en

Helping Hands

Omnichannel AI in Customer Service: Why Every Channel Answers Differently

Omnichannel AI in customer service explained simply: why chat, email, phone and search share one knowledge base but each need their own kind of answer.

Blog post thumbnail

Omnichannel AI in customer service means that an AI answers questions on every channel customers use to reach a company, such as chat, email, phone or search. The knowledge behind it is the same across channels. But the answer has to look different on each channel, and sometimes it even needs a different AI model. In this article, we take a closer look at four typical channels.

In Part 1, we looked at how a language model sees the world: as a stream of tokens, whether the input is text, audio or an image. Now it’s time for practice: what happens when real customers ask a question? It quickly becomes clear that the model alone isn’t enough.

This is Part 2 of our fundamentals series, created in collaboration with our Machine Learning Engineer Dr. Tae-Gil Noh. Every example comes from automations running in production at OMQ.

Key Takeaways

  • Medium vs. channel: A medium is the data type a model handles (text, audio, image). A channel is the path a customer arrives on (chat, email, phone, search).
  • Every channel has its own rules: rhythm, expected response time, natural length, and what it can carry.
  • Inform and perform: Almost every support task either answers a question or triggers an action in a system.
  • Model plus design: New models make a lot possible. To work in everyday service, they need good content and thoughtful conversation design.
  • One brain, many shapes: A single knowledge base is the strongest foundation. The real work is shaping its answers for each channel.
  • In practice: myphotobook answers 83% of its service emails automatically with OMQ. At KKT Kolbe, service requests dropped by almost 80% in peak times.

Medium or channel: what’s the difference?

A medium is the data type an AI model processes. A channel is the path a customer arrives on. The two terms often get mixed up, yet this whole article turns on the difference between them.

When we talk about the AI model, we talk about media: the kind of data a model processes – text, audio, images or video.

When we talk about customers, we talk about channels: the way someone gets in touch with a company. Every channel uses a medium, but it also comes with its own rules. Our guide communication channels gives an overview of the most common ones.

On the left, the four media text, audio, image and video; on the right, the four channels chat, email, phone and search with the medium each one uses

A channel uses a medium but brings its own rules: pace, length, links and images.

It’s easiest to see when the same question arrives through different routes:

  • Email: one detailed message, often with screenshots, photos or PDFs attached, then a wait for the reply.
  • Chat: text too, but a quick back-and-forth of short messages and follow-up questions. Because chat lives on a website, it can also show links, forms and file uploads.
  • Phone: a back-and-forth like chat, but with voice only. Nothing to click, nothing to show, and nothing to read back once it’s been said.
  • Help-page search: like Google – type in a few keywords and get matching articles. There’s no conversation.
Remember: The medium gives a channel its character. But in customer service, a channel is always more than its medium.

How do chat, email and phone differ?

Every channel has its own rules. They decide what a good answer looks like there.

ChatEmailPhoneSearch
Rhythmmany short messagesone long message, reply hours laterreal-time conversationsearch, no conversation
Expected response timea second or twominutes to hoursinstantinstant
Natural lengthshortlong and detailedshort, step by stepkeywords in, results out
Hand over a link or fileeasyeasyimpossibleas a result
Receive an imageyesyes, often severalnono

Four cards for chat, email, phone and search showing flow, response time and length, with ticks for whether links can be shared and images received

Every channel has its own rules. They decide what a good answer looks like.

On the phone, nobody can scroll back. Spoken sentences have to be short and come in a clear order. An email, on the other hand, can be long, because it’s read rather than heard.

What this means: every channel needs its own implementation – and sometimes even a different model. A voicebot has to listen and speak in real time. An AI email bot can use a slower model that takes more time to think. So the channel’s rhythm decides not only how the model answers, but sometimes which model fits at all.

People are far more startled by an AI voice than by a chatbot that handles a hundred languages.
Dr. Tae-Gil Noh, Machine Learning Engineer at OMQ

Fancy a quick chat? Our Customer Success Lead Marcel Gerold is happy to show you what this could look like for your channels, with no strings attached.

What tasks does AI handle in customer service?

Almost every customer service task you can automate falls into one of two groups:

  • Inform: The AI looks up an answer in the FAQ or the knowledge base and passes it on. “Yes, you can return it within 30 days. Here’s how.”
  • Perform: The AI gets something done: placing an order, registering someone, starting a return or booking an appointment. For that, something has to actually happen in another system, such as the shop or the calendar.

