Before you build an AI assistant, prepare material — not budget

Most people think step one is discussing price and picking a model.

It isn't. Step one is turning what you know into something a machine can use.

The same project ships in seven working days with material ready, and takes twice as long — with worse results — when material is scattered. The difference isn't technical.

TL;DR

Five things to prepare:

Item	Minimum	What it determines
Q&A pairs	30	What it can answer
Past conversations or tickets	50	How it sounds
Product/service description	1 doc	Factual grounding
Boundary list	Required	What it must not say
Escalation rules	Required	When it hands off

The fourth is the most skipped and the most consequential.

---

1. Q&A pairs: 30 minimum

The knowledge base.

Format

Simplest is two columns — question and answer:

Q: How long does shipping take?
A: Next-day within the metro area, 2–3 working days elsewhere, 4–5 days for remote regions.

Better, with synonyms:

Category: Logistics
Canonical question: How long does shipping take
Synonyms: when will it arrive / delivery time / how fast is shipping
Answer: Next-day within the metro area, 2–3 working days elsewhere.

Where to get them

Easiest first:

1. Existing FAQ page — copy it
2. Support chat logs — the top 20 questions
3. Your support staff's head — have them dictate
4. Complaints — the most accurate signal

How many

• Under 20: too thin, the assistant says "I don't know" constantly
• 30–50: covers ~80% of enquiries — a reasonable start
• Over 100: use retrieval rather than cramming into the prompt, or cost climbs

---

2. Past conversations: 50 minimum

This isn't for knowledge. It's for tone.

The same answer can read:

Dear valued customer, regarding your enquiry about shipping times, we wish to clarify as follows…

or:

Metro area is next day, everywhere else two or three days.

If your customers are young, the first gets mocked. If you're in healthcare or government, the second reads as careless.

Fifty real transcripts tell the developer how you actually speak. If none exist, ask support for examples of replies they consider good.

---

3. Product/service description: one document

Doesn't need to be pretty. Needs to be accurate.

• Name and one-line positioning
• Price range (doesn't need to be exact, but must answer "how much")
• Scope (what you do, what you don't)
• Delivery timeline
• After-sales policy

Most common failure: the price in the document is stale, the assistant quotes last month's number, the customer screenshots it.

Anything that changes — price, promotions, stock — either says "check the page" or comes from a live API.

---

4. Boundary list: non-negotiable

The most important item, and the one people skip.

It answers: what must it never say?

Type	Example	Handling
Commitments	"It will definitely arrive in 3 days"	Soften to "typically"
Price promises	"I can give you 20% off"	Route to sales
Medical/legal advice	"You should take…"	Block, escalate
Competitor commentary	"They're worse than us"	Speak only about yourself
Internal info	Cost, suppliers, staff names	Block

Then decide what happens when it does:

• Hard block: keyword filter before output
• Soft redirect: detect the topic, escalate
• Audit: flag in logs, weekly review

At least one of the three.

A project shipped without a boundary list usually gets suspended in week one for "the AI promised something we can't deliver". That's not a model failure — nobody told it where the line was.

---

5. Escalation rules: required

Decide when a human takes over:

Escalate immediately:
- Customer asks for a human
- Refunds, complaints, legal matters
- Two consecutive unresolved turns

Suggest escalation:
- Three questions in and still unclear
- Amounts over a defined threshold
- Negative sentiment detected

And how:

• To whom (which queue, which channel)
• By what means (chat handoff, email, ticket)
• Whether the assistant stays in the conversation (it should exit — otherwise it talks over the human)

---

What happens without preparation

Four real outcomes:

1. Timeline doubles. Gaps mean pausing to wait, or guessing and reworking. Seven days becomes fifteen.

2. Nobody dares turn it on. No boundary list means unpredictable answers, so it never leaves internal testing.

3. Failure gets misattributed. "AI doesn't work" — when the knowledge base was eight Q&A pairs that didn't cover what customers actually ask.

4. Maintenance explodes. Without real transcripts, tone is guesswork. Every "it doesn't sound like us" means another prompt revision.

---

Copy-paste preparation checklist

□ Q&A pairs (≥30)
   □ From existing FAQ
   □ From support logs (top questions)
   □ Synonyms noted

□ Past conversations (≥50)
   □ Including examples you consider good
   □ Tone preference noted (formal / casual)

□ Product/service description
   □ Name and positioning
   □ Price range
   □ Scope, including exclusions
   □ After-sales policy

□ Boundary list
   □ Prohibited commitments
   □ Prohibited topics
   □ Sensitive keyword list

□ Escalation rules
   □ Trigger conditions
   □ Recipient
   □ Channel

---

One honest note

If reading this makes you think "we haven't figured these out ourselves" — then the bottleneck isn't AI, it's that the business isn't structured yet.

In that case I'd suggest doing three things first:

1. Turn frequent questions into an FAQ page (which also helps SEO)
2. Standardise support scripts
3. Define refund and after-sales rules clearly

All three are worth doing whether or not you ever add AI, and the AI project gets much easier afterwards.

---

Next

Once prepared, read the cost breakdown to judge budget and timeline.

What's included and what isn't is on the AI integration service page. If your material exists but you're unsure whether it's enough, send it over.