Before you picture a sci-fi receptionist, picture your inbox at eleven at night: three people asking about your opening hours, one asking whether an item ships abroad, one complaining that an order arrived late. None of those needed you. That is the gap an AI chatbot fills for a business with a handful of employees — not a replacement for judgment, but a filter that keeps judgment for the messages that deserve it.
What a bot actually replaces — and what it must never touch
A full-time support agent costs somewhere between $3,500 and $5,500 per month, and for most operations under fifty employees that line item was never realistic. Platforms like Intercom Fin, Zendesk AI, and open-source frameworks running on modest cloud instances now absorb the bulk of routine inquiries for businesses that size — the original guide puts it at sixty to eighty percent.
"Absorb routine" is the operative phrase. A bot answers "what time do you open?" four hundred times a day without fatigue, typos, or a raise request. It should not negotiate refunds, handle an angry escalation about a damaged delivery, or improvise policy. Hide the human escape hatch and every point of deflection you gain turns into resentment; the healthiest setups deliberately route ambiguity to a person.
Four ways to deploy one — pick by how you already sell
Rather than comparing feature grids, start from where your customers already talk to you:
- If you run a full support suite (Intercom Fin, Zendesk AI Agent), the AI rides inside the seats — budget roughly $55 to $115 per agent per month depending on plan, or around $39 per seat to start with Fin alone.
- If you run an online store, stay in-platform: Shopify Inbox with AI costs nothing beyond the plan itself, and Gorgias with AI starts at $60 per month for merchants who want tighter order integration.
- If a developer is on retainer or you enjoy server admin, open-source frameworks cost software plus hosting; expect ten to twenty hours of setup and ongoing maintenance, with a hosted option near $30 per month at small usage levels. Rasa remains the hardest of the bunch; builder-style tools sit friendlier for non-developers.
- If your sales happen on Instagram, WhatsApp, or Facebook — fashion, food, local services — a social-first bot is the natural fit: ManyChat's free tier covers up to 1,000 subscribers, paid plans start at $15 per month, and heavier setups begin near $79 per month.
The category matters less than honesty about your own technical patience. The most expensive chatbot is the one deployed in a tier you will never maintain.
The math that decides whether this is worth it
Take a business handling about 500 conversations a month. On a Shopify plan, the AI inbox adds nothing to the bill — the plan itself runs near $468 a year — against a human-agent equivalent you would price at $42,000 to $66,000. The social route through ManyChat lands between $180 and $540 a year.
Now the working case: a bot at $100 per month deflecting sixty percent of those 500 conversations saves roughly ten to fifteen hours of staff time monthly. At $25 per hour that is $250 to $375 back — before counting faster replies, night coverage, and calmer customers. That is the whole business case in four lines, and it either works for your volume or it does not.
A week of setup, not a quarter
Scope discipline beats tool choice, and the original roadmap compresses to seven days:
- Days one and two: list the fifteen to twenty questions customers actually repeat. Cover the eighty percent that is routine volume; ignore the exotic scenarios.
- Days three and four: connect the platform to your store or social inbox and load those questions as intents.
- Day five: write answers in your voice, then attack the top five failure cases — ambiguous phrasing, misspellings, multi-part questions.
- Day six: soft launch to twenty to thirty percent of incoming traffic.
- Day seven: full launch with a fifteen-minute daily review for the first two weeks.
A seventy-percent-quality bot that improves weekly beats a ninety-percent project that ships next quarter.
The numbers to watch after launch
Deflection rate is the headline: fifty to seventy percent is healthy for small-business support. Below forty percent means poor training data or genuinely complex use cases; above eighty percent usually means the bot is blocking legitimate escalations. Pair it with speed — a good implementation cuts seventy to ninety percent off resolution time for routine questions — and one rule for satisfaction: if the bot's CSAT trails the human benchmark by more than ten points, it is creating friction rather than removing it. For stores, watch conversion on sessions where the bot answers product questions in real time; a ten to twenty percent lift for engaged shoppers is a realistic ceiling, not a promise. After thirty days, review deflection, satisfaction, and hours saved together — any single number alone always lies.
Four failures that make customers hate bots
- Hiding the human option. Every flow needs a visible path to a person.
- Over-training before launch. The ship-early approach above exists precisely because of this trap.
- Ignoring multilingual needs. A bot that answers Vietnamese customers in English — or the reverse — burns more trust than having no bot at all.
- No review routine. The short daily habit is not optional; unwatched bots drift fast.
Where the full breakdown lives
The original walkthrough pairs every category above with per-platform pricing, a transparent cost table for the 500-conversation month, the ROI metrics worth tracking, and the complete day-by-day roadmap — read the full guide on tanit365.com.