How far informing can go automatically is shown by the myphotobook case study: OMQ’s AI email bot answers 83% of all service emails there fully automatically.

Either way, what the answer says is settled. How it looks is decided by the channel. And one more pattern keeps coming up: new AI models make some things possible for the first time. But for them to work for customers, they need good content and thoughtful conversation design.

How does the same answer look on different channels?

The same answer looks different in chat than in an email or on the phone: sometimes it’s a click, sometimes text, sometimes a spoken sentence. Three real-world examples show what matters.

1

The simplest case: registration only exists as a link. In chat, it’s a click. In an email, it’s an underlined word. On the phone, there is no link.

The obvious fix would be to text the link by SMS. We tried it and dropped it, for two reasons. First, the voicebot would have to ask for and store a mobile number – extra personal data it otherwise doesn’t need. Second, some people call from a landline, where an SMS never arrives.

What worked in the end was three things together. None of them would have been enough alone:

  1. The model had to get good enough to read out a web address and spell it when needed. That’s harder than it sounds. Until recently, models were surprisingly bad at spelling and counting characters. The famous example: ask an older model how many r’s there are in “strawberry” and it would confidently say “two”. Remember Part 1 on tokens: a model doesn’t see letters, only tokens. “Strawberry” arrives as two or three chunks, not ten letters. It’s like counting the bricks in a wall you’ve only ever seen in a single photo.
  2. The content was simplified: the web address is short and made of ordinary words (such as “/register”) instead of random characters, so it’s easy to say out loud.
  3. The conversation was designed so the AI reads the link slowly, repeats it, and finally spells it out – until the caller has written it down. A voicebot is patient in a way no human agent can afford to be.

The three building blocks model, content and conversation design add up to a link that gets through on the phone; below, the steps read it out, repeat slowly and spell it

Model, content and conversation design: only all three together get a link safely through a phone call.

Good to know: Most of the time, a phonetic alphabet (“R for Romeo”) isn’t needed, because the parts of the URL are everyday words. A second, slower pass is usually enough. Spelling is the safety net, not the main event.

Remember: “Read out a link” isn’t a feature you switch on. It takes a capable model, the right content and good conversation design. Remove any one of them and it breaks.

2

“Where’s my order?”: a looked-up result

Some answers come live from another system, such as an order status, parcel tracking or a fee. The lookup is the same on every channel. Only the presentation changes:

  • Web and chat: a card, a button, a small tracking widget or a table.
  • Email: formatted text and a link.
  • Phone: a spoken summary, with no table and no link: “Your parcel left the warehouse this morning and should arrive tomorrow before noon.”

One order-status lookup branching into three presentations: a tracking widget in chat, text with a link in an email and a spoken sentence on the phone

The lookup stays the same. Only the presentation adapts to the channel.

The clearest example from our own work is a fee calculator that more than a dozen organisations use on their websites. The same calculator also runs in their voicebot. On screen, you read the figure in a small table. On the phone, you hear “Your annual fee is …“. Same calculation, different delivery.

3

“Here’s a photo of my problem”: an incoming image

Some questions arrive as pictures: a photo of a broken part (so the right spare part gets ordered) or a screenshot of an error message (for IT support). That’s completely normal in email, where people put everything into one message: description, PDFs, screenshots. It works in chat too. On the phone, no image can arrive.

Here too, model progress helps. Current models can look at a screenshot and understand it: “This is Windows, this is a browser, the error says such-and-such.” Then they check the knowledge base and reply with a real fix. That wasn’t possible with the previous model generation.

But the capability alone isn’t enough. With email, the AI has to process the attachment before the first reply. Then it either answers directly or prepares a draft for the service team. Chat can do the same but relies more on follow-up questions. On the phone, all that’s left is to have the caller describe the problem or point them to a channel that accepts images.

How does AI get things done on each channel?

If the AI is meant to get something done, it first needs every required detail. How it collects them depends on the channel again.

  • On the website: simply show the form. The customer is already there.
  • In chat: ask step by step. What has already been said in the conversation isn’t asked again. Nobody should have to give their name or address twice.
  • By email: weigh it up. If the email already contains everything, a short confirmation is enough. If details are missing, emails quickly bounce back and forth (“What’s your order number?” … “You still didn’t include it.”). Then it’s kinder to point to the right form, which only needs filling in once.

