ai agents for business: what to build on whatsapp, website and instagram, and the guardrails that keep them honest
puneetraj chavanfounder, rapchik media22 oct 2026 · 11 min read
an ai agent is a sales process written down precisely enough for software to run the first half of it. here is what to build, which meta option fits, how to score leads, what to feed it and where agents fail.
in this article8 sections
- 01what are ai agents for business in india, and how are they different from chatbots?
- 02meta business agent on whatsapp: use meta's, a provider's bot or a custom agent?
- 03ai chatbot for website or whatsapp: where do indian buyers actually talk?
- 04how does ai lead qualification decide who sales should call first?
- 05how do you train an ai chatbot on your data without leaking it?
- 06ai chatbot mistakes: where do agents fail, and which guardrails stop them?
- 07what should a 30-day ai agent pilot look like?
- 08who wrote this, how and why
- this is the agents chapter of our guide to ai automation for business in india, which ranks what to automate first. voice has its own chapter in our ai voice agent guide.
- meta now offers its own agent: business ai went live in the whatsapp business app in india in may 2026, and the meta business agent launched globally in june 2026 across whatsapp, messenger and instagram (meta).
- meta, citing a 2025 kantar study, says 91% of online adults in india chat with a business every week (meta).
- a canadian tribunal held air canada liable for its chatbot's wrong refund advice in 2024. your bot's answer is your answer (aba business law today).
- for sports for life, our whatsapp ai agent delivered 5x faster first response and 200+ trial bookings in the first month (case study).
what are ai agents for business in india, and how are they different from chatbots?
a chatbot replies. an ai agent replies and then acts: it books the slot, scores the lead, writes to the crm or calls a human. automation acts too, but only on fixed triggers. most useful business builds combine all three, with buttons for fixed paths, a model for messy text and rules deciding what happens next.
the word "agent" is stretched by every vendor right now. here is how we use the terms when scoping a build.
| type | what it does | example | fails when |
|---|---|---|---|
| automation | runs a fixed step when a trigger fires | form submitted, lead goes to the crm and a reminder is scheduled | the input doesn't match the rule |
| rule-based chatbot | follows a scripted menu of buttons | "pick 2 bhk or 3 bhk" | the customer types something off-script |
| ai chatbot | understands free text and drafts a reply | answers "is parking included?" from a facts sheet | it has no facts sheet and guesses |
| ai agent | understands, decides and takes an action through tools | reads a hinglish message, books a trial, updates leadsquared, alerts sales | nobody defined when it must stop and hand over |
in microsoft's 2025 work trend index, 93% of indian leaders said they intend to use ai agents to extend their workforce in the next 12 to 18 months (microsoft, august 2025). intent is not a working build.
where ai agents earn their keep
- whatsapp: first reply, qualification, booking and reminders. this is where we start most clients.
- website: answering pre-purchase questions on high-intent pages and pushing the buyer to whatsapp or a form.
- instagram dms: replies to story and ad comments, then a handover to whatsapp for the booking.
- internal ops: pulling ad, crm and sales data into one daily view. in our own agency ops, ai systems cut manual reporting effort by 60%+.
the sports for life agent shows the shape: manychat runs the whatsapp conversation, gpt-4o reads free text, n8n applies the rules, leadsquared and google sheets hold the record, and ringg ai voice calls no-shows. it qualified 60%+ of conversations with 0 manual first-touch. the work is in the rules, not the stack.
meta business agent on whatsapp: use meta's, a provider's bot or a custom agent?
use meta business agent if you are a small business answering catalogue, pricing and booking questions inside the whatsapp business app. use a provider's chatbot if your flows are mostly buttons and one inbox is enough. build a custom agent when the bot must score leads, write to your crm and follow strict handover rules.
coverage mixes up the names, so here is the timeline. meta introduced business ai in the whatsapp business app for indian small businesses on 14 may 2026, available in indian languages with no coding; it answers questions, recommends products, captures leads and books appointments, and the owner can take over a chat (meta). meta said upi payments inside the ai chat were coming soon, not live. on 3 june 2026 meta announced the meta business agent across whatsapp, messenger and instagram, free to start, with paid subscription offerings to follow (meta; cnbc). it was unveiled for india at whatsapp's mumbai business summit in july 2026, with access through a waitlist (uni; adgully).
