Customer conversations
Should AI send your customer replies for you?
The case against auto-replying to customers, what a drafted reply does instead, and how nine channels become one queue without anything leaving unread by a human.
How MarketWar handles it
An AI reply here is a DRAFT and is never auto-sent: a human presses send on every message, across all nine channels the inbox collapses into one queue — email, SMS, WhatsApp, web form, reviews, Facebook, Google, chat and missed calls. Summaries, priorities and forecasts are labelled estimates, and the engine will not fabricate a message, a contact or a metric to fill a gap.
There is a version of this feature every vendor sells: connect your channels, switch on AI, and let it answer. It demos beautifully.
Here is what it looks like three weeks in. A customer asks whether you can do a job on the 14th. The model does not know your diary, so it produces a sentence that sounds like a yes. You find out when they arrive.
The rule, and it is not negotiable here
An AI reply is a draft. A human sends.
Not as a setting you can switch off. The engine's doctrine says replies are labelled drafts and are never auto-sent, and that is enforced in the engine rather than suggested in the documentation.
The reason is an asymmetry. Across fifty routine enquiries, autonomy saves you an hour. On the fifty-first — the one about a refund, a deadline, an allergy, a legal question — it costs you the customer, and possibly more than the customer. Nobody notices the hour they saved; everybody remembers the sentence a machine sent in their name.
What you actually get, which is the typing
The slow part of replying is not sending. It is reading the thread, remembering the context, and composing the first sentence. That is the part that is drafted for you:
- every thread in one queue across nine channels,
- prioritised, with how long that person has been waiting and whether you have breached your own promised response time,
- a drafted reply in your own brand voice, ready to edit,
- a pipeline view with a weighted forecast, labelled an estimate, because a forecast is a guess with arithmetic in it.
You read, you adjust, you send. The work goes from twenty minutes to three, and a person is still accountable for every word that leaves.
Nine channels, one queue
Email, SMS, WhatsApp, web form, reviews, Facebook, Google, chat and missed calls. A missed call is in that list on purpose: for a trade business it is the highest-intent message of the day and it arrives in a place nobody is watching.
Nothing is fabricated to fill a gap
This matters more than it sounds. An inbox summary is a perfect place for a model to invent — a contact's name, a figure from a thread it misread, a metric that makes the dashboard look complete.
The engine does not. It never fabricates a message, a contact or a metric, and every derived number is labelled an estimate rather than presented as a count. A thread with no information in it summarises as a thread with no information in it.
And it is not a way around the sending rules
If a reply turns into marketing follow-up, it is marketing. Consent applies, and so does the cap of five touches in seven days — the same cap the campaign engines enforce. An inbox that quietly became a bulk sender would put every tenant on the shared sending domain at risk, which is the one cost that is not yours alone.
Related reading: how to ask for reviews without buying them, and which channel actually produced the sale.
The question to ask
Ask a vendor selling autonomous replies one thing: what happens when it is wrong, and who is accountable?
If the answer is a confidence score, you are the accountability.
What this does not do
It will not answer your customers while you sleep, and that is a refusal rather than a missing feature. If what you want is an autonomous agent talking to buyers unsupervised, this is the wrong product and you should know that before you pay for it.
Common questions
Why won't it just send the reply?
Because the cost of one wrong answer to a real customer is higher than the time saved on fifty right ones. A draft gets you the typing; a send gets you the liability.
Which channels does it collapse into one place?
Nine: email, SMS, WhatsApp, web form, reviews, Facebook, Google, chat and missed calls. One queue, prioritised, with how long each person has been waiting.
Does it invent a summary when a thread is unclear?
No. Summaries and priorities are labelled estimates and the engine never fabricates a message or a contact. An unclear thread reads as unclear.
Can it still be used for marketing follow-up?
Where a reply becomes an outbound marketing touch, consent applies and a cap of five touches in seven days applies with it. The inbox is not a loophole around the sending rules.
Related
How do you ask for reviews without breaking the rules?
When to ask, what to say, where to send them, and why filtering out the unhappy ones breaks the rules and costs more than a bad review.
What happens when the AI provider fails mid-job?
AI providers have outages, rate limits and bad days. What that costs you, and why a gateway that reserves before charging bills nothing.
Which marketing channel actually made the sale?
Last-click credits whoever stood nearest the till. What a position-based split does instead, and where in the funnel the money leaks.