Automating your WhatsApp, Instagram and TikTok messages means handing an AI agent the repetitive questions — price, stock, size, order status — and keeping for yourself the conversations that call for judgement. The human touch is not lost because a machine replies. It is lost when the reply lands the next day, or when it reads like a form letter.
What an AI agent handles alone, and what it should hand back
An AI agent handles alone anything with a factual answer already sitting in your catalogue: the price of an item, whether a size is in stock, an order confirmation, a delivery update, an abandoned-cart nudge, a question asked at two in the morning. These exchanges are the volume, not the value, and none of them needs a decision from you.
Three kinds of conversation belong to a human, and a good agent hands them over instead of improvising. An automation that refuses to step aside is exactly what makes a customer feel they are talking to a wall, and that is the moment the sale disappears:
- Haggling over a large basket, where any discount is a case-by-case call
- A complaint loaded with emotion, which needs listening to before solving
- Complex advice, such as picking a face cream for a particular skin type
The line between the two is not fixed. It moves as you give the agent feedback the way you would an employee on the cases it got wrong, much as you would coach a new salesperson through their first weeks. Start generous on the human side, then hand back the cases the agent has learned to handle.
Bringing WhatsApp, Instagram and TikTok into one inbox
Running WhatsApp Business on one phone, Instagram on another and TikTok Shop on a third costs you one to two hours a day, purely in switching between apps. That time produces nothing: it does not sell anything, it does not pack any orders, and it does not make a single reply better.
One inbox also fixes a quieter problem: consistency. When every channel keeps its own history, the same customer is quoted one price on Instagram and another on WhatsApp, and nobody remembers who promised what. A tool such as Bedones Moderator brings five channels into a single interface, with one agent answering everywhere in the same voice and from the same information.
A single inbox changes how you measure the business, too. You can see at a glance which channel brings questions and which brings orders, and that is where the difference between a chatbot and an AI agent starts to show: one of them replies, the other knows what it is selling.
Plugging the agent into your catalogue
An AI agent plugged into your catalogue answers like your best salesperson; an agent without one answers like a switchboard. That is the only difference the customer actually feels, because it decides whether the conversation ends in an order or in a vague promise to get back to them later.
Four blocks of information are enough for an agent to carry a sale from the first message through to payment. They have to be current: an agent announcing a promotion that ended last week does more damage than an agent that stays silent, because it commits you in front of the customer:
- Products and their variants: sizes, colours, options
- Current prices and running promotions, with their end dates
- Delivery zones you cover and the lead time attached to each
- Accepted payment methods and how a payment gets confirmed
The catalogue is also what makes the agent correctable. When a reply annoys you, you do not rewrite a script: you fix the data or the instruction, and the agent applies the change across all five channels at once. It is also what lets an agent recognise your products from a screenshot a customer sends.
Tone is what separates a bot from a salesperson
Tone does most of the work: with identical content, two phrasings do not convert the same way. “Hi Mariam, we’ve got the M in stock, want me to hold it for you?” gets an answer. “Dear valued customer, we hereby confirm the availability of the requested item” gets a polite silence, followed by an order placed somewhere else.
Four rules are enough to set the agent’s voice, and they hold across all three channels. Decide them once, write them into the agent’s instructions, then leave them alone: consistency matters far more than which option you picked in the first place:
- Formal or casual address to match your brand, but the same one everywhere
- The customer’s first name in every reply, never “dear customer”
- One emoji or two where the context invites it, never six
- Short sentences, two or three lines per message at most
Good tone does not excuse slowness, though. A warm message sent the next day arrives after the order has already gone to a competitor, which is why replying in under five minutes is the first thing to fix, well before you polish a single sentence.
The three numbers to watch in the first week
Three numbers tell you whether your automation is working, and all three can be read in the first week. Do not build a full dashboard on day one: these figures already say whether the agent is saving you time or quietly creating extra work for the rest of your team.
- First response time, target under one minute
- Share of conversations turning into orders, target +20% versus before automation
- Human takeover rate, target steady around 15 to 20%
The human takeover rate is the most informative of the three. If it drops very low, the agent is probably answering things it does not actually know; if it climbs week after week, your catalogue is incomplete or your instructions are too narrow. A steady 15 to 20% means the split between the machine and you is set correctly.
The best automation stays invisible to the customer. They should never think “ah, a bot” — they should think “these people reply fast and they know their stock”. Those three numbers tell you, with figures rather than impressions, which side of that line your conversations are sitting on.