Delivering a working AI customer-service bot for a small business in 3 days, at a $99 entry price point. Here is the actual delivery playbook, not vendor slides.
Last week I wrote about why small businesses should look at AI customer service in 2026. The most common follow-up question was: "What does the cheapest real engagement actually look like?"
So here is exactly what goes into a $99 (¥699) entry-tier delivery.
Who this tier is for
A shop, studio, or small online store with under 200 customer conversations per month, on one channel (WeChat, Instagram DM, TikTok DM, or a website widget). Restaurants, nail salons, tutoring centers, local services, small e-commerce.
It is not for call centers, multi-location brands, or anyone needing CRM/ERP integration. Those start around $700-$1800.
What is included (and what is not)
Included:
- A knowledge base of 20-30 real Q&A pairs, written from the business's actual materials
- One channel connected
- Industry-tuned tone and greeting
- Human handoff for complaints and big orders
- 7 days of tuning after go-live
Not included: multi-channel, analytics dashboards, API integrations, phone/voice. Scope creep is the number one reason cheap AI projects fail — keep the first build deliberately small.
Day 1: knowledge mining (this is 80% of the value)
The model is rarely the differentiator. The knowledge base is.
- Export or collect the last 1-2 months of real chat records
- Cluster questions — in practice 60-80% are the same 20-30 questions: price, hours, location, sizing, returns, booking, parking
- Write each answer in the owner's voice, short and concrete. "How much is it?" must get a real answer with a price range, not "please contact customer service"
- Mark the 5-10 questions that must escalate to a human: complaints, refunds under dispute, bulk orders, anything the answer sheet doesn't cover
A bot with an empty or vague knowledge base hallucinates and is worse than no bot. A bot fed real chat records answers like the owner would.
Day 2: build, connect, fail-test
- Build the bot on a mainstream agent platform (Coze/Dify/FastGPT-class) — no code needed
- Upload the Q&A as the knowledge base, set retrieval-first answering so it does not invent facts
- Write a strict fallback: "Good question, let me have the boss confirm — one moment" → notify human. Never bluff
- Connect the one channel
- Run 30 adversarial tests: rephrase questions 5 different ways, ask half-questions, typos, angry messages, questions slightly outside scope. Fix every wrong or evasive answer
Day 3: handoff, go-live, measurement
- Train the owner (15 minutes): how to read the handoff queue, how to tell when the bot answered wrong
- Agree on the one metric that matters this month: self-service resolution rate — conversations fully closed without a human. Target: 60%+
- Go live, then watch real conversations daily for 7 days and patch the knowledge base as new questions appear
Why it is priced at $99
It is a proof-of-value wedge, not a profit center. At this price the business owner can say yes without a committee. When resolution hits 60%+ and they want the second channel, booking flows, or customer records, that is the $700 full build.
The dirty secret of this market: most vendors sell a $2000+/year SaaS seat and leave the knowledge base empty. The actual work small businesses need — mining their chats, writing their answers, tuning for a week — is exactly what nobody wants to do, and exactly what creates the result.
If you want to try this yourself
Start with your own or a friend's business. Pull their real chats, build the 20-30 answer knowledge base, and ship it in a weekend. The delivery skill is worth far more than the platform subscription.
Questions about fitting a specific industry — drop it in the comments with the industry and rough daily message volume.
Top comments (0)