How much a good form helps is shown by KKT Kolbe. Complaints used to arrive by email, often without details such as the serial number. Today a form on the help page captures everything in one step, and the complete case goes to the service team.

In peak times, there was a drop in service requests of almost 80%.
Thomas Raab, KKT Kolbe

Not every task can or should be handled by the AI alone. Often its greatest value lies in compiling the case. It asks for whatever is missing, completes the case and passes it to a person, who can act right away.

For important actions such as a cancellation or anything involving money, a person deliberately makes the call (see human-in-the-loop). The AI doesn’t replace them there. It makes sure they receive a complete case rather than a half-finished one.

Website, chat and email collect details into one complete case; then the AI handles it right away or a person decides, with handover via call-back, live chat or ticket

Collect the details first, then act or hand over. Nothing gets lost in the handover.

How varied this is shows in the automations running at OMQ: ordering spare parts for coffee machines, parcel tracking, stock checks, appointment booking, account cancellations for bike sharing, and tickets created directly in other systems. Some things the AI answers on the spot, such as a stock check or a fee. Many it collects and forwards. Which approach fits is the company’s decision, based on risk and speed.

The quickest way to see the difference is with your own customer questions.

When does AI hand over to an agent?

Some requests belong with a person. That’s why every channel needs a clear path to one. In chat, this step is also called human takeover. Because the AI has already collected every detail, they travel with the handover. Nobody has to start over.

What the handover looks like is the company’s choice, matched to how its team works:

  • Call-back: for a traditional service that lives on the phone.
  • Live chat: almost real time and much faster than email, while one agent can still handle several conversations at once.
  • Ticket: everything is compiled into one case and worked from a queue.

What matters is a smooth transition. Customers get most answers automatically, on the channel they chose themselves. When they genuinely need a person, they’re guided to the channel the service team works on. And “I want to talk to someone” is honoured the moment it’s said.

Conclusion: one brain, many shapes

A central knowledge base is the best foundation: you write an answer once and use it on every channel. But behind that sits everything we’ve seen above: the sayable link, the result in several formats, the screenshot that gets read, the form for each channel, the compiled case and the clean handover.

That adaptation to every channel is what omnichannel AI in customer service is really about. New models make it possible. Good content and thoughtful conversation design make it work in everyday service.

The next part tackles another exciting question: how can a German knowledge base answer a chat in Chinese, a call in Japanese or an email in Swedish?

Curious how your channels would work together with OMQ? In a no-obligation call, we’ll look at your most common enquiries together.

Read more:

Frequently Asked Questions (FAQ)

What's the difference between a medium and a channel?

Why doesn't AI work the same on every channel?

Can a voicebot share a link?

What do “inform” and “perform” mean in customer service?

Does AI replace people for cancellations?

What's the difference between a voicebot and a chatbot?

What's the difference between multichannel and omnichannel?

What is an omnichannel chatbot?

Glossary

  • Omnichannel AI: AI that answers enquiries on all of a company’s channels from the same knowledge.
  • Medium: the data type an AI model processes, such as text, audio or images.
  • Channel: the path customers use to communicate with a company, such as chat, email or phone.
  • Inform: a task in which the AI gives an answer from a knowledge source.
  • Perform: a task in which the AI triggers an action in another system, such as a shop or calendar.
  • Handover: passing a request, with its full context, from the AI to a person on the service team.

About OMQ

OMQ is a Berlin-based software company that has been building AI for customer service since 2010. Its AI chatbot, AI email bot, AI voice bot and central knowledge base automate support in 32 languages for more than 250 customers. OMQ is a member of Bitkom.

Last updated: November 2026. Reviewed quarterly.

What Is a Token? AI Basics Explained Simply

Blog post thumbnail

What Is a Token? AI Basics Explained Simply

ChatGPT, Claude and the like seem like true all-rounders at first glance. They answer questions, write texts, summarise documents and, on top of that, operate other programs…

What Happens When an AI 'Reads' Your Knowledge Base?

Blog post thumbnail

What Happens When an AI 'Reads' Your Knowledge Base?

When an AI answers a customer using your knowledge base, what does it actually read? Not the whole base. Not even close. For anyone responsible for data protection or IT governance…