| question | meta business agent | provider chatbot | custom agent (e.g. n8n + model + crm) |
|---|---|---|---|
| setup time | hours, inside meta's apps | days | weeks, including transcript review |
| best for | shops, salons, clinics with simple faqs | button flows, small sales teams | lead-heavy businesses with a crm |
| writes to your crm | limited; check what your account supports | often one-way sync | two-way, field by field |
| custom lead scoring | no | basic tags | yes, your rules |
| handover control | owner takes over in the app | inbox assignment | rules by trigger, owner and timer |
| cost | free to start, paid tiers announced | monthly plan plus messages | tools, model usage and a maintainer |
| who owns the logic | meta | the provider | you |
whichever you choose, keep it a business bot. whatsapp has barred general-purpose ai assistants from its business api since 15 january 2026; support, booking and sales bots remain allowed. we cover that policy, rule-based vs ai bots and handover triggers in our whatsapp chatbot guide.
ai chatbot for website or whatsapp: where do indian buyers actually talk?
for most indian businesses, buyers talk on whatsapp first, so the agent should live there. a website ai chatbot still earns its place on pricing, product and contact pages, where it answers the question blocking a form fill. treat the website bot as the door and whatsapp as the room where the sale continues.
we don't have a trustworthy india-wide figure comparing website chat with whatsapp chat, so we won't invent one. what is verified: meta, citing kantar, says 91% of online adults in india chat with a business weekly (meta). a website chat ends when the tab closes; a whatsapp chat is still there tomorrow, number attached.
in our accounts, click-to-whatsapp ads produce conversations and landing pages produce form fills. a website bot earns its place when it shortens the path to either.
- put a website ai chatbot on: pricing, product comparison, service area and contact pages.
- give it one exit: "continue on whatsapp" with the page context, or a short form.
- measure: chats that become a whatsapp thread or a form, not chat volume.
how does ai lead qualification decide who sales should call first?
ai lead qualification reads each conversation, pulls out the facts sales needs and turns them into a score with a reason. high scores go to a named salesperson within minutes, middle scores get a whatsapp nurture, and low scores get a polite close. the model extracts the facts; your rules assign the points.
don't let the model "feel" how hot a lead is. it extracts fields; fixed rules you can read score them. for sports for life, "qualified" meant four known fields (sport, age group, nearest centre and preferred time) plus a bookable slot. a points table adds priority on top when volume outgrows the sales team.
| signal | points | why |
|---|---|---|
| budget inside your price range | +25 | the commonest reason good-looking leads die |
| timeline within 30 days | +25 | urgency decides who gets a call today |
| location you serve | +15 | out-of-area leads waste a visit slot |
| asked for a visit, trial or callback | +20 | a stated next step beats any inferred intent |
| replied to every question | +10 | engagement predicts a picked-up call |
| came from a product-specific ad | +5 | context for the salesperson, small weight |
| job seeker, vendor or student project | -50 | route elsewhere, never to sales |
starting thresholds: 70+ to sales within 10 minutes, 40 to 69 to nurture, below 40 a polite close.
illustrative example, not client data: a parent messages at 11pm about weekend football for a nine-year-old near your andheri centre and asks if saturday works. location +15, timeline +25, asked for a trial +20, replied to every question +10: 70. the agent offers two saturday slots and books one, and the morning team sees a scored, booked lead with a two-line summary. budget scores zero because the trial is free.
send the score, reason and transcript to the crm together. field mapping is covered in our whatsapp crm integration guide.
how do you train an ai chatbot on your data without leaking it?
you don't retrain the model. you give it a small, approved knowledge base and make it look up the answer before replying, then refuse when the answer isn't there. feed it prices, policies and faqs you would print on a poster. never feed it customer lists, payment details or anything you would not show a stranger.
the technique is retrieval-augmented generation, or rag: an open-book exam. the customer asks, the system finds the two or three most relevant passages in your documents, and the model answers using only those. in the sports for life build, fees, timings and centres live in a sheet the team edits. when an answer is wrong, we fix the sheet, not the prompt.
| train it on | never train it on |
|---|---|
| faqs with approved answers, owned by a named person | customer names, phone numbers or chat exports with personal data |
| price list with an "as of" date | payment, bank, card or upi details |
| refund, cancellation and delivery policies | aadhaar, pan or other id numbers |
| catalogue, configurations, timings and locations | health, financial or children's records |
| handover rules and what the bot must refuse | internal margins, discount floors and negotiation limits |
| your top 50 real questions, anonymised | claims legal hasn't approved (returns, rera details, outcomes) |
dpdp: what changes for the data your agent collects
india's dpdp rules were notified in november 2025, and the core duties, including clear notices, purpose-specific consent, security safeguards and breach intimation, apply from 13 may 2027 (pib). for an agent that means:
- collect only what the job needs. a trial booking needs a sport and a slot, not a date of birth.
