At 11:20 on a Thursday night, Riya was watching a reel of a sage linen co-ord set from a small label she had never bought from. She sent the brand a DM asking whether it came in medium and what it cost. An auto-reply pointed her to the link in bio, which opened a homepage with 300 products and no obvious way to find the set she had just seen. She closed the tab. The brand's reply reached her at 10 the next morning, and by then she had already ordered something similar from a store that answered her in the chat.
Riya was trying to buy inside the conversation, as more shoppers now do, and conversational shopping AI exists to let them. In practice, it gives every product question in a DM or WhatsApp chat a real answer from the live catalog, and moves the conversation straight to a cart at any hour.
On the brand's side, that one reel brought the three-person team 140 DMs over the weekend. Sixty-two of them asked about a specific size, colour, or price, which is about as clear a buying signal as a store ever gets. The team replied to everyone by Monday afternoon and closed nine orders at a $72 average. That's $648 of revenue from a post that had earned the attention of dozens of ready buyers. Nothing was wrong with the product, the price or the reel. The gap was that the conversation happened in one place and the checkout lived somewhere else, with a delay and a homepage link in between.
This guide is for founders, heads of ecommerce and CX leads who sell through Instagram and WhatsApp and are deciding how much of that selling to hand to AI. It covers:
what actually happens, message by message, when someone buys inside a chat
why Instagram and WhatsApp need different playbooks, and how Meta's rules shape each one
the four things an AI must get right before it can sell in a chat
what running both channels costs, in dollars
what goes wrong when AI sells, and the guardrails that prevent it
where to start, channel by channel
What actually happens when someone buys inside a chat?
Messaging has become the main place where shoppers talk to brands. When Meta launched its Business Agent globally in June 2026, it said there were already more than one billion active threads with businesses on WhatsApp, Messenger and Instagram every day. AI is moving into those threads quickly too. Meta told investors in April 2026 that its business AI was handling about 10 million conversations a week, up from 1 million at the start of the year.
Volume only tells half the story. What matters for a store is what shoppers are trying to do in those threads, and most chat purchases pass through the same five messages. A good human seller handles each one in a way a scripted auto-reply can't.
| The message | What the shopper is really asking | What a good seller does | Where a scripted auto-reply breaks |
|---|---|---|---|
| "Do you have this?" | Is the exact thing I saw available to me right now? | Checks the variant (size, colour) in live stock | Sends a generic link or a "DM us your size" loop |
| "Will it fit me?" | Will I regret buying this? | Asks one or two questions, then answers from the size chart and product measurements | Pastes the whole size chart and leaves the shopper to work it out |
| "What goes with it?" | Help me get the look right | Suggests one or two in-stock pieces that work together | Recommends nothing, or recommends something sold out |
| "How do I pay?" | Make this easy before I change my mind | Sends a checkout link with the item already in the cart and states shipping up front | Points to the homepage or the link in bio |
| "Where's my order?" | Can I trust you with my next order? | Sends a proactive update in the same thread | Redirects to email or a tracking page |
The fifth message gets underestimated because it comes after the sale. It still decides whether there's a second sale, and "where is my order?" questions are expensive when nobody answers them in the chat. We broke down that cost in The WISMO Tax.
Seen together, the five messages give a working definition. Conversational shopping AI is AI that carries a shopper through all five inside the chat they started. It answers from live catalog, inventory, and order data, builds the cart, and keeps the same customer record when the conversation moves between channels. A tool that only handles the first message is an auto-reply. A tool that handles all five and writes the order back to your store is doing a salesperson's job.
Ephanti's manifesto sums up the difference in one line: "a chatbot deflects. An assistant suggests. An agent acts." In a shopping chat, acting means the cart, the order, and the update actually happen, which is more than producing a convincing reply.
Our take: We think most brands still staff their DM inbox the way they staff a support queue, with macros, assignment rules, and a reply time measured in hours. After a post lands, that inbox behaves like a shop floor at closing time, and every hour of delay costs the sales that were easiest to win.
