Selling in Chat: What It Does to Sales and to Support Costs

Chat automation gets sold two ways. To a support manager it is a cost saving: fewer
messages reach a person. To a founder it is a sales channel: more conversations end in an
order. Both pitches are true and each one on its own is misleading, because the two
numbers move together and sometimes against each other.

This is an attempt to set out the actual arithmetic, including the parts that do not
flatter the idea.

The sales side

Three things produce revenue in a chat that a website does not.

The question that was blocking the sale gets answered while the customer is
still there. Will it fit. Does it arrive before Saturday. Can I return it. Is
there one in black. A shopper with an unanswered question at the point of purchase either
asks or leaves, and most leave. Answering in seconds converts a share of them. Answering
the next morning converts almost none.

Recommendation happens in context. A customer describing what they
need can be pointed at the right item, which is a conversation a product grid cannot have.
One Shopify App Store review, from a solo watch retailer in Singapore, puts it plainly:
“Love the fact that it can recommend the right watch to my customer”.

The sale can complete in the thread. No checkout, no account creation,
no delivery charge appearing at the last step. The documented average cart abandonment
rate across published studies was 70.22% as of September 2025 (Baymard Institute). A chat sale does not
enter that funnel.

How much is that worth? Nobody can tell you from outside your business, and you should
be suspicious of anyone who offers a number. What is available is directional evidence.
The e-Conomy SEA 2026, Google, Temasek and Bain & Company, published on 21 October 2026, reports that 69% of consumers in the six main
Southeast Asian markets now use AI to compare product options and 93% trust the
recommendations they get from it, and that AI is delivering order volume uplifts “as large
as 75%” for some commerce players. Read the second figure as a ceiling observed somewhere,
not a forecast for you.

The useful point in that report is not the uplift. It is the adoption: 243 million
people across those six markets, 58% of the working population, started using generative
AI within three years. Your customers have already changed how they shop. Most businesses
they buy from have not changed how they answer.

The support side

Here the arithmetic is more tractable, because the work is countable.

Take a month of messages and sort them into three piles.

Answerable from a document. Hours, delivery areas, return window,
parking, materials.
Answerable from a system. Where is my order, is this in stock, what
slots are free on Friday, what is my balance.
Needs judgment. Complaints, exceptions, negotiations, anything
clinical or legal.

Pile one is what most chatbots handle. Pile two is usually larger, and it is where tools
built only on document retrieval fail, because the answer is not written anywhere. It is
in your order records, your calendar or your stock. Pile three should never be automated,
and attempting it is how businesses lose customers to automation.

Your realistic saving is piles one and two. Not the whole inbox. Any vendor quoting you
a single resolution percentage is quoting their best customer. For scope, here is ours
with its limits attached: for one bakery merchant, AeroChat handles 87 to 94% of
conversations with no person, through festive peaks. One merchant, one business type, one
period. We do not present it as a platform-wide figure, and you should not accept one from
anybody else either. See resolution rate.

Where the two numbers interact

This is the part the two sales pitches both leave out.

Selling in chat creates support volume. Every order taken in a
conversation produces a “when is it coming” message in the same conversation, usually
within a day. If you add chat selling without connecting delivery updates, you have bought
yourself a new support queue. See WISMO.

Support quality is a sales input. The question that gets answered at
10pm is sometimes the sale. Treating the two as separate budgets makes the trade-off
invisible.

Automation that cannot hand over costs you both. A customer stuck with
a bot that will not fetch a person does not buy and does not come back. When we read 2,793
public reviews of 67 customer service and chatbot apps on the Shopify App Store, support
quality was the single largest complaint theme at 19%. Control over what the AI says
ranked fifth, at 8%. People are angrier about not being helped than about being helped by
software.

The cost line nobody budgeted for

One change this month makes the arithmetic sharper. Since 1 October 2026
Meta charges for service messages on WhatsApp, which are the freely worded replies you
send inside the 24-hour window a customer’s message opens. There is a free tier of 1,000
service messages per business phone number per month, and it does not roll over (Meta’s WhatsApp Business Platform pricing documentation).

Those replies used to be free. Now the number of messages it takes you to finish a
conversation has a price. Two consequences worth planning around:

Brevity is now worth money. Three short messages where one clear one would do costs
three times as much.
Proactive updates are not automatically cheaper than answering questions. Sending
three delivery notifications to avoid one “where is my order” is a calculation, not an
obvious win. Work it out on your own volumes. See conversation-based pricing.

A worked example, with the unflattering parts left in

Take a store getting 600 customer messages a month. The sort comes out at 180
document-answerable, 300 system-answerable and 120 needing judgment. Say automation closes
most of the first two piles and none of the third.

On the support side, that is roughly 400 messages a month that no longer wait for a
person. At three minutes each including the interruption, call it 20 hours. Whether 20
hours is worth anything depends entirely on what the person doing it would otherwise do.
For an owner-operator it is 20 hours back. For a team of four it is not a headcount
saving, it is a backlog saving, and it will show up as faster replies rather than as a
smaller wage bill. Be honest with yourself about which one you are buying.

On the sales side, the number to find is different. Count the conversations last month
where a customer asked something, you answered after a delay, and they never replied. Not
the ones who said no. The ones who went quiet. If that is 25 conversations and your average
order is worth something you can state, you now have a range for what late replies cost
you. Most people running this count for the first time are surprised by it, which is why
it is the number worth starting with.

Then subtract the costs that are easy to forget. Meta’s per-message charges on the
replies you send. The time to get your data into one consistent state. And the first month,
in which you will be correcting the workflow rather than saving time.

How to work out what it is worth to you

Four numbers from your own data, none of which requires buying anything.

Messages per month, by pile. From the sort above.
Minutes per message. Time yourself on ten. Include the context
switch, which is the real cost.
Conversations that went quiet after a question you answered late.
Count them for one week. This is your visible lost-sales number and it is usually the one
that changes minds.
Replies per finished conversation. You need this to compare any
vendor’s pricing, and to estimate your WhatsApp message bill. See conversation-based pricing.

Run those four and you will know whether this is a support project or a sales project
for you. It is usually both, in a ratio nobody could have guessed from the outside.