Day 2 — AI Text Generation & Customer Acquisition
AI & Marketing
The Bot Decision in One