Why Instagram and WhatsApp can't be run the same way
Most content treats "social DMs" as one channel. For a store, Instagram and WhatsApp do different jobs in the purchase, and Meta runs them under different rules. A playbook that works on one can cost money or break policy on the other. The table compares the two on the points that decide how selling works.
| Instagram DMs | ||
|---|---|---|
| Role in the purchase | Discovery and impulse from reels, stories and comments | Decision, follow-up, order updates and repeat purchase |
| Who starts the chat | Almost always the shopper | The shopper, or the brand once the shopper has opted in |
| Reply window | 24 hours for most actions; up to 7 days after a Click-to-Direct ad | 24-hour customer service window; outside it, only paid template messages |
| Checkout | Meta's Shops now send buyers to the brand's own website checkout | A checkout link shared in the chat, with catalogs browsable in the thread |
| Cost to message | No Meta per-message charge for replies inside the window; tool costs are usually billed per contact | Meta charges per delivered message by category and country |
| Where automation trips up | Missing the window, or misusing message tags to reach people outside it | Messaging people who haven't opted in, and sending marketing templates that are billed on every delivery |
Each rule in this table comes from Meta's own documentation, and each one limits what an AI can do on that channel. The sections below go through the two channels one at a time.
Instagram is where buying intent starts, and where it fades fastest
Instagram is where the impulse happens. A shopper sees a product in motion, wants it, and asks a question while the wanting is still fresh. Adobe's January 2026 recap of the holiday season shows how much this channel now drives sales: social media's share of US online holiday revenue rose to 4.6%, up 40.3% year on year.
The catch is that Instagram no longer lets the purchase finish inside the app. Meta's Business Help Center confirms that Shops on Facebook and Instagram now use website checkout, so every DM sale has to hand the shopper over to your site. That handover is where much of the loss happens, especially when the shopper discovers costs they weren't expecting. A Bizrate Insights survey of 1,010 US shoppers in March 2026 found that 42% often or very often abandon a cart because of unexpected costs such as shipping, taxes, or fees. Only 12% said they rarely or never do.
A clean Instagram DM to checkout flow deals with both problems. The shopper gets a link with their exact item, size, and quantity already in the cart, and the chat states the total, shipping included, before they tap. None of this needs AI. What AI adds is doing it at 11:20 on a Thursday night, for all 140 DMs.
The window is the other constraint. Meta's messaging policy, last updated in April 2026, gives businesses 24 hours to respond for most user actions. The window extends to up to 7 days only when the shopper's first message comes after clicking a Click-to-Direct ad. Once it closes, the brand can't message that shopper again unless they write in first, so a Monday reply to a Friday DM may simply fail to send. We cover the day-to-day mechanics in our guide to selling in Instagram DMs (Coming out soon).
A WhatsApp commerce platform closes the sale; a broadcast tool only announces
WhatsApp plays a different role in the purchase. Shoppers move there when they're closer to deciding, after they've opted in. It's also where order updates and repeat purchases happen. Meta's pricing documentation notes that businesses can only message WhatsApp users who have opted in, which makes WhatsApp a channel you earn rather than one you react to.
Many stores start with a broadcast tool that sends offers and order updates through templates. A WhatsApp commerce platform goes further and handles the two-way part of the sale. It answers product questions from the live catalog, builds the cart, recovers abandoned checkouts through conversation rather than a single reminder, and writes the order back to the store. That's why WhatsApp tends to outperform email at the recovery stage, where shoppers usually have a question that needs an answer. We compared the two directly in email vs AI agents for cart recovery.
WhatsApp’s pricing has also changed in ways that matter for automation. Meta has billed per delivered message since July 2025, and its current North American rates are $0.025 for a marketing message and $0.0034 for a utility message. Since October 1, 2026, Meta also charges for service replies inside the customer service window, once a business phone number has used its 1,000 free replies for the month. A WhatsApp strategy built on frequent blasts costs more than it did a year ago, and even a chat-heavy one is no longer entirely free. The brands that benefit are the ones whose WhatsApp messages actually move a sale or an order forward. That's also where the repeat-purchase case is won, which we cover in Why the Second Purchase Is Where Ecommerce Is Won.
Shoppers shouldn't have to repeat themselves when they switch apps
The most valuable moment in this setup is often the move between channels. A shopper finds you on Instagram, moves to WhatsApp to finish buying or to track the order, and comes back weeks later through either one. If each channel runs on a separate tool, each one knows only part of that history, and the shopper ends up repeating their size, their order number, or their complaint.