- separate purposes. consent to book a visit isn't consent to send offers.
- know where transcripts go. check how long your providers retain conversations.
- plan for breaches. ey reports a 72-hour window for breach notification under the rules (ey).
this is general information, not legal advice.
ai chatbot mistakes: where do agents fail, and which guardrails stop them?
agents fail in predictable ways: they invent a policy, get pushed off-topic, promise a discount, loop instead of handing over, or quietly stop writing to the crm. each has a cheap guardrail. set them before launch, test them with a real phone, and read transcripts every week, because the failures show up there first.
two public cases. in moffatt v. air canada (2024), the airline argued its chatbot was responsible for its own words; the british columbia civil resolution tribunal rejected that and held air canada liable for the wrong bereavement-fare advice (aba business law today). in january 2024, a customer got parcel firm dpd's chatbot to swear and criticise the company after a system update; dpd disabled the ai element (hcamag, as reported).
| failure mode | what it looks like | guardrail |
|---|---|---|
| invented answer | quotes a refund rule or price that doesn't exist | answer only from the knowledge base; fallback "let me get someone to confirm" |
| pushed off-topic | writes poems, insults the brand, answers general questions | scope the prompt to your business; refuse everything else; test after every update |
| unauthorised promise | offers a discount to end the chat | no pricing authority; discount requests go to sales |
| handover loop | asks the same question three times | hand over after two failed understandings, with the transcript |
| stale facts | last season's fees, a closed centre | dated knowledge base with a named owner and monthly review |
| silent failure | chats happen, the crm stays empty | daily count of chats vs crm records; alert on any gap |
| over-messaging | follow-ups to people who never opted in | opt-in stored with source and timestamp; one follow-up rule |
voice adds its own rules. ai calls in india must be declared to your telecom operator under trai's september 2026 amendment, which our voice guide covers.
what should a 30-day ai agent pilot look like?
a 30-day pilot proves one job on one channel with real leads. week one measures today's response time and writes the facts sheet. week two builds and tests the handover. weeks three and four run on a slice of traffic while someone reads transcripts daily. scale only when the numbers beat the baseline.
- days 1 to 5: measure first-response time and lead-to-booking rate by source. pick one job.
- days 6 to 10: write the knowledge base, the "never answer" list and the scoring table. get the owner to sign off on every price.
- days 11 to 15: build on the channel where leads arrive, usually whatsapp. connect the crm. run 30 test chats, including rude, off-topic and hinglish ones.
- days 16 to 23: go live on 20 to 30% of traffic. read every transcript daily. fix the knowledge base, not the salesperson.
- days 24 to 30: compare against the baseline on first response, qualification rate, bookings and handover time. decide to scale, change the job or stop.
more builds are on our work page.
who wrote this, how and why
who: puneetraj chavan, founder of rapchik media, with 8+ years running performance marketing and automation for 30+ brands across india and the us. how: meta's announcements, the dpdp rules, survey data and the case reports linked above were checked on 22 october 2026. platform-reported figures, such as meta's kantar statistic, were not independently audited, and the scoring example uses illustrative numbers. why: most writing on ai agents comes from people selling seats or tokens, and we wanted a version that starts with the handover rules, because that is where our builds succeed or fail.
- meta: business ai on whatsapp in india · meta business agent · cnbc on the launch · uni on the india launch
- adoption: microsoft work trend index 2025, india
- regulation: pib on dpdp rules · ey on dpdp rules
- failure cases: aba on moffatt v. air canada · hcamag on dpd's chatbot
want an ai agent that books, scores and hands over without making things up? see our services or talk to us.
founder of rapchik media. 8+ years scaling 30+ brands across india and the us, from agri-drones to a us job portal. strategy and execution, same brain.
more about me →
