Search for conversational AI in Estonia and you get three different products wearing one label. A chat bubble in the corner of a website. A voice agent that answers the phone. An internal assistant that reads your documents and answers staff questions. They cost different money, fail in different ways, and fix different problems. Buying the wrong one is the most common expensive mistake we see, and the sales decks do not help, because all three are pitched with the same sentence.
What conversational AI means once the pitch stops
Conversational AI is any system a person talks to in plain language and gets a useful answer from. That is the whole definition. It covers text and speech, customers and staff, scripted flows and generative models. The category tells you almost nothing about whether it will pay.
The distinction worth holding is between a system that answers and a system that acts. A model that recites your opening hours has answered. A system that checks the calendar, holds the slot, writes the booking to your CRM, and sends the confirmation has acted. The second one earns money. The first one is a nicer FAQ page with a monthly invoice attached.
So the first question is not which technology. It is which conversation your business keeps losing, and what closing it is worth. We wrote the short self-serve version of that exercise in the closed loop score framework.
The three things Estonian companies mean by it
Almost every enquiry we get is one of these three. Naming yours out loud shortens the project by a month.
The website chatbot
A text assistant on your site. It answers product questions, qualifies a visitor, and hands the good ones to a human. It is the cheapest of the three to stand up and the easiest to do badly. A chatbot trained on a thin website repeats the website, which the visitor already read before they opened the chat.
It pays when your traffic is real and people leave with an unanswered question. It does nothing at all if your problem is that nobody visits, or that the buyer never uses the site because they pick up the phone instead.
The voice agent on the phone
A system that answers a call, understands the caller, and finishes the task. It books, reschedules, takes the order detail, or routes to a human with the context attached. This is the one that most often pays for itself in Estonian service businesses, because the phone is where the urgent buyer goes and the phone is where small teams run out of hands.
The arithmetic is unglamorous. Count the calls that ring out during busy hours, evenings, and weekends. Multiply by your close rate and your average job. That number is usually larger than the quote for fixing it, which is the point we make in what a missed call costs your business. A voice agent is not the same product as the old phone menu, and the difference is covered in AI voice agent vs IVR.
The internal assistant
An assistant that reads your own material: contracts, price lists, technical manuals, past quotes. Staff ask it questions instead of interrupting the one colleague who knows. This is the quietest of the three and often the highest value in firms where knowledge sits in a few heads.
It is also the one with the sharpest data question. Whatever it reads, it can repeat. Deciding what goes in the index is a governance decision, not a technical one, and it belongs to you rather than to your vendor.
Choose by where the money leaves
The honest way to pick is to follow the loss. Where does a conversation start and then die? If it dies on the website, build the chatbot. If it dies on the phone, build the voice agent. If it dies inside the company, in the gap between a question and the person who knows, build the internal assistant.
The failure pattern we see most often is a company that buys the visible one because it is visible. A chat bubble looks like progress on the homepage. Meanwhile 30 calls a month ring out after 17:00 and nobody counts them, because a missed call leaves no trace in any system you own. What you cannot see does not appear in the board pack, which is exactly why it keeps costing money.
What breaks in Estonian
Estonian has roughly a million speakers. Speech models are trained where the data is, so Estonian gets treated as a minor language, and that shows up in ways a demo will never surface.
Case endings are the first thing to go. A model that hears the nominative and needs the partitive produces something that sounds almost right and reads as foreign. Place names are the second. Street names, parish names, and small town names are exactly the words a general model has seen least of, and they are exactly the words a booking depends on. Company names are the third, especially the ones that mix Estonian and English.
None of this makes Estonian conversational AI a bad idea. It makes the tuning the job rather than a finishing touch. The practical defence is simple: ask for a live call in Estonian, with your own address and your own product names in it, before any money moves. A recorded reel proves nothing, because reels are recorded until they are good.
The stack underneath
A working setup is four parts, and none of them are exotic. There is a speech layer that hears and speaks. There is a model that decides. There is the plumbing that connects to your calendar, CRM, and stock. And there is the escalation path for the moment the system is out of its depth.
We build the voice side on Vapi with ElevenLabs for speech and Twilio for the telephony, so your existing number keeps working. The connecting work runs on Make when a hosted builder your team can read is the right answer, or self-hosted n8n when the data has to stay under your control. The part that decides whether any of it is useful is the integration, and we set that out in the voice agent CRM integration guide.