Ephanti's omnichannel platform is built around this moment. In one scenario on Ephanti's site, a customer asks about an order on Instagram and follows up on WhatsApp the next day. The agent recognises the same customer, loads the Instagram thread, and carries on without asking them to repeat anything. When the customer later calls in frustrated, a human picks up with the full two-day history already in front of them. Whichever platform you choose, use this as the test for the handoff: the shopper should never have to tell you something they already told you on another app.
Our take: We don't think "Instagram or WhatsApp?" is the decision worth agonising over. Instagram is where you earn the conversation, and WhatsApp is where you get to keep it, so the costly mistake is running them as two tools with two separate memories of the same customer.
Four things conversational shopping AI must get right before you let it sell
Putting AI on both channels is easy. Putting on AI that actually sells, without creating returns, policy strikes, or angry customers, depends on four capabilities. Each one comes with a test you can run in a vendor demo.
1. Your AI product recommendation agent suggests only what's in stock
Recommendations are where AI earns its place in a shopping chat, because they're the hardest part of a human seller's job to scale. A good AI product recommendation agent works from three live inputs: what is in stock in the shopper's size, what the shopper has said or bought before, and which items your catalog marks as going together. It asks one or two short questions ("dressy or daywear?") and then suggests two or three options, which helps far more than a scrolling carousel.
Shoppers who use AI to narrow their choices tend to arrive ready to buy. In its January 2026 analysis of the holiday season, Adobe found that visits referred by generative AI tools converted 31% more than visits from other sources. Adobe also found that 65% of shoppers who used AI felt more confident in their purchase. The same effect is available inside your own chat, provided the suggestions come from real stock. We've argued before that shoppers would rather discover than search, and a recommendation in the chat is discovery with a buy button attached.
Demo test: ask for a product in a size you know is sold out. A good agent offers an in-stock alternative or a back-in-stock alert, while a weak one recommends the sold-out item anyway or invents a variant.
2. It answers fit questions honestly, or hands over to a person
Fit questions make or break apparel and footwear sales in the chat, and they're where a guessing AI does the most damage. The agent should answer from your size chart, product measurements, and return policy. When a question goes beyond that data, it should say so and bring in a person rather than improvise a confident-sounding answer.
Shoppers have started to expect more from AI than answers, and they punish AI that gets in the way. Gartner's August 2026 survey of 3,566 customers found that 58% of those who use generative AI have used it to complete a task on their behalf. Gartner also cautioned that customers forced through several unsuccessful AI exchanges before reaching a person are less likely to use that tool again. For a store, that means one confident wrong answer about fit can cost both the sale and the shopper's willingness to try the chat next time.
Demo test: ask a fit question the product page can't answer, such as how the fabric behaves after washing, and see whether the agent admits it doesn't know or invents something.
3. It turns an Instagram DM to checkout in two taps, without a homepage link
Building the cart is where a conversation becomes revenue, and every step it takes affects whether the sale completes. The flow should go from "I'll take the medium" to a checkout page with that item already in the cart. The total cost should be stated in the chat, and the order recorded against the same customer profile that started the conversation. Every extra screen between intent and payment gives the shopper another chance to drift off. The Bizrate finding on surprise costs shows how often that drift ends in an abandoned cart.
Demo test: count the taps from the shopper saying yes to the shopper seeing a payment screen. More than two is a problem.
4. It follows up without breaking Meta's rules
The follow-up is where AI selling most often runs into Meta's rules. On Instagram, Meta's policy warns that using message tags outside their approved purposes may result in restrictions on your ability to send messages. The Human Agent tag exists for manual replies within 7 days and can't be used for automated sales nudges. On WhatsApp, follow-ups outside the 24-hour window must be paid templates sent to opted-in shoppers.
A well-designed agent plans around these limits rather than discovering them later. It makes its best offer while the window is open. If the shopper has opted in to WhatsApp, it moves the conversation there before the Instagram window closes, and it sends a template only when the message is worth the fee. For recovering carts on WhatsApp within these rules, our WhatsApp cart recovery guide covers templates, timing, and compliance in detail.
Demo test: ask the vendor what their agent does at hour 23 of an Instagram conversation, and again at hour 25.
Our take: We would rather an AI say "let me check that with the team" than guess a size. A slow, honest answer costs a few minutes, while a confident wrong one costs a return, a review, and often the customer.
