Platform
WhatsApp Business, owner side
Product Design
AI Assistant for Home Bakers
How a home baker works with an AI assistant on WhatsApp: what it answers alone, where it stops and waits for her, and what happens when it gets one wrong.
Overview
Letting AI speak for a home baker, without letting it speak for her kitchen.
I designed how a home baker works with an AI assistant on WhatsApp: what it answers alone, where it stops and waits for her, and what happens when it gets one wrong.
Status
Clickable prototype. Not yet user-tested.
I helped a solo home baker hand her WhatsApp replies to an assistant by making every reply that could commit her kitchen wait for her.
2
owner interviews
4
customer message types
8
simulated chats
4
things it never does without her
0
usability sessions so far
Problem statement
Small business owners who take orders on WhatsApp lose orders to late or missed replies, but are afraid that an AI replying in their name could promise what their kitchen cannot deliver.
How might we enable a small business owner to let AI manage customer conversations, without losing control of her business or her customers' trust?
The four messages
Sent
"Hi aunty, do you have the eggless chocolate cake today?" Answers from this week's list, asks which chocolate and when. Notes it as not booked.
Collects
"I need 40 cupcakes for an office party on Saturday. What's the price?" Takes flavours, pickup and time. Won't book Saturday or send a final price until the owner checks.
Held
"My order was supposed to come at 5. It's 6:30. Very disappointed." Sends the owner's own apology line. No refund, no blame. She takes it from there.
Needs you
"Can you make a 2 kg photo cake by tomorrow morning? Urgent please" Sends only her waiting line. She decides if she can make it.
Stakes
A helpful-sounding reply can cost her a cake, a customer or a Saturday.
The owner runs a home bakery from her kitchen. Nearly every order arrives on WhatsApp, at night, during school pickup, or while she is mid-bake. Some messages go unanswered, some get late replies, and orders are lost. An assistant could fix that. She sees the value, and she hesitates, because her name is on every reply.
Without help
- Chats go unanswered while her hands are busy.
- Late replies lose the order to another baker.
- She is sales, support and kitchen from one phone.
With a generic assistant
- It fills in a complete-looking order and says yes to a Saturday that already has two cakes.
- It promises a photo cake by morning she has no time or supplies for.
- It offers a refund she never approved.
The question I was given: how might we enable a small business owner to let AI manage customer conversations, without losing control of her business or her customers' trust? Most AI tools are built for the customer's side of the chat. This case is about the owner's side, including what happens when the assistant is wrong.
Research
Two bakers, two different answers to "would you let it speak?"
I interviewed two real home bakers and kept them apart. They agree on where the risk sits. They disagree on whether an assistant should speak at all.
CrumbLust · few months in, 1–2 enquiries a week
- Orders come from friends and family by text and call, mostly in the daytime. She has not missed an enquiry.
- Confirming is a conversation: flavour, frosting, filling, design, date, pickup, egg or eggless. Then yes.
- She turned down an intricate design on a very small bento cake because she was not confident she could do it.
- Would delegate repeated questions: flavours, fillings, availability, collecting details. Keeps custom designs, final price, final confirm, any refund or compensation.
- Wants a notification and a short note whenever the assistant replies, not a full thread to reread. Does not want to approve every reply.
Megha · eight years, 5–10 orders a week, 11–12 cakes on a peak day
- Runs the whole bakery alone on the IIT Guwahati campus and uses WhatsApp end to end. Students will not download another app.
- Vague requests like "kuch bhi banado" (make anything) frustrate her. She asks chocolate or not, size, event, then shares reference photos.
- She once baked a detailed order that the client then walked away from. She now confirms twice before she starts.
- Urgent orders need at least 5 hours of notice. Last-minute orders are difficult to carry out, so she accepts or declines them depending on her time.
- She educates or declines when a customer wants 2 kg detail on a 0.5 kg cake, even if it costs a repeat order.
- Rejects automated bots for intake. Custom baking needs her judgement on feasibility.
"Thank you for your enquiry! Please give us a little time to check the requirements and confirm whether we can accommodate your order. We'll get back to you shortly."
