AI AutomationBy Nacho Nayar · 8 min read
AI for customer service over email and phone: beyond WhatsApp
Almost every small business that automates customer service starts with WhatsApp, and for good reason: that's where the volume is and where customers expect an instant reply. The problem is that most of them stop there. While the chatbot answers questions at three in the morning, the info@ inbox piles up with unsorted requests and the phone rings after hours with nobody knowing who called or why.
Email and phone aren't just a slower WhatsApp: they're channels with different rules, and automating them with the same playbook is the fastest way to get it wrong. Here's what AI can do in each one today, what's still worth leaving alone, and how the investment actually adds up.
Key takeaways
- Email tolerates delay: between the question arriving and the answer going out, there's room for a human to review. That makes it the safest channel to automate first.
- In email, start in draft mode: the AI writes the reply using real business information and a person approves it before sending. You capture almost the same time savings at a fraction of the risk.
- Sorting and routing pays off more than answering: knowing whether an email is a complaint, a quote request or an invoice — and getting it to the right person — solves the actual bottleneck.
- Phone calls forgive neither latency nor robotic voices: if the assistant hesitates or sounds artificial, people hang up. Start with the after-hours window, not by replacing your morning shift.
- You have to disclose that it's an automated assistant. It isn't only a legal matter: conversations work better when people know what they're talking to and how to reach a human.
- All three channels are built within the same AI automation service, starting at USD 1,000, and they share a single knowledge base about your business.
Why email is the best place to start
On WhatsApp, whoever writes expects an answer in seconds, and that expectation forces the assistant to reply on its own with nobody watching. Email works the other way around: nobody expects an instant response from a contact inbox. Those minutes of tolerance are exactly what you need to put a person in the loop, which is why email is where a small business can automate with the least possible risk.
The mode that works best isn't the auto-reply but the assisted draft: the AI reads the question, pulls from real business information — current prices, stock, lead times, policies — and leaves a reply ready for review. Whoever handles support goes from writing every email from scratch to reading it, fixing a line and hitting send. Time per request drops sharply and no answer goes out without a human seeing it.
Sorting and routing: the highest-yield piece
In most small businesses the email problem isn't writing the reply: it's that everything lands in the same inbox and someone has to open them one by one to find out what each is. A classifier solves that without writing a single reply: it reads every incoming email, decides whether it's a quote request, a complaint, a supplier invoice or spam, tags it and routes it to the right person or system.
It's the email automation with the best impact-to-risk ratio: if the classifier gets it wrong, the worst case is that an email lands in the wrong inbox and someone forwards it. Nobody received an incorrect answer, because nothing was answered.
Pulling data out of attachments
The other silent job in email is attachments: purchase orders, delivery notes, payment receipts, order spreadsheets. A workflow can read the attachment, extract the data and load it into your system, flagging for human review anything that doesn't match the expected format. This doesn't call for a conversational agent: a rule-based workflow is enough, and it's cheaper and doesn't over-interpret.
The phone: the newest and the trickiest
AI phone support is no longer the touch-tone menu of twenty years ago. A voice assistant today understands someone speaking normally, without pressing any numbers, and can take the reason for the call, book an appointment, report an order status or route the call to the right person with the context already loaded. Voice quality has improved to the point where the accent is no longer the problem.
The problem is elsewhere, and it's twofold. First, latency: in a spoken conversation, two seconds of silence feel like a dropped line, and if the assistant hesitates, the caller hangs up. Second, expectations: people who call instead of writing usually do so because they're in a hurry or because the issue is complicated — the two worst conditions for running into a robot.
Where to start without breaking anything
The after-hours window. Right now those calls are lost entirely: nobody's there and no record is left of who called. An assistant that picks up at ten at night, understands the reason, takes the details and leaves the case ready for the next morning isn't competing with your team — it's competing with silence. It can't make service worse, because there's no service to make worse, and it lets you measure how real people respond before touching business hours.
The natural second step is overflow: when three calls come in at once and two are left on hold, the assistant takes the ones that would have dropped anyway. Only after that, if the numbers hold up, does it make sense to discuss front-line coverage during business hours.
What not to automate on any channel
Complaints and negotiations. Not because of a technical limitation: a model can draft a response to a complaint just fine. The point is that someone complaining is already upset, and finding out a machine replied turns a service problem into a brand problem. Same goes for any conversation involving money, terms, or an exception to policy: a person closes those.
The practical rule is simple: AI keeps what's repetitive and has a single correct answer, and people keep what takes judgement, context or the authority to decide. A project that tries to automate one hundred percent of support doesn't end up one hundred percent automated: it ends up with customers typing "I want to talk to someone" in their first message.
An assistant that says "a person will look at this and get back to you today" serves the customer better than one improvising an answer to save face. Knowing when to hand off is a feature, not a failure.
One knowledge base, three channels
The most expensive implementation mistake is treating each channel as a separate project: a WhatsApp chatbot with its answers, an email assistant with different ones, and a voice agent with its own. When a price changes you have to update it in three places, and in practice it gets updated in one: two months later each channel is saying something different.
What you want is a single knowledge base — prices, stock, lead times, policies, hours — and three channels that query it. What changes between channels is the form: on voice, answers have to be short and list-free; on email they can run long and detailed; WhatsApp sits in between. The information is the same and gets updated once.
What it costs
Automated email and phone support is built within the AI automation service, which starts at USD 1,000 and scales with the number of channels and integrations with your systems. On top of that comes the monthly cost of the automation platform and model usage, which for a small business typically runs between USD 20 and USD 150.
Voice carries an additional cost of its own: you pay per minute of conversation, and that number depends on call volume and provider, so it's worth quoting separately rather than assuming it inside the monthly fee. It's the main cost difference between automating text and automating voice, and the reason starting with email almost always makes sense.
Frequently asked questions
Can AI answer emails on its own, with nobody reviewing them?
Technically yes, but it's not where you should start. Draft mode — AI writes, a person approves and sends — captures nearly all the time savings at a fraction of the risk. Fully automatic replies make sense only once you've measured months of drafts and know which request types always come out right.
Do I have to tell customers they're talking to an AI?
Yes. Beyond being the right thing to do and increasingly required by regulation, it works better: when people know they're talking to an assistant, they adapt how they ask and request a human when they need one, instead of discovering it midway through a difficult conversation.
Can a voice assistant handle business hours from day one?
It can, but that's not where to start. The after-hours window doesn't compete with your team — those calls are lost today — and it lets you measure how your actual customers respond before touching front-line coverage.
What happens when the assistant doesn't know the answer?
It has to hand off, not improvise. A well-scoped agent routes to a person with the full conversation context when a question falls outside its knowledge base. An agent without that rule makes things up, which is worse than not having answered at all.
Is it worth it if my company gets only a few requests a day?
At low volume, the classifier and the assisted draft still pay off because they save your team time from day one. Automated phone support, on the other hand, only makes sense once there's a measurable volume of missed calls.
If WhatsApp is already sorted and the info@ inbox is still a black hole, or the phone rings after hours with nobody knowing who called, tell us how you handle support today and we'll tell you which part to automate first and what to leave with your team. At loco22 we build all three channels on a single knowledge base, with a fixed quote from the first assessment.
Found this useful? Add us as a preferred source in Google.