What does it really cost to sell inside Instagram DMs and WhatsApp?
Cost conversations about chat commerce usually stop at subscription prices. That hides two bigger numbers: what the manual version costs you today, and what a stitched-together tool stack costs in a good month. The worked example below uses one store, so every figure can be checked line by line.
The example store is a US apparel brand doing about 2,500 orders a month at a $72 average order value. It receives around 1,800 Instagram DM conversations a month and has 8,000 opted-in WhatsApp subscribers. These store figures are our assumptions; each benchmark used in the calculations is linked.
Answering DMs by hand costs more than most founders realise
Before buying anything, it helps to price the current manual process. Assume each DM thread takes about 4 minutes of handling, once you include reading it, checking the size in Shopify, finding the right link, and replying. At 1,800 threads, that's 120 hours a month. The median US customer service representative earned $21.53 an hour in BLS's May 2025 occupational wage data, published in May 2026. At that rate, the store spends about $2,584 a month, or roughly $31,000 a year, on wages alone before benefits. All of that time falls in working hours, while much of the buying intent arrives in the evening.
The second manual cost is the window. Suppose even 10% of those 1,800 threads, or 180 conversations, get their first reply after Meta's 24-hour window has closed. The brand then loses the ability to message those shoppers at all. At a $72 order value, every ten of those 180 who would have bought is $720 the store never had a chance to earn.
The usual tool stack runs about $16,000 a year
Most stores that automate both channels end up with four tools plus Meta's message fees. The table below prices a common setup at public US list prices from October 2026. It covers a month in which one reel lifts Instagram engagement to about 4,000 active contacts.
| Tool category | What the store gets | Monthly cost |
|---|---|---|
| Instagram DM automation, mid tier | 2,500 active contacts included; 1,500 extra at $0.05 each | $39 + $75 = $114 |
| WhatsApp business messaging platform, mid tier | 5 users, unlimited broadcasts, 2,000 automation triggers (monthly billing) | $149 |
| Meta WhatsApp message fees | 16,000 marketing messages (two broadcasts) at $0.025 = $400; 1,200 cart-recovery templates at $0.025 = $30; 7,500 order updates at $0.0034 = $25.50; 5,000 service replies beyond the free 1,000 at $0.0034 = $17 | $472.50 |
| Ecommerce helpdesk, 2,000-ticket tier, plus its AI agent add-on | Shared inbox for email, chat and social; AI add-on from $250 | $360 + $250 = $610 |
| Total | $1,345.50 a month, or about $16,146 a year |
Exclusions: extra seats beyond those included, the messaging platform's own markup on Meta's rates (we used Meta's list rates), helpdesk ticket overages, AI interaction overages, ad spend, and SMS or voice. Annual billing lowers several of these lines, and a busier month raises at least three of them.
The bill spikes in the month your reel takes off
Every tool in that stack runs on a different meter, and the meters all climb together when something works. Merchants have already noticed. One leading Instagram automation tool's own August 2026 pricing explainer addresses complaints, gathered from Reddit threads and comments, that a viral reel can leave a business with a huge bill at month-end. In reply, the company acknowledges that a viral spike is charged in the month it happens. These are the metrics to watch:
Contact-based Instagram billing. Every person who interacts with your automations that month counts toward the bill, so a reel that pulls in thousands of new people raises your cost for that month.
Markups on WhatsApp messages. Many WhatsApp business messaging platforms bill messages from a prepaid balance at their own rate card rather than at Meta's list rate. Compare the two line by line before you commit.
Paid service replies. Since October 2026, the free tier covers only 1,000 service replies per number per month, so a brand that moves its whole DM operation to WhatsApp will pass it early in the month. If you use Meta's own business agent, Meta prices its replies by tokens at roughly 4 to 5 cents a message, by its own estimate.
Metered AI add-ons. Helpdesk AI is usually billed per automated interaction, with an allowance and an overage rate, so its cost grows exactly as fast as it succeeds.
Beyond the invoices, there's an operational cost that never shows up on one. When the Instagram tool, the WhatsApp tool, and the helpdesk each hold part of the customer's history, your team spends hours reconciling three inboxes and the shopper spends time repeating themselves.