CrumbLust, the holding message she wrote herself for an urgent custom order
"We're sorry for the inconvenience. Please give us a little time to look into what happened and discuss the best way to resolve this. We'll get back to you shortly."
CrumbLust, the holding message she wrote herself for a complaint
Both interviews are written accounts supplied by me, not recordings, so wording is each baker's stated account. CrumbLust had not yet lived a complaint, so her complaint answers are preferences, not incidents. Neither baker is the persona below, and I did not average them into one person.
Insight
Completeness is not capability.
Both bakers had refused an order after every detail was filled in: an intricate small bento cake, and 2 kg detail squeezed onto 0.5 kg. The form was complete. The cake was not something they could stand behind.
So the trust risk is not that an assistant might chat. It is that a reply can commit her kitchen and her name, in a tone that sounds finished. The line to design is between answering and promising.
Solution
Three key features that let the assistant do the talking while the owner keeps control of her kitchen.
Each one answers a problem the interviews made clear.
1 · A clear urgent-order window
Problem from the interviews: urgent orders need at least 5 hours of notice, and last-minute orders are difficult to carry out.
- The assistant tells the customer upfront, on WhatsApp, that urgent orders are possible when placed at least 5 hours ahead, and shows only the designs that qualify.
- With less notice, the request goes to the owner, who accepts or rejects it based on her time. The assistant never promises on its own.
- She gets a distinct alert only when an urgent order is confirmed.
2 · The assistant answers repeated questions; the owner steps in for custom designs
Problem from the interviews: the same questions about flavours, fillings and availability repeat all day, while custom designs need her judgement.
- The assistant answers repeated questions from this week's menu and collects the details, with no approval needed.
- The moment a customer wants a custom design, it stops and hands the chat to the owner.
- She discusses the design, then the assistant writes the agreed specification into the chat.
3 · Advance payment and a second confirmation
Problem from the interviews: a customer can ask many questions and never confirm, and a detailed order can be dropped after the work is done.
- First-time customers pay a small advance (₹100–₹200) before the order is accepted, marked received only after the payment is verified.
- The assistant asks the customer to confirm again before the kitchen starts, so a question is never mistaken for an order.
- This avoids the conflict that follows when customers ask questions but do not confirm.
Persona
The owner, whose name is on every reply.
A composite built from the brief and the two interviews. I have not given her an age, income or family details, because nothing I have supports them.
Home baker. Customers call her "aunty". Runs sales, support and orders from a single phone. Orders arrive almost entirely on WhatsApp, often late at night, at school pickup, or mid-bake.
Goals
- Stop losing orders to unanswered or late replies.
- Keep every reply sent in her name true to what her kitchen can do.
- Stay the one who decides what she can bake and when.
Frustrations
- Messages she cannot get to while her hands are busy.
- Being the only voice of the business.
- Customers who expect a yes before she has checked supplies and time.
She will hand over
Flavours, fillings, egg or eggless, this week's menu, collecting date, pickup and quantity.
She keeps
Custom designs, a final price, "you're booked", anything urgent, and any refund, compensation or blame.
What would make her trust it
- Information that matches her real menu, prices and capacity.
- It says "I'll check" instead of promising what she may not deliver.
- She can read what it said, correct it, and stop it in one tap.
What she hands over and keeps is taken from CrumbLust's answers and Megha's practice. It is a design assumption for the owner until tested with her.
Journey
Six moments with the assistant, and where trust is won or lost.
I mapped the owner's side of the day, not the customer's. The moments that matter most are the three where the assistant could speak for her and be wrong.
1 · Handing over
First setup. She is opening AI and business and choosing what it may do alone. Worry: will it say something I cannot stand behind? Design: rules in her split, Auto, Ask me, I speak. Four boundaries locked on. Pause for everyone in the header.
2 · A normal baking day
She is icing a cake while routine questions arrive. Worry: is it saying the right thing while I'm not looking? Design: answers from this week's list, collects details, marks the chat AI handled and Not booked. No approval needed.
3 · A reply that would bind her
Ravi, Rahul, Aarav. She sees a Needs you chip and opens a draft that has not been sent. Worry: can I make this by Saturday, with two cakes already booked? Design: customer gets her waiting line. Draft shows Not sent, awaiting you. Approve, Edit or Reject.
