An auto-reply confirms the email arrived. A chatbot tries to resolve the case. That is the whole difference, and it decides which of them deserves your money.
The auto-reply costs practically nothing and takes ten minutes to set up. The chatbot costs money every month and requires someone to keep it current. Below is when one is enough, when the other pays off, and how you work out which applies to you.
There are four levels, not two
Putting chatbot against auto-reply is a simplification. In reality there are four steps, and most small firms jump from the first to the last without looking at the two in between.
| Level | What it does | To set up | Stops working when |
|---|---|---|---|
| 1. Auto-reply | Confirms receipt, states a response time | 10 minutes, SEK 0 | Never — but it solves nothing either |
| 2. Auto-reply with content | Answers the three most common questions inside the confirmation and links onwards | 2–4 hours, SEK 0 | The text goes stale and nobody notices |
| 3. Rule-based bot | Buttons and decision trees, always the same path | 1–2 weeks, SEK 300–1,500/mo | The customer wants something that is not a button |
| 4. AI bot on your material | Reads the question in free text and answers from your own documents | 3–6 weeks, SEK 500–3,000/mo plus the build | The material is poor — then it invents |
Level two is the most underrated. An auto-reply that says "we answer within one working day — if this is about delivery time, here it is; if it is about returns, do this" removes a surprisingly large share of the emails, costs nothing, and can never answer wrongly in a way that harms you.
How to tell which level you need
Volume decides, not the technology:
- Under 20 cases a week. Level 1 or 2. A bot would not have time to learn anything before you had answered yourselves.
- 20–100 cases a week, mostly the same questions. Level 3 or 4. This is where the time starts to cost real money.
- Over 100 cases a week, or customers in other time zones. Level 4, and it pays off quickly.
There is one exception the other way. If you sell something where the customer abandons the purchase at ten in the evening because nobody answers, a bot can pay off well below twenty cases a week. Then it is not about saved time but about orders you would otherwise not get.
What an AI bot actually requires from you
The technology is the easy part. Four things have to be in place, and if one of them is missing the result is worse than an ordinary auto-reply:
- Source material. Written answers to what keeps coming back: terms, delivery times, prices, exceptions. A bot cannot know anything you have not written down.
- An owner. A named person who reads the logs once a month and corrects what went wrong.
- A route to a human. The customer must always be able to move on with one click, without guessing passwords or repeating the whole case.
- An honest "I don't know". A bot that would rather guess than admit it does not know will sooner or later promise something you have to honour.
The most common mistake
Building level four when level two would do. You see it clearly at firms where the bot answers ten questions a week beautifully and costs two thousand a month to own. The reverse exists too: a rule-based bot with forty buttons nobody can be bothered to click through, built because someone wanted to avoid the monthly cost of a real model.
A third mistake is more expensive than both. A bot that answers questions about price, contracts or the right of withdrawal does so in your voice, and what you promise in the chat stands. Do not let it comment on such things. There it should collect details and hand over.
Personal data in the chat
A chat that collects names, email addresses or order numbers processes personal data, and the same rules apply as for the rest of the business. Three practical decisions go a long way: pick a supplier with data centres inside the EU, decide how long logs are kept and then delete them automatically, and enter the processing in your records. Never ask for a personal identity number in a public chat.
What to do next
Take the last fifty cases and sort them into three piles: what a better auto-reply would have solved, what needs a lookup in your systems, and what needs a human. The size of the piles tells you which level to build, and it takes half an hour to find out.
If the middle pile is the largest, level four is what applies, and we are glad to help — see AI for Business for what such an assignment looks like, the pricing page for what operation costs, or the courses if you would rather build it yourselves.
Frequently asked questions
Is a chatbot the same as putting ChatGPT on the website?
No. A raw language model answers from everything it read during training and knows nothing about your company. A customer service bot is the same model framed by your own documents and clear rules for what it may say. Almost all the work lies in that framing.
How often does the bot answer wrongly?
It depends almost entirely on the source material. With well-written answers to the questions actually being asked, a bot normally handles the bulk of recurring cases unaided. With thin material that figure drops fast, and the errors are harder to spot because they sound reasonable. Read the logs weekly for the first month.
Can we start with an auto-reply and switch later?
Yes, and that is usually the right order. Writing a good auto-reply is the same work that later becomes the bot's source material, so nothing is wasted. You also find out how many cases disappear from clearer text alone.
Do we need a bot in both Swedish and English?
An AI bot handles both languages from the same source material, so the extra cost is small. A rule-based bot has to be built twice, because every button and every path needs its own translation. If you have customers outside Sweden, that argues for level four.
What does it cost to do nothing?
Count cases times five minutes times your internal hourly cost. Thirty repetitive emails a week is ten hours a month, so around SEK 4,000 at SEK 400 an hour. If you are under five hours a month, it is too early to build anything.