One platform makes sense once DMs outgrow a single person
The case for consolidating is that one platform covers the conversation on both channels, with one customer record and one bill. According to Ephanti's E-Commerce solution page, typical customers consolidate 4 to 6 point tools within 60 days. Ephanti bills on the interactions handled, with unlimited users and no per-seat fees. It doesn't publish prices, so we won't put a number next to the stack above. The E-Commerce ROI calculator lets you run your own volumes, and the head-to-head on how usage-based pricing compares with ticket-based helpdesks shows where the meters differ.
A full WhatsApp commerce platform isn't the right first purchase for everyone, though. If your store gets fewer than about 20 DMs a day, and one person who knows the catalog can answer them inside the 24-hour window, Instagram's saved replies and the free WhatsApp Business app may be enough for now. The case for conversational shopping AI becomes strong once a meaningful share of buying questions arrives outside working hours, or once a single post can create more conversations than your team can answer in a day.
Our take: We think the sticker price is the least important number in that table. The bill that hurts arrives in the month your reel works, when contact-based, message-based, and resolution-based meters all spike together, so ask every vendor to price your best month rather than your average one.
What happens when the AI says yes and you didn't?
AI selling carries different risks from AI support, because a selling AI makes commitments. It quotes prices, confirms stock, and can be talked into discounts. Shoppers are open to AI but want a way out of it: Gartner’s August 2026 survey found that 87% of customers consider it essential that companies using generative AI offer a way to reach a human agent.
The most instructive case so far came from a small UK online store in February 2026. Its owner described on Reddit how a customer spent about an hour coaxing the shop’s after-hours chatbot into inventing a 25% discount code, then pushed it up to 80%, before placing an order worth more than £8,000 and demanding the discount be honoured. The bot had never been given authority to discount anything. A legal analysis from Syracuse University’s technology law journal used the case to argue that businesses can be bound by such deals under the doctrine of apparent authority, because a branded chat window leads shoppers to reasonably rely on what it says.
Regulation is tightening too. The EU AI Act's transparency rules have applied since August 2, 2026. They require that people be told when they are interacting with an AI system such as a customer-service chatbot. Any store selling to EU shoppers through an AI in the chat now has a disclosure duty, alongside the disclosure Meta already requires wherever local law demands it.
| Guardrail | Both | ||
|---|---|---|---|
| Hard limits | No discount codes or prices outside approved rules | Same, plus no templates sent without opt-in | Refund and discount ceilings above which a person must approve |
| Source of truth | Stock, price and variants read live from the store | Order status read live from the store and carrier | Nothing quoted that didn't come from a connected system |
| Disclosure | The AI says it's an AI at the start of the conversation | Same | Required in the EU since August 2026, and wherever local law demands it |
| Window and message discipline | No message-tag workarounds outside the 24-hour window | Templates only for opted-in shoppers, and only when worth the fee | One follow-up schedule across both channels, so shoppers never get the same nudge twice |
| Audit | Every reply logged | Every message and its cost logged | Every action, override and escalation searchable |
Ephanti builds these limits into the platform. Escalation triggers include a refund or dispute above a configured dollar threshold, and a customer asking for a person. The human who takes over receives the full history across channels, along with the reason the agent escalated. Ephanti's Trust Center describes full audit logs of every agent decision, system action, and human override. Its WhatsApp agent is designed to flag uncertainty and hand off rather than guess. We explain the design thinking behind this in designing an accountable agent architecture.
Our take: We believe the safest AI in a shopping chat is one that can't invent anything, meaning no price, code or stock level that didn't come from a live system. The UK store's bot was only ever meant to answer questions, and that intention protected nobody, because a chat window is judged by what it says rather than what it was built for.
Where should you start: Instagram, WhatsApp or both?
There's no universal first step, because the right one depends on where your sales are leaking today. The table matches the starting point to the symptom, with the question to ask vendors and the numbers to watch in the first 30 days.
