Case study: Pi Business Ads
Pi Business Ads lets a shop owner try AI ad creation without a designer or a brief. Seven answers in a chat produce a Facebook ad preview, sent to them on WhatsApp. I designed the flow from the landing page to the delivery states.
Featured atGlobal Fintech Fest 2026Mumbai. This year’s theme: Potential to Impact.
Most small businesses in India do not advertise digitally. In one survey of 3,249 MSMEs, only 13% used digital marketing.1 A report cited by WION says over 70% of SMBs find the range of ad options overwhelming.2
A chat of seven questions. Three answers are taps and four are typed. Pi works out the reason to run the ad from the answers, shows what it will use, then returns the ad as a Facebook post and sends it on WhatsApp.
I owned the conversation flow, the screens and the states for waiting, success and failure.
A business owner with no design skill should be able to see what an AI-made ad for their shop looks like, in about a minute. The same visit gives the company a lead. Pi Business Ads is built as a guest demo: no sign-up, one chat, one result.
A shop owner who has never made an ad faces two blank pages: a tool with many options, and a brief they do not know how to write. They need an ad that looks professional and carries their shop’s name, without learning the vocabulary of advertising.
I combined published surveys on small businesses in India with research on chat interfaces and a look at existing AI ad tools. This is secondary research. Where a figure comes from a company or an unnamed report, I say so, and where two figures disagree, I say that too.
Share of MSMEs that currently use digital marketing.
Google and India SME Forum survey of 3,249 MSMEs, June 2026. 72% were micro enterprises1Share of SMBs who say the variety of ad options is overwhelming.
A report cited by WION. The article names no publisher or sample2Share of businesses on Meta platforms in India that are MSMEs.
Meta, December 2025. A company claim7TechCrunch reported in 2024 that India has more than 65 million small and medium businesses, and about 5 million of them sell online.3 The 13% survey and the WION report’s claim that nearly 80% of SMBs use AI ad tools do not agree. The survey names its sample and leans towards micro firms in smaller cities. The other source names neither. I rely on the survey.
Kirana stores’ share of the grocery market in 2025, projected to be about 86% by 2030.
RedSeer, January 20264Estimated number of kirana shops. Estimates vary widely.
CB Insights, 20225Share of internet users who accessed content in Indic languages, among 886 million users in 2024.
IAMAI and Kantar6Preferred a chatbot over an online form for a health questionnaire. The chatbot took longer.
Frontiers in Digital Health, 2022. 206 analysed. Not shop owners8Net Promoter Score for the chatbot against the form. Usability scores were close: 69.7 against 67.7, not a significant gap.
Same study8Options a chat can list before choosing becomes unworkable, in one team’s view. They also note about 2 to 3 seconds of delay per AI reply.
UX Planet, SAP Emarsys authors, from their own testing9If over 70% of SMBs find ad options overwhelming, the first screen should not show any. The flow asks one question at a time and shows no ad option until the result.
Most businesses on Meta platforms in India are MSMEs, so the ad appears as a Facebook post. WhatsApp carries it back to the owner, because it is the channel they already use with customers.
People preferred the chatbot but took longer with it. So the chat stays at seven questions, and three of them are taps.
Business type, city and colour each set a rule set, a line of ad copy or a colour code. They are cards and swatches, not free text, so the answer always lands on something valid.
With 98% of internet users reading Indic languages, an English-only chat is a real gap. I list it in the review. The ad prompt already supports Hindi and Hinglish.
These descriptions come from vendor help pages and press coverage. I did not test the tools, and Meta’s features were described in 2024 and may have changed.
| Tool | What you start with | What it makes | What you do next |
|---|---|---|---|
| Meta Advantage+ creative10 | An existing ad | Image versions inspired by the original, text overlays, headline and primary-text variations | Choose among variations in Ads Manager |
| Google Performance Max asset generation11 | A landing page URL | Headlines, descriptions and images, from a product summary drafted from the page | Edit the summary, keep or discard assets, generate more |
| Canva Magic Design12 | A text prompt, or an uploaded image or logo | A range of matching layouts | Customise the layout you pick |
| Pi Business Ads | Seven answers in a chat | One finished ad, shown as a Facebook post | Receive it on WhatsApp |
Each of the other three assumes something a first-time owner may not have: an existing ad, a website, or a prompt worth writing. That is my reading of the pages above.