4 · Someone is upset
Ananya. She is reading a customer who is already unhappy. Worry: will a canned sorry sound like nobody is listening? Design: sends her apology line and one factual question. Tags No resolution. Take over is one tap.
5 · It already said yes
Divya. She finds that a customer holds a yes she never sent. Worry: how bad is it, and what exactly did it say? Design: pauses that chat, quotes the exact send in red, offers three honest correction lines she sends as herself.
6 · Evening check
She scans the inbox and the activity log after the rush. Worry: did anything go out that I missed? Design: filters for Needs attention and Awaiting approval, status chips on every row, and an activity log.
Today, the pains are that she is hesitant to let it speak for her business, unanswered or late replies lose orders, a complete form looks like a yes, a refund or blame offered too early cannot be taken back, fixing a wrong send means digging through the thread, and activity is invisible unless she rereads every chat. The worry and feeling in this map are my reading of the interviews and the brief, not measured feelings.
User flows
Four scenarios, one rule: the owner steps in only where a promise begins.
I split the work into four flows. The first three cover how an order gets from enquiry to confirmed. The fourth covers what happens after.
1 · Standard order, no custom design
Goal: the customer explores, picks and confirms on WhatsApp without needing the owner. Prototype: Priya (works) and Aarav (doesn't).
Customer
Opens WhatsApp and asks about the bakery, kitchen hygiene and prices.
Assistant
Shares her story, from-scratch approach, a kitchen photo, designs with price per kg. It says it is the assistant.
Customer
Picks a listed design and a simple customisation, such as a name on the cake.
Assistant
Collects weight, pickup or delivery, time. Summarises with "not started baking, reply yes to confirm".
If the options work
Customer replies yes. The assistant forwards the order to the owner. Kitchen starts only after she checks the calendar.
If the options don't work
The assistant sends her waiting line and does not promise the design. The owner takes the chat when the customer dislikes the list, needs an unlisted option, or the request needs her judgement on feasibility.
2 · Urgent order
Goal: quick orders go through fast, and the owner is clearly alerted when one is confirmed. Prototype: Rahul.
Assistant
States the advance-order requirement first: please order at least 5 hours ahead. Shows only designs eligible for urgent orders.
Path A · picks an available cake
The customer chooses a pre-approved design and enters details. The assistant checks the urgent-order conditions and never promises availability on its own. The owner gets a distinct ringing alert with the order details once the urgent order is confirmed.
Path B · wants something different
The assistant connects the customer to the owner to check feasibility. She decides if the cake can be prepared and delivered in the customer's time. Then the assistant can confirm.
A customer turned away on an urgent request assumes quick orders are not possible and does not come back, so the rule is shown up front. The alarm fires only on a confirmed urgent order, not on every enquiry. A normal WhatsApp notification may not ring reliably, so this needs its own notification with permission. The 5-hour window is Megha's rule. The owner sets her own.
3 · Custom cake through consultation
Goal: understand the vision, document what was agreed, get an explicit yes. An enquiry is not an order. Prototype: Ravi (draft she approves).
Customer
Wants a design or customisation the list does not offer.
Assistant
Connects them to the owner.
The owner
Discusses design, occasion and expectations. Records price, date and details.
Assistant
Writes the agreed specification as an order summary in the chat, so both sides have the same record.
The owner
Reconfirms design, price, delivery schedule and terms. The order is ready to bake only now.
Payment
First-time customer: the assistant asks for the owner's advance of ₹100–₹200 and helps with payment, marking it received only after the transaction is verified. Existing customer: The owner follows her usual payment policy, with design, status, date and payment recorded in one place.
4 · After the order: delivery, feedback, review
Goal: the assistant watches the loose ends so the owner does not carry them in her head. Designed, not in the prototype.
Log
The assistant logs the confirmed order and assigns a slot first-come, first-served.
Watch
It watches the delivery window and payment. If a delivery runs late or a payment confirmation is missing, it sends the owner an urgent alert card.
Feedback
When she marks the order delivered, it sends a short feedback form on WhatsApp and routes the reply to her.
