Home / Blog / Global playbooks
AI Chatbot for Customer Service: What It Costs and When It's Worth It
Somewhere between "chatbots are magic" and "chatbots are the reason I hate this airline" is the truth: a well-built AI support bot deflects most repetitive questions and pays for itself quickly — and a badly built one actively costs you customers. The difference is scope and honesty. Here's what the good version costs in 2026.
The three tiers, priced
| Tier | What it is | Setup cost | Monthly |
|---|---|---|---|
| Off-the-shelf widget | SaaS bot trained on your site/FAQ pages | $0–$500 | $20–$100 |
| Custom RAG bot | Built on your docs, policies and product data, with designed escalation | $1,500–$8,000 | $30–$150 (hosting + API) |
| Full support agent | Takes actions: order lookups, bookings, refunds via your systems | $8,000–$30,000+ | $100–$500 |
Market estimates. RAG = retrieval-augmented generation: the bot quotes from your real content instead of improvising, which is what keeps answers accurate.
The ROI math (do this before buying anything)
Count last month's support conversations. Classify a sample of 50: what percentage were routine — shipping times, pricing, "how do I reset…", opening hours, return policy? For most small businesses it's 50–70%.
- 300 conversations/month × 60% routine × 6 minutes each = 18 hours/month of a human's time.
- At $25/hour, that's ~$450/month — against $50–$150/month of bot costs after setup.
- Add the invisible win: those routine answers now arrive at 2am on Sunday too, when a decent share of buying decisions happen.
If your routine share came out under 30%, stop — a chatbot is the wrong purchase; fix your FAQ page instead.
Why most chatbots fail
- No escalation path. The cardinal sin. Every conversation needs a visible exit to a human.
- Stale sources. The bot learned your 2024 prices. Assign an owner and a monthly refresh, or accuracy rots silently.
- Pretending to be human. Users always find out, and they punish you for it. Label the bot; nobody minds a good robot.
- Scope greed. Launch on your top 20 questions done perfectly, not 200 done vaguely. Expand from transcripts.
A sane implementation path
- Mine your inbox. Your last 200 support emails are the training spec — cluster them into topics and rank by frequency.
- Write canonical answers for the top 20 topics. This document improves your human support too — it's pure upside.
- Pick the tier the math justifies (usually: start off-the-shelf, upgrade to custom RAG once volume proves it).
- Design the handoff: what the bot says when unsure, what context transfers, response-time promise for escalations.
- Review transcripts weekly for the first month. Fix the top three failure patterns; accuracy climbs fast.
We build custom RAG bots with designed escalation as part of our AI & automation consulting — and we'll tell you honestly if the off-the-shelf tier is all you need. Ask us for the audit and bring your inbox stats.
Frequently asked questions
Will a chatbot make my business feel impersonal?
Only if it pretends to be human or traps people. The pattern customers actually like: instant accurate answers to routine questions, an always-visible "talk to a human" path, and the bot handing over full conversation context so nobody repeats themselves. Deflect the repetitive 60%, and your humans give better service on the 40% that matters.
How long does a custom chatbot take to launch?
An off-the-shelf widget trained on your site: a day or two. A custom RAG build on your docs, policies and product data with escalation flows: typically two to five weeks including testing. The timeline driver is almost always how organized your source content is.
What happens when the bot doesn't know the answer?
That must be designed, not left to chance: the bot should say it doesn't know, collect the question and contact details, and create a ticket or live handoff. A bot that invents answers is worse than no bot — insist on seeing the escalation flow in any demo.