24/7 AI customer service: how it actually works and where it breaks
Why "AI customer service" is different now
Old chatbots were bad because they didn't know anything about you. A modern AI customer service agent has your order database, your returns policy, your product catalogue, your calendar and your CRM plugged in. When a customer asks "where is my order?", the agent actually looks and answers with the tracking number.
What it handles alone
Based on shipped agents, the breakdown for a typical Finnish SME:
- Order status: 100% of the time, reads courier API
- Return and refund questions: 90%, reads policy plus the specific order
- Business hours, location, contact info: 100%
- Product availability and specs: 85%, reads catalogue
- Simple troubleshooting ("resetti ohjeet"): 80% if you have decent docs
- Booking or rescheduling: 90% if the calendar is connected
What it escalates
- Angry customers - detected by tone, handed to a human immediately
- Complaints about a person by name
- Refund requests above a threshold you set
- Anything the agent's not confident about (below a configurable score)
- Legal, safety or medical questions
The escalation isn't "sorry, I don't know". It's "let me connect you with someone - here's what we discussed so far", passed to the human with full context.
The two-week ramp-up
We don't turn 24/7 AI service on cold. The pattern that works:
- Week 1-2: Agent reads incoming messages and drafts every reply. Human sends. This trains it on your voice and catches gaps.
- Week 3-4: Agent sends the top three question types on its own (order status, hours, availability). Everything else still drafted.
- Week 5+: Confidence-based cutover. If the agent scores itself above the threshold, it sends. Below, it drafts.
What Finnish customers actually think
Surveying users of two live agents we shipped in 2026:
- 72% preferred an immediate AI answer over a 4-hour email wait
- 84% said they'd rather talk to AI than sit in a phone queue
- 61% wanted the AI to say clearly it was AI (we do, always)
- 12% said they'd feel misled if the AI pretended to be human
::: callout Our rule
Agents always identify themselves as AI on first message. We think it's the only workable long-term posture.
:::
The economics
For a Finnish SME with 200 - 800 monthly customer messages:
- Build: €3,000 - €6,000
- Runtime: €50 - €120 / month
- Removed workload: usually one part-time customer service role
- Payback: 4 to 8 months
See [AI agent pricing in Finland](/blogi/ai-agentti-hinta-suomessa) for the full price grid.
Where it still breaks
- Regional Finnish dialects: some slang trips up even good models. We add a phrase glossary.
- Photo attachments ("here is the broken part"): possible but expensive. Handoff is often cheaper than automation.
- Very personal complaint handling: technically possible, ethically better to hand off.
- Cross-language mid-conversation: FI/EN switches work fine. FI/RU/EE gets shakier - test carefully.
What we ship
Every customer service agent we ship comes with:
- Handoff button always visible ("Puhu ihmiselle" / "Talk to a human")
- Full conversation log emailed to the assigned agent on handoff
- Weekly report of top 20 questions and the agent's confidence per category
- Kill switch you can flip in one click if something goes wrong
Ready to price your case? See our [AI customer service agent page](/tekoalyagentti-asiakaspalveluun) or [request a proper quote](/yhteystiedot).
Published · Kaspar Eevald
Bluefour – Helsinki digital agency | hello@bluefour.fi | +358 45 850 4236