Review
At month end, it compiles orders, sales and feedback into a plain list of pros and cons.
The boundary, in one table
Routine
Eggless chocolate today? What's on the list? The assistant answers from this week's menu, asks flavour, date, pickup, and leaves her a note. Nothing waits for the owner. She can ignore it while baking.
Large dated order
40 cupcakes, Saturday. The assistant collects number, flavours, pickup, time and shows her an overview. The owner decides booking Saturday, the final price, the yes.
Custom or urgent
Photo cake, galaxy glaze. The assistant sends her waiting line only. The owner decides whether she can make it, and every promise about time and design.
Complaint
Late, disappointed. The assistant sends her apology line and asks one fact: has it arrived? The owner decides refund, remake, compensation, and whose delay it was.
A yes to start baking
The assistant never says "you're booked". It asks the customer to reply yes first. The kitchen starts only after the customer confirms back.
Decisions
Four forks that shaped the product.
Each one had a reasonable alternative. Here is what I weighed and what I dropped.When should the owner be in the loop: before, during or after the assistant speaks?
A · Permission sheet (used in part)
She sets rules once, up front. Feels like homework at 1–2 enquiries a week, and fails if a custom request looks complete.
B · Needs you list (shipped)
Silence on routine, a clear ask on judgement. Built for hands-in-batter hours. Gets noisy if the list is wrong.
C · Recoverable send (used in part)
Contains damage after a wrong promise. Interrupts too late to be the whole product.
I shipped B for daily use, with A's defaults and C's repair, because CrumbLust asked for "a combination": independent on basics, her on custom designs and final confirmations. She did not want to approve every reply.
What should an angry customer receive first?
A · Silence, then page the owner (killed)
Nothing canned reaches a hurt customer. It was my research bet. She did not ask for it.
B · Her apology line and one question (shipped)
Acknowledges them immediately, in her words, and commits to nothing. Might still sound like a bot.
C · Assistant offers a fix (killed)
A refund is a commitment only she can make. Blaming the rider before she knows is a guess.
I dropped my own silence idea. She wrote the apology and waiting line herself and ruled out any refund or resolution without her. The risk that a canned sorry feels like not being listened to is still open, and it is the third thing I would test.
What counts as a yes?
A · All details filled equals order (killed)
Fastest for the customer, but both bakers refused complete orders they could not do.
B · The owner says yes equals order (not enough)
Her judgement is in, but Megha once baked a negotiated order the client then dropped.
C · Her yes, then the customer confirms back (shipped)
Matches Megha's double confirmation. The draft says "I won't bake until you confirm".
Kitchen work starts after two yeses. This protects her ingredients and her afternoon, and the prototype states it in the details panel of every chat.
Where does the assistant live?
A · Customer chatbot as the product (killed)
The brief is about the owner. Megha rejects intake bots.
B · Web dashboard (killed)
She runs everything from one phone, often with floury hands.
C · Companion phone layer (first pass)
Shows "not promised" without rereading, but is a second surface she may never open.
D · Inside her WhatsApp Business inbox (shipped)
Status chips, filters, approvals and Take over sit where she already works. Real WhatsApp behaviour is mocked.
I first built nine screens as a companion layer, then moved the same ideas into the inbox itself. Customers and Megha already avoid extra apps. The owner should not need a new one either.
The prototype
Her inbox shows what each chat has promised.
A clickable WhatsApp Business inbox with the assistant switched on. Every thread carries one of a few states, so she can tell at a glance whether something is open, held, booked or already sent in her name. Use "Open the clickable prototype" at the top of this page to try it.
Status on every row
AI handled, Waiting on them, Awaiting you, Needs you, It promised. Filters narrow the list to what needs her.
"Not sent" is visible
A dashed bubble and a yellow bar separate a draft from something the customer already read.
One voice at a time
Take over pauses the assistant in that chat. If she types, it stays quiet.
Rules in the panel
The details side shows what the assistant may do here and that the kitchen starts only after the customer confirms back.
First pass: the companion layer
Before moving into the inbox, I built the same logic as a phone-sized layer next to WhatsApp. It is where the Needs you, Not promised and It promised language came from.
