AI customer support, explained without the hype.
AI customer support means using language models to read incoming customer messages and write replies, on chat widgets and email. Done well, it handles the bulk and steps back when it shouldn't be the one talking. Done badly, it hallucinates policies and breaks trust.
AI customer support reads a customer message, looks up the relevant facts from your own knowledge sources, drafts a reply that cites those sources, and escalates to a human when it is unsure. The good ones are honest about what they do not know.
- ✓Reads your website, FAQs and documents, and answers from those sources only. Not the open web.
- ✓Cites the page or paragraph it used, so operators can verify every claim.
- ✓Escalates with a clean summary when confidence drops, instead of bluffing through an answer.
- ✓Lives on email and the chat widget. No customer logs into your support tool, only your team does.
What AI customer support actually is
AI customer support is the practice of letting a language model read inbound customer messages and write replies on your behalf. The model is grounded in your own content, your help center, product pages, return policies and FAQs, and produces answers in your tone of voice. Good systems cite the source they used. Bad systems guess. The difference between the two is whether you trust the AI in front of your customers.
Where it sits in your stack
Most teams plug an AI agent into two surfaces: the chat widget on the website, and the support inbox that handles email. The agent reads the message, looks up the relevant policies and product facts, writes a reply, and either sends it or hands the conversation to a human. The operator inbox stays the same, your team works the conversations the AI could not finish, with the full context already attached.
What separates a good system from a chatbot
Traditional chatbots match keywords to scripted answers and break the moment a customer phrases a question slightly differently. AI customer support reads the meaning. It can handle a refund question, a shipping question and a product question in the same conversation. Crucially, it knows when it does not know. A grounded AI agent will say so and escalate. A guess-bot will invent a return window and you find out from a chargeback.
How to evaluate one in a week
Skip the demo, paste your URL. A serious AI customer support product can ingest a website in minutes and start drafting replies in your sandbox. Test it with twenty of your real recent tickets. Look for three things: does it cite the source it used, does it refuse cleanly when the answer is not in your content, and does it sound like your brand. If any of those three fail, the deflection numbers on the marketing page do not matter.
Where Keloa fits
Keloa is AI customer support and sales, hosted in the EU, with citations on by default and an honest handoff when the AI is unsure. Three channels live today: chat widget, email, and Slack — the AI agent answers end to end in all three (including DMs, @mentions, and threads inside Slack). WhatsApp, Instagram, and Messenger are on the roadmap. Starter is free with fifty replies a month, no credit card. Growth is €49 a month for fifteen hundred replies, Business is €149 a month for six thousand replies, and Scale is €449 a month for twenty-five thousand. Top up at €30 per extra thousand.
Frequently asked questions about AI customer support.
Is AI customer support actually worth it?
For most small and mid-sized teams, yes. The math is simple: the AI handles the repetitive sixty to seventy percent of tickets, your team handles the rest. You pay for replies, not for seats, so the cost scales with traffic instead of with headcount. The catch is that it only works if the AI is honest about what it does not know. A guess-bot that invents policies makes things worse, not better.
Will AI customer support replace my team?
No, and serious vendors will tell you the same. AI customer support absorbs the volume floor: the password resets, the shipping ETAs, the return policy questions. Your team is freed up for the conversations that need judgment, empathy, or a decision the AI is not allowed to make. Most teams that go live with Keloa keep their staff, redirect them to upsell, retention, and product feedback.
How long does it take to set up?
If your help center is online, ten minutes is realistic. Paste your URL, the AI reads your site and builds the initial knowledge base. Test in a sandbox, tweak the tone, drop the widget snippet on your site or forward your support inbox. Going from sign-up to first live AI reply inside ten minutes is the common path on the Starter plan.
What happens when the AI does not know the answer?
It escalates. A good AI customer support system tracks its own confidence per reply. Below the threshold, default seventy percent in Keloa, it does not send. It writes a short summary of what the customer asked and what it could not find, and the conversation lands in your team's inbox with that context attached. Your operator picks it up and answers properly.
Can it handle multiple languages?
Yes. Modern AI customer support is multilingual by default. Keloa detects the customer's language from their first message and writes the reply natively in that language, not a translated draft. One hundred plus languages are supported, with Dutch and English the most polished out of the box.
Is AI customer support GDPR-compliant?
It can be, but it is not automatic. Check three things: where the data is stored, who the sub-processors are, and whether personal data is redacted before any AI call. Keloa stores data in the EU, signs a Verwerkersovereenkomst (DPA) per customer at no extra cost, and strips emails, phone numbers, IBANs and card numbers before any model call. The sub-processor list is public.
Try AI customer support on your own site.
Free Starter plan, fifty AI replies, no credit card. Paste your URL and see your AI colleague answering real customer questions in ten minutes.