| Start with Instagram | Start with WhatsApp | Run both together | |
|---|---|---|---|
| Start here if | Posts regularly trigger more DMs than your team can answer within a day | You already have opted-in subscribers and lose carts at checkout | Shoppers routinely move between the two, and your team keeps re-asking for details |
| First use case | Answer stock, size, and price questions and send a pre-filled checkout link | Conversational cart recovery and order updates | One customer record across both, with a single follow-up schedule |
| Vendor question to ask | "What does your agent do at hour 23 and hour 25 of a DM?" | "Do you bill messages at Meta's rate or your own rate card?" | "Show me a shopper moving from Instagram to WhatsApp without repeating anything." |
| Measure in the first 30 days | Median first-reply time, DM-to-checkout rate, conversations lost to the window | Recovered revenue per message sent, opt-out rate | Repeat questions per customer, total cost per order across both channels |
How you roll it out matters as much as where you start. Ephanti recommends a phased migration instead of a weekend cutover. You connect your store, helpdesk and channels in the first week. The agent then handles a slice of volume, such as one channel or one queue, while your current tools keep running everything else. You widen the scope as confidence builds and retire tools only after the numbers prove the case. On Ephanti's site, most e-commerce teams go from contract to live within 30 days. The full plan, with timelines and common pitfalls, is in the E-Commerce Playbook.
Our take: We'd start where the store is losing sales today, which is rarely where the most exciting demo is. A month of honest data from one channel will tell you more about the second channel than any vendor deck.
Riya's next order took four minutes
A few weeks later, the same label posts a new colourway of the co-ord set, and Riya DMs again just after 11 pm. This time the reply arrives within seconds, confirming that medium is in stock in the new colour. The agent asks whether she wants the set for daywear or evenings and suggests matching sandals that are in stock in her size. It then sends a checkout link with both items already in the cart and the total, shipping included, stated in the chat. She pays before midnight. She had opted in to WhatsApp updates last time, so her order confirmation and dispatch message arrive there, in the thread she'll use if anything goes wrong.
The brand's team starts the next morning with the orders already in Shopify. Three conversations are waiting for a person: an alteration request, a bulk enquiry and a refund above the auto-approval limit. That's what conversational shopping AI should deliver. Shoppers buy at the moment they want to, and the team spends its hours on the conversations that need judgment.
As Sri Sabesan, Founder & CEO of Ephanti, puts it: "Every customer conversation should end in something finished, not disappear into a tool that only shows you the last message."
Bring your ugliest workflow. Name the job nobody on your team wants to do, whether it's the Friday-night DM pile or the cart that stalls at checkout. See it handled in your own systems, on the channels your customers already use. Book a demo.
Frequently asked questions
Can customers buy directly inside Instagram DMs?
Not fully. Meta's Shops now use website checkout, so the payment happens on your site. The conversation can still do almost everything else: confirm stock, answer sizing questions, and send a checkout link with the item already in the cart, so the shopper only sees the payment step.
How does an Instagram DM to checkout flow actually work?
In a strong Instagram DM to checkout flow, the shopper asks about a product, and the AI checks the exact variant in stock and answers any fit questions. It then sends a link with that item and size already in the cart. The total, shipping included, is stated in the chat before the link, and the order is recorded against the same customer profile that started the conversation.
What does a WhatsApp commerce platform cost?
Expect three layers: the platform subscription, Meta's per-message fees, and any add-ons. In the US, Meta charges $0.025 per marketing message and $0.0034 per utility message, and service replies are charged after 1,000 free per number each month. A mid-sized store commonly spends several hundred dollars a month on messages before the subscription, depending on how many broadcasts it sends. A WhatsApp commerce platform that recovers carts through conversation rather than repeated blasts usually keeps the message bill lower.
How does an AI product recommendation agent decide what to suggest?
A good AI product recommendation agent works from three things: live stock in the shopper's size, the shopper's stated needs and past purchases, and which products your catalog pairs together. On Ephanti, recommendations draw on customer behaviour, browsing activity, purchase history, and shopping intent. The aim is to help shoppers make decisions faster and increase average order value through relevant add-ons.
Is conversational shopping AI allowed under Meta's messaging rules?
Yes, within the rules. Meta's WhatsApp documentation notes that service replies can be powered by third-party AI, and Instagram allows automated replies inside the standard messaging window. Conversational shopping AI must respond within the window and disclose that it's automated where required by law, which now includes the EU. It must not use message tags for sales follow-ups, and on WhatsApp it can only message shoppers who have opted in.
Should I start with Instagram, WhatsApp or both?
Start where you're losing the most sales today. If posts create more DMs than your team can answer within a day, start with Instagram. If you have opted-in subscribers and high checkout drop-off, start with WhatsApp. Run both together once shoppers regularly move between them, because that's when a shared customer record pays for itself.