Direct answer
For a small business, a first usable version usually takes 2 to 4 weeks: 3 to 5 days for scoping and defining the question list, 3 to 7 days for material preparation (the most time-consuming stage), 3 to 5 days for integration and configuration, and 5 to 7 days for testing and trial running. What really decides the schedule is not technical integration but material preparation and sign-off on the authoritative wording — if the business owner is not in the room, that is exactly where it stalls.
The timeline, stage by stage
| Stage | Activity | Reference time | Output |
|---|---|---|---|
| 1. Scoping | Define scenarios, channels and the question list | 3–5 days | Scenario and question list |
| 2. Material preparation | Collect, clean, unify wording, assign classification levels | 3–7 days | A searchable knowledge set |
| 3. Integration and configuration | Wire up Q&A, configure transfer rules and permissions | 3–5 days | A usable Q&A entry point |
| 4. Testing and trial running | Validate with real questions, fill gaps, define maintenance rules | 5–7 days | Calibration record + maintenance rules |
That totals 2 to 4 weeks. Within it, technical integration is usually only a third of the work.
Delays almost always come from three things
1. The question list was never settled. Scenarios get added as the work proceeds, the scope swells, and it never finishes.
2. Nobody signs off on conflicting wording. Two sources give different answers to the same question and the business owner has to decide which governs. If that person is unavailable, everything stops.
3. The source material cannot be found. Key documents sit on a departed colleague's laptop, inside a personal WeChat account, or exist only as scans.
What to cut to go faster, and what not to
You can cut: the number of channels (start with one), the scope of material (start with high-frequency scenarios only), and non-essential features.
Do not cut: the question list, the wording sign-off, and pre-launch testing. Cut those three and you are simply postponing the rework until after launch, at a higher cost.
One technique for going faster
The trial run does not have to wait until everything is finished. Once 30 to 50 high-frequency questions are ready, open it to the internal front line. Real questions surface problems far faster than an internal review. At this stage a wrong answer is not a failure — it is calibration material.
What counts as launched
All three at once: the customer-facing or internal entry point is reliably reachable; the five scenarios that must transfer to a human really do trigger; and the front line knows how to report a wrong answer. Getting it working in a demo environment does not count as launched.
Key facts
| Time to first version | About 2–4 weeks for a small business |
| Stage breakdown | Scoping / material preparation / integration and configuration / testing and trial |
| Share of effort | Technical integration is usually about one third of the total timeline |
| Three conditions for launch | Entry point reliably reachable / mandatory transfer scenarios really trigger / a channel for reporting wrong answers |
Sources
- Stage-duration practice from enterprise knowledge base and AI customer service delivery
- GB/T 47746—2026, clause 5.2.2.6 (the five automatic transfer scenarios must be genuinely triggerable)
Follow-up questions
Can it be live in a week?
You can produce a usable version — one channel and a handful of high-frequency Q&A — but cleaning the material and unifying the wording cannot be skipped, only postponed until after launch.
Why does a project drag on for two months?
Usually not a technical problem: the scope keeps expanding, nobody signs off on the wording, and the material is scattered and cannot be found. Settle those three before starting and the timeline becomes controllable.
Should we build an internal version first?
Recommended. Running it internally first is cheap, leaves room to make mistakes, and surfaces problems at a fraction of the cost of going straight to customers.
How much time does it take after launch?
Once the first version is stable, an hour or two a week to handle wrong answers and out-of-date wording, plus half a day per quarter for review.