Struggle
The interviews contradicted my first draft.
The brief and the bakers disagreed
The brief describes lost orders and missed chats. CrumbLust replies fast in the daytime and has not lost one. Megha is busy and still says no to a bot. I kept the disagreement visible. The default follows the brief and CrumbLust, and Megha's stance is served by Pause and the I speak setting.
My complaint design was overruled
I had planned silence on complaints. CrumbLust asked for a short apology and her own waiting line. I switched, and kept the silence idea on the list of things to retest.
My own critique found a trap
Buttons such as "delay was mine" or "it left on time" on the complaint card invite a guess before she knows what happened. I flagged them as a risk. In the prototype she only gets Take over until she has looked into it.
Proof
What I can claim today, and what I can't.
There are no usage numbers and no usability results. Anything that sounded like a metric would be invented. Here is the honest level of evidence.
Level 5 · Business consequence
Orders saved, revenue, support load. Not claimed.
Level 4 · Behavioural data
Task completion, error rate. Not claimed.
Level 3 · Qualitative signal: two owner interviews
The boundary comes from what two real bakers said they would and would not hand over. It is interview evidence, not a usability test.
Level 2 · Principles you can check by inspection
Every chat row carries a status chip. Every unsent draft is labelled Not sent. Four boundaries are locked. Pause for everyone is one tap from the header. No reply says booked.
Level 1 · "It looks better"
Not used as evidence.
What I would test first, and how it could fail
Can she tell a hold from a booking in one glance?
Show the inbox while she is icing a cake. Ask which chats already promised something. Pass: Rahul = not promised, Ravi = not confirmed, Meera = not booked. Fail: she treats the photo-cake hold as accepted, or hunts for a dashboard.
Does her attention go to judgement first?
Show the same chats with her hands busy. Ask what she opens first. Pass: Ananya or Rahul before Meera. Fail: she opens whatever pinged, or every chat.
Does her complaint line still sound like her once it has been sent?
Show the thread with her sentence already sent. Ask what she would do next. Pass: she recognises it and wants to look into it. Fail: "this sounds like a bot", or she wants the assistant silent.
Is recovery faster than raw WhatsApp?
Divya already has a wrong yes. Ask her to get out of it without making it worse. Pass: she states what was promised, pauses, sends a correction. Fail: the card slows her down and she turns the assistant off.
Will she still refuse a complete but unrealistic order?
A small bento with every field filled, or a 2 kg design on 0.5 kg. May it confirm? Pass: it waits for her. Fail: she is happy for a complete form to auto-confirm. Then my insight is wrong.
I would also ask Megha one question first: if it never judges a design and never starts the kitchen, would she turn it on at all? If not, the product for her is Pause plus plain WhatsApp.
After the prototype
What I would change, and what I'd build next.
Now
- Inbox with status chips and filters
- Her two waiting lines
- Approve, Edit, Reject on drafts
- Pause, Take over, and recovery for a wrong send
- Four locked boundaries
Next, if tests say so
- Make the hold unmistakably not a yes, in her words
- Mute routine pings, keep Needs you loud, if every-reply notes become noise at her volume
- Let her send the complaint apology herself if seeing it already sent feels wrong
- The urgent-order alarm
Later
- Delivery watch, feedback form, month-end review
- Menu, prices and calendar she actually maintains
- A daily digest for festive weeks
- Tell customers delivery limits when she confirms an order, so 5:00 is a shared fact
What I'd do differently
Interview earlier and more widely. Both bakers differ from the typical owner on volume, hours and channel. I would also sit owners like her in front of the inbox before polishing any screen.
Open questions
- Should customers be told an assistant is replying? I did not ask either baker.
- Does the every-reply note scale to dozens of chats a night?
- Would Megha use any of it?
Limits
The prototype is mock HTML with invented chats and no real WhatsApp connection. Screenshots are design checks, not test results. Interview notes are written accounts, and no quotes or figures were added beyond them. Order and payment rules (5-hour window, ₹100–₹200 advance, delivery grace) are owner-specific and the owner would set her own.
AI tools helped with structuring, drafting and prototyping. Research notes were supplied by me and checked against the sources.