Chatbot, voice bot, AI agent: what's actually different

Three different things get sold under roughly the same words: chatbot, voice bot, AI agent, automation. They solve different problems, cost money in different shapes, and fail in different ways. This post sorts them into three categories you can actually use, so you can work out which one fits before anyone shows you a demo.

The three categories, one line each

• Text chatbot — answers typed questions, usually on your website or in a messaging app. The customer is reading.
• Voice AI agent — picks up the phone, talks, and does something with what it hears. The customer is talking, in real time, and will hang up if it's slow.
• Workflow automation — no customer-facing conversation at all. It moves information between systems when something happens: a form lands, so a text goes out and a row appears in a spreadsheet.

Nearly every "AI" product sold to a small business is one of those three, or two of them bolted together. The marketing word on the website tells you almost nothing. The channel and the timing tell you everything.

Text chatbots: cheap, patient, easy to ignore

A chatbot sits in the corner of a web page or inside WhatsApp and Instagram, and replies to typed messages. Older ones were decision trees — buttons and canned answers. Newer ones run on a language model and can answer in their own words from whatever you've fed them: your prices, your service area, your FAQ page.

What chatbots are genuinely good at: answering the same twelve questions forever without getting bored. Do you cover Luton. Are you Gas Safe registered. How much is a full head of highlights. Can I bring my dog. These are questions people would rather type than ring up about, and a chatbot gives an instant answer at two in the morning for pennies.

What they're bad at: being noticed. A website chatbot only works on people who are already on your site and already minded to engage with a little pop-up. If your enquiries arrive by phone, a chatbot catches none of them. It also can't chase. It waits. If the visitor closes the tab mid-conversation, that's the end of it unless you captured an email address early.

The messaging-app versions are a different beast because the conversation persists — the customer can come back to it tomorrow and you can reply. Meta's free in-WhatsApp agent is worth knowing about if that's where your customers already are; we covered what it does and where it stops.

Voice AI agents: the hard one, and usually the valuable one

A voice agent answers the phone and holds a spoken conversation. Technically it's three things chained together: speech-to-text, a language model deciding what to say, and text-to-speech saying it. Each link adds delay, and delay is the whole game. A chatbot can take four seconds to reply and nobody minds. Four seconds of silence on a phone call and the caller says "hello? hello?" and hangs up.

That latency problem is why voice costs more per conversation than text and why the engineering is fussier. It's also why the category has consolidated around a handful of platforms rather than everyone rolling their own — we went through who's building what if you want the landscape.

The reason voice is usually worth more to a small business is blunt: phone calls are where the money is. Someone with a leaking cylinder rings. Someone whose tooth has cracked rings. They don't fill in a form. If you're on a roof or mid-colour, that call goes unanswered, and an unanswered call is a job that goes to whoever picks up next. The arithmetic on missed calls is uncomfortable once you actually do it.

The honest weakness of voice agents: accents, bad lines, background noise, and people who talk in long loops. A chatbot can ask the customer to rephrase and nobody's embarrassed. On a call, a misheard postcode is a real problem. If you want the caller's experience in detail, there's a walkthrough of what people actually hear.

Workflow automation: no conversation, no intelligence required

The third category gets lumped in with AI because the tools that sell it now have "AI" on the homepage, but most of what people buy is a trigger and an action. When a quote is accepted, send a thank-you text and create the job in the calendar. When a review comes in below four stars, email me. When an invoice goes 14 days overdue, chase it.

These are the cheapest wins in the whole list and the least glamorous. They don't need a language model at all. If your problem is "things fall through the cracks between my inbox, my diary and my invoicing", you want automation, not a bot, and buying a chatbot will not help you even slightly.

Where it gets muddy is that the word agent has been stretched to cover automations that make a decision or two along the way. If you want the strict version of the definition, we wrote what "AI agent" actually means — the short version is that an agent completes a multi-step task on its own, choosing the steps, rather than running a fixed script.

Side by side

The three questions that sort it quickly

1. Is a human waiting right now? If yes and they're on the phone, you need voice. If yes and they're typing, chat. If nobody's waiting, automation.
2. Does the job need judgement, or just steps? Fixed steps every time means automation, and it'll be cheaper and more reliable than anything with a model in it. Judgement — working out what someone wants, deciding whether it's urgent — means you need the model.
3. What happens after the conversation ends? This is where most people get caught out. A bot that has a lovely chat and then does nothing with it has just created a transcript. The useful part is the handover: the job logged, the booking made, the text arriving on your phone with the caller's number and what they want.

Point three is worth dwelling on. Plenty of tools are strong at the conversation and weak at what follows. When you're being shown a demo, ask to see the output — the actual message or calendar entry or CRM record — not just the chat window.

Where the labels genuinely blur

A phone receptionist that answers a call, qualifies the job, checks your diary and texts you for approval is doing all three categories at once: voice conversation, decision-making, and automation at the end. Calling it a "voice bot" undersells it; calling it an "AI agent" is accurate but tells a buyer nothing concrete.

So ignore the noun and ask the three questions above. The useful distinction isn't chatbot versus agent. It's: which channel, how fast must it respond, and what does it leave behind when it's finished.

When none of these is your answer

Three situations where buying any of this is a mistake.

Your problem is volume, not handling. If you get four enquiries a week and you answer all four, no bot will help. You have a marketing problem. Automating a quiet phone line just gives you a quiet automated phone line.

Your conversations are genuinely sensitive. Safeguarding calls, bereavement, complaints that have already gone wrong, anything where the caller is distressed. A machine answering first makes those worse, and a fair number of people are actively fed up with being routed to AI — more on that here. Route those to a person, even if it means voicemail with a promise of a call back within the hour.

Your underlying process is a mess. If you don't know what you charge, what your availability is, or who does what, a bot will broadcast that confusion faster and to more people. Sort the process on paper first. Automation amplifies whatever it's pointed at.

You need something done once a month. A five-minute manual task done monthly is sixty minutes a year. Building, testing and maintaining an automation for it will cost you more than that in the first week. Automate the daily things.

What to ask any vendor, in any category

• What's the pricing shape — per minute, per message, per run, per seat — and what happens in a busy month?
• Where does the output land? Show me the text, email or calendar entry it produces.
• What does it do when it doesn't know the answer? Guessing is worse than handing over.
• Who maintains it when my prices or hours change — me, or you?
• Can I listen to or read real transcripts from the first week, before I let it loose unsupervised?

That last one matters more than any feature list. The first week of transcripts tells you whether the thing understood your customers or just talked at them — and that's the only test that counts.