Direct answer
A small business works through four phases and typically has a first usable version in 2–4 weeks: 1) scoping and classification (3–5 days); 2) document cleaning and ingestion (3–7 days — the slowest and most underestimated step); 3) question-answering integration and permissions (3–5 days); 4) pilot run and calibration (5–7 days). What actually causes the delay is never the technical integration. It is document preparation: old files in unusable formats, versions that contradict each other, and the person who owns the answer not being in the room.
Four phases and how long they take
| Phase | Main work | Indicative time | Deliverable |
|---|---|---|---|
| 1. Scoping and classification | List the business questions, define the document scope, label classifications | 3–5 days | Question list + classification table |
| 2. Cleaning and ingestion | De-duplicate, reconcile wording, convert into a retrievable format | 3–7 days | A retrievable knowledge set |
| 3. Integration and configuration | Connect question answering, configure permissions and external isolation | 3–5 days | A working question-answering entry point |
| 4. Pilot run and calibration | Test with real questions, fill gaps, settle maintenance rules | 5–7 days | Calibration record + maintenance rules |
That totals 2–4 weeks to a first usable version. A first usable version, not a finished one — a knowledge base keeps growing, so the idea of completing it in one pass does not hold up.
Phase two is the one everyone underestimates
In practice, delays almost all happen during document cleaning:
- Old files in mixed formats (scans, images, legacy Office documents) that have to be converted into readable text first.
- The same question answered differently across several files — two versions of a price list, an out-of-date process — so “which one is authoritative” has to be settled before anything else.
- Key material that lives on a departed colleague's laptop or in one person's chat app, with the original file nowhere to be found.
The real effort in phase two therefore depends on how well the company has managed its documents all along, and very little on which AI tool is used.
The key move: define the question list first
The most important task in phase one is to list the questions you want the AI to be able to answer, then work backwards to the documents those questions require. Doing it the other way round — pile in everything first — has two consequences: retrieval gets diluted by large amounts of unusable material and answer quality drops, and nobody can say when the work is finished.
Keep the first list to 30–50 high-frequency questions, covering the situations your customer service meets most often.
Who has to be involved
- A business owner — to define the question list and the priority of each version, especially which one is authoritative.
- Front-line staff — to supply the way people really ask, and the wording they use.
- The technical side or supplier — to do the integration and configuration.
An absent business owner is the most common cause of delay — with nobody to call a decision on conflicting versions, document cleaning stalls.
What success looks like after launch
Not how many documents went in. Three things: front-line staff actually start using it; the share of answers they adopt keeps rising; and the problems people run into have somewhere to be recorded and get fixed. Ingesting 5,000 documents nobody uses is the same as doing nothing.
Key facts
| Number of phases | Four (scoping and classification / cleaning and ingestion / integration and configuration / pilot run and calibration) |
| Time to a first version | About 2–4 weeks for a small business |
| First-round question list | 30–50 high-frequency questions recommended |
| Most common cause of delay | Document cleaning (mixed formats, conflicting versions, missing originals) |
Sources
- Phase structure and timing conventions from enterprise knowledge base delivery work
- Three-tier document classification and cleaning-and-ingestion conventions
Follow-up questions
Can it go live in a week?
You can get to “usable” but usually not to “good”. What fits in a week is the integration and a small set of high-frequency answers; document cleaning and reconciling wording cannot be skipped, only pushed past launch.
Do I have to ingest every document?
No. The more you ingest, the more retrieval gets diluted. Work backwards from the question list to the documents you actually need — that beats a full ingest.
What about old scans?
Convert them into readable text before ingesting, or retrieval will never hit them. Format conversion is usually the most time-consuming part of document cleaning.
How much maintenance effort after launch?
Once the first version is stable, typically 1–2 hours a week: handling wrong and missed answers reported by the front line, and updating out-of-date wording. The amount depends on how fast the business changes, not on the number of documents.