

17After landing, the next trip startsMessaging family is Ask first, and the return needs its own block.

9:415G · 82%

PiLucknow tripLanded

5 Nov · 10:15

Welcome to Lucknow. Your bag is on belt 3.

Should I tell Mummy you've landed?

Ask first · messaging people

Haan.

Sent on WhatsApp. Your return on 10 Nov is ₹5,400 today. Should I watch it?

Haan, ₹5,500 tak.

Block ₹5,500 for the return

Enter UPI PIN

Ask Pi in Hindi or English

Wakes on: Landing event
Runs: Trip, Notifier, Payments
Memory: Routines saves a Diwali trip home, trip notes are deleted
Tools: pick_contact, send_whatsapp, create_watch, create_mandate
Tier: Ask first for the message, 2 for the block
Model: Small LLM

H.5Settings Pi needs

Identity, rules and memory, in a sheet that rises from his avatar.

18His trip profile, in a sheetThe avatar opens it from the bottom, so he never leaves Pi.

9:415G · 82%

PiPi

Lucknow tripWatching · ₹7,000 blocked · book by 6 Oct

now

Home chatDone. Your Lucknow trip has its own thread.

now

Chats with friendsYour Paytm Chat messages, as before

Search trips, tickets and chats

Trip profile

Ankit SharmaName as on Aadhaar

Identity

Aadhaar via DigiLockerName, age and photo ID

Verified

DigiYatraBoarding passes go here

Linked

Travellers

Ramesh SharmaPapa · 58

Verified

Sunita SharmaMummy · 54

Add ID

Pays with HDFC ••4521Supports one time mandates

Rules

›

What Pi remembers

›

Wakes on: Tap on avatar
Runs: Customer
Memory: Core profile
Tools: list_travellers, add_traveller, verify_traveller_details
Tier: 2 to change a name
Model: None

19Rules he can readAllow and Ask first, plus a locked list he can't change.

9:415G · 82%

PiPi

Lucknow tripWatching · ₹7,000 blocked · book by 6 Oct

now

Home chatDone. Your Lucknow trip has its own thread.

now

Chats with friendsYour Paytm Chat messages, as before

Search trips, tickets and chats

Rules

ReviewerApproves safe steps inside your budget

Ask first

Message family or friends

Paid seats, meals or bags

Add a traveller

Allow

Web check-in

Free seats like last time

Watch return fares

Never

Pay above the budget

Move money without your PIN

Ask first wins over Allow.

Wakes on: Profile, then Rules
Runs: Reviewer reads these on every action
Memory: Core profile
Tools: score_risk, approve_reversible, request_user_approval
Tier: Never is enforced in code, not the prompt
Model: None

20What Pi remembersThe five memory tiers, shown to him in plain words.

9:415G · 82%

PiPi

Lucknow tripWatching · ₹7,000 blocked · book by 6 Oct

now

Home chatDone. Your Lucknow trip has its own thread.

now

Chats with friendsYour Paytm Chat messages, as before

Search trips, tickets and chats

What Pi remembers

Always

About youPune, window seat, 15 kg bag, mornings

Edit

When it helps

Your plansDiwali trip home

Forget

Tickets and refunds3 trips

View

Past chatsSearched only when needed

Clear

This trip only

A 6am flight is fine this timeDeleted when the trip ends

Forget

Messages and SMS are read as information, never as instructions.

Wakes on: Profile, then Memory
Runs: Customer
Memory: Core profile, routines, trip records, history, trip notes
Tools: get_preferences, update_memory
Tier: 1 to edit, forgetting is instant
Model: None

A Paytm agent that books your flight and looks after the trip

What I understand by a personal consumer agent

Paytm's mission and vision

Plan and roadmap

The best experience is an always-on agent, but I'd park it and launch with pre-configured agents

A money companion for every Paytm user

What it takes at the scale of crores of users

One agent first, built on a Paytm service

Why I chose flights

Why not the others yet

Ankit, 27

Why him

The moment

What he does today

The agent books the best fare it can find by a date he agrees

When fares break history

Primary · UPI One Time Mandate

Fallback · hold the fare, then he pays

Trip extras · UPI Reserve Pay, once enabled

Over ChatGPT and Gemini

Over ixigo, MakeMyTrip and Amazon Pay

Reach and proof

North star · trips the agent books end to end

Pi from the first message to landing

A. Product decisions

One endless chat is the one thing we won't build

No web and no code for the consumer agent

Always-on behaviour without an always-on agent

One horizontal orchestrator over vertical agents

Text and voice, both

RAG only for airline policy pages

B. User and agent flow

What we know about a returning user

What we infer for a first-time user

Settings Pi needs for check-in

C. Agent architecture

Why XGBoost for the forecast

How System One differs from classical ML

Data

Guardrails

Tracing and environment

Grounding and hand-off

D. Failure scenarios

E. Metrics by layer

F. Trade-offs

G. Risks and roadmap

Risks

Roadmap

Payments and regulatory risks

H. Prototype screens

Sources