The fourth part is the one vendors skip. Every model is wrong sometimes. What the system does at that moment, hand to a human with context or fail quietly and lose the customer, is a design decision made before the build. Ask about it early.
What it costs
Market rates for AI consulting sit between 150 and 350 dollars per hour for senior independent talent, and a scoped build runs from 20,000 to 150,000 dollars. Those bands cover the whole discipline, which is why they are so wide. The per-conversation running cost is a separate line, and we broke it down in the voice agent cost guide.
Our own model is deliberately plain. The audit is free and takes about 30 minutes. Ongoing build-and-run starts at 600 euros per month plus VAT, and the exact figure comes after the audit, based on call volume, integrations, and workflow count. We quote after looking, because quoting before looking is guessing with a straight face. The code and the accounts stay in your name, which means you can leave without your operations stopping.
The rules you inherit by being in the EU
Estonia is in the EU, so the EU AI Act applies to what you build. It sorts systems into four risk tiers, from unacceptable through high risk and limited risk to minimal, and it sets transparency duties for systems that talk to people. In plain terms: a customer should know they are speaking to a machine.
Most customer service and booking automation lands in the lower tiers. Hiring, credit decisions, and biometric identification do not. For the governance side, the NIST AI Risk Management Framework is the reference most serious teams borrow from. It is voluntary and American, and it remains the clearest public description of how to think about AI risk. A consultant who can name a framework has read one.
When the answer is neither
Sometimes the right recommendation is to build nothing. If the process behind the conversation is broken, automating it produces a faster broken process and a louder complaint. If your calendar is not trustworthy, a booking agent will book into fiction. If nobody has agreed what your prices are, no model will invent a consistent answer.
We say this on audits more often than people expect, and it is the cheapest advice we give. Fix the calendar, agree the price list, then automate. The order matters more than the tooling.
How we work
Audit first. A free session of about 30 minutes where we walk the business and name, in plain terms, where conversations are dying and what closing them is worth. Then we build the specific fix: a voice agent, an automation, or the connected data layer underneath. Then we run it, because the version that works in month one and the version that works in month twelve are not the same system.
For a transport company in Tartu that work started with the website rather than a bot. The structured data went from 4 valid schema blocks to 75 and the hero image dropped from 60 KB to 14 KB, which is unglamorous work nobody had done. The Estonian-language voice agents and the workflow automations came after, once the foundation was worth talking to.
You can read how we work as an AI consultancy, compare the kinds of provider you will meet here in AI consultants in Estonia, see the Estonian side of the business at kratt Eestis, or skip ahead and book the audit through contact.
What is conversational AI for a business?
Conversational AI is any system a customer or a colleague talks to in plain language and gets a useful answer from. In practice it arrives in three shapes: a website chatbot, a voice agent on the phone, and an internal assistant that reads your own documents. The commercial question is not which label you buy. It is whether the system can finish the conversation without a human, because a system that only answers is a nicer FAQ page.
Do I need a chatbot or a voice agent?
Look at where the business loses the conversation. If people leave your website with a question, a chatbot pays. If calls ring out during busy hours, evenings, and weekends, a voice agent pays and a chatbot changes nothing. Most Estonian service businesses we audit lose more on the phone than on the site, because the phone is where the urgent buyer goes.
Does conversational AI work in Estonian?
It works, and it needs more tuning than English. Estonian has roughly a million speakers, so speech models treat it as a minor language. Case endings, place names, and company names are where it slips. Ask any vendor for a live call in Estonian rather than a recorded demo reel, because the demo is usually built and tested in English.
How much does conversational AI cost in Estonia?
Market rates for AI consulting run from 150 to 350 dollars per hour for senior independent talent and 20,000 to 150,000 dollars for a scoped build. Our model is different: the audit is free and takes about 30 minutes, and ongoing build-and-run starts at 600 euros per month plus VAT, quoted after the audit because the honest number depends on call volume, integrations, and workflow count.
Do I have to tell customers they are talking to AI?
Yes. The EU AI Act sets transparency duties for systems that interact with people, so a customer should know they are speaking to a machine rather than a colleague. Estonia is in the EU, so this applies here. It is also the setting that costs you the least: the disclosure line takes two seconds and removes the complaint that follows when someone finds out later.
If you want this scoped to your business instead of a generic comparison, start where we always start: book the free audit. Thirty minutes, your call logs and your website, and a straight answer on which of the three is worth building first.