One article lists specific ways a chat falls short of a form. It is the authors’ own view from their testing, with no sample sizes, so I used it as a review checklist and not as proof.9
| Published point | What this flow does |
|---|---|
| Chat hides progress, so users cannot see how much is left | Not addressed yet. Nothing shows how far along the owner is. See the review |
| Chat is hard to review, because key inputs scroll out of view | A summary card lists every answer before the ad is made |
| Option lists get unworkable above three to five items | Every closed question offers four options |
| Each AI reply adds a delay, and users need to see system status | “Getting things ready for you…” with the input disabled, then a skeleton of the post while the ad is made |
| Recognition beats recall | Cards carry icons and examples, and the description field has an example and a suggestion button |
These three segments are hypotheses built from the product spec and the research above. I would confirm them in interviews.
I organised the chat by what the owner has to tell Pi. Each answer exists because something later needs it, and a few things the ad needs are worked out instead of asked.
Worked out, never asked: the compliance rule set from the business type, the reason to run the ad and who to aim it at from type and city, and the mood of the photograph from the chosen colour.
From the product’s field spec, which lists what each answer feeds.13
| Answer | How it is asked | What it feeds |
|---|---|---|
| Name | Free text. “What should I call you?” | The chat, which uses the first name once, and the lead record |
| Business type | Four cards, each with examples | The compliance rule set, the brand type line, the colours offered, and the reason to run the ad |
| Business name | Free text. “I’ll use this as the brand name on your ad.” | The logo plate, the ad copy, and the page name in the feed preview |
| City | Four cards, each with its state | The contact bar, the ad copy, and the targeting line |
| Description | Free text, with an example and a suggestion button | The ad copy and the scene in the photograph. The biggest lever on quality |
| Theme colour | Four swatches named for the trade | The colour of the contact bar and headline accent, and a mood phrase that steers the photograph |
| Mobile number | Free text, ten digits | The contact bar on the ad, the lead record, and where the ad is sent on WhatsApp |
There is no offer or discount question, no logo or product photo upload, no street address, and no language choice. The spec cut the offer to keep the chat short, and notes that the offer badge is the loudest element on a poster, so it may be worth adding back.13 I list it in the review.
Work out the rest from type, city and colour.
Cards and swatches, never free text for type, city or colour.
Each question carries one line on how the answer will be used.
Show every answer, then make the ad.
The map follows a kirana owner through seven steps, from the landing page to the ad on WhatsApp. It is built from the flow and the research, and the feelings are hypotheses to test, not measurements.
Swipe sideways to see all seven steps.
| Row | Step 1Arrive | Step 2Introduce | Step 3Describe the business | Step 4Describe the look | Step 5Give a number | Step 6Confirm | Step 7Receive |
|---|---|---|---|---|---|---|---|
| Goal | See what this is | Get started without effort | Say what kind of shop it is | Make the ad look like my shop | Let customers reach me | Be sure it is right | See my ad and keep it |
| What they do | Read the promise and tap Let’s get started | Type a name | Tap a type and a city, type the shop’s name | Describe what the shop sells, tap a colour | Type a ten-digit number | Read the summary and tap Create my Ad | Wait, view the preview, tap Send Message |
| Screens | Landing | Name, working | Type, name, city | Description, keyboard, colour | Number | Summary | Generating, ad ready, sent, failed |
| How they feel |
Solid green line: the ad arrivesDashed red line: delivery fails |
||||||
| Risks | Not believing an AI can make a usable ad | Wondering why a name is needed | A business that fits none of the four types | Not knowing what to write, and giving a vague line | Not knowing what the number will be used for | A wrong answer that cannot be changed | A failed message, with no clear next step |
| Design response | Four sample ads and one button | One question, and the name used warmly afterwards | Cards with examples, and each question echoing the last answer | An example, a “Suggest me a description” button, and swatches for the trade | A line saying the number goes on the ad so customers can call | A summary of every answer before the ad is made | A skeleton while waiting, then a Send Message button and a retry |
| How I would measure it | Landing to first answer rate | Drop-off at the name | Drop-off by question, “none of these” reports | Description length, suggestion use | Invalid-number rate, drop-off at the number | Back or edit use, Create my Ad rate | Delivery success, resend and create-new rates |
Typed answers: name, business name, description and mobile number.
Tapped answers: business type, city and colour.
Optional shortcut: a suggested description the owner can accept and edit.
Every screen in the file, in order. Scroll sideways.
Why. An owner cannot picture an AI-made ad from a description. The landing page shows four sample ads, one for each supported trade, and a WhatsApp badge that hints where the ad will arrive. One button starts the chat. The first question says what is coming: “Share few business details and I’ll create an ad for your business.”
Trade-off. Four samples tell an owner of any other trade that the tool is not for them. See review finding 2.
Why. Business type, city and colour each map to a rule set, a line of copy or a colour code, and free text would not resolve. Each card carries examples, such as “Dine-in, Take Away, Cafes”, so a bakery owner can see where they belong. Every closed question has four options, within the range a chat can handle.9 The chosen card keeps its outline and the rest fade.
Trade-off. A business outside the four types has no route, and the spec warns against forcing it into the nearest box.13
Why. After the owner picks Kirana Store, the next question reads “What’s your Kirana Store called?” After they type a name, it becomes “Where is Laxmi General Store located?” Each question adds one line on its use, such as “I’ll use this as the brand name on your ad.” The chat feels like it is listening, and each ask is justified.
Trade-off. Echoing the type only works because the type is asked first. The order of questions is part of the design.
Why. The description is the only source of anything specific in the ad, and it shapes the photograph too. The spec’s own example makes the point: “we do haircuts” gives a generic ad, while brands, packages and two senior stylists give a specific one.13 So the question names what to cover, the field shows an example, and a “Suggest me a description” button gives something to accept and edit. The field grows to two lines as the owner types.
Trade-off. An owner who accepts the suggestion unchanged gets a generic ad. The spec says to push for two sentences and ask again if the answer is one line.13
Why. A wrong business type applies the wrong rule set, and the phone number is printed on the ad. The number is asked last, with a line saying it goes on the ad so customers can reach the shop. Then one card lists every answer, with a single Create my Ad button. The summary also answers a known weakness of chat: key inputs scroll out of view.9
Trade-off. The card shows values but no way to change one. See review finding 1.
Why. While the ad is made, a skeleton in the shape of a Facebook post stands in for it, with “Reading your Business Details” above. The result appears in that same frame, with the shop’s name, “Sponsored”, the copy, the picture and a button, so the owner judges it the way a customer would. “Your ad has been sent on WhatsApp” tells them where to go next, with Send Message to try again and Create New Ad to start over.
Trade-off. The page promises the ad was sent before the owner knows it arrived.
The file covers the happy path well. It shows seven waiting, input and delivery states, and no error for any answer. I list both so the gap is plain.
Nothing shows how far along the owner is, and each answer leaves the screen as the chat moves on. Add a small “3 of 7” marker, and let the owner tap an earlier answer to edit it. These are two of the published weak points of chat.9 I read this from the frames, so I would confirm it in the build.
The frames offer Pune, New Delhi, Mumbai and Kolkata. The spec lists Mumbai, Delhi, Bengaluru and Pune, so the two need aligning. Either way, an owner elsewhere has no route, so add “Other city”. The chat is in English while 98% of internet users read Indic languages.6 The ad prompt already supports Hindi and Hinglish, so generation is ready. The chat and a language choice at the start still need design.13
| Metric | What it tells us |
|---|---|
| Landing to first answer rate | Whether the promise and the sample ads get owners into the chat |
| Completion rate | Whether seven questions is short enough to finish |
| Drop-off by question | Which question costs the most owners, likely the description and the number |
| Description length, and use of the suggestion | Whether owners write enough for a specific ad, and whether the suggestion replaces their own words |
| Summary edits and back taps | Whether the summary catches mistakes |
| Time from Create my Ad to Ad ready | Whether the wait fits the minute the spec expects |
| WhatsApp delivery and resend rate | Whether the ad reaches the owner |
| Create New Ad rate | Whether owners want to try another |
| Leads captured | The commercial output of the demo |
Moderated sessions with five to eight owners per trade, on the devices they use. Tasks:
I would record whether each task is finished without help, time on task, what owners expect to receive, and where they hesitate or misread a screen.
Pages read in October 2026. Company figures and unnamed reports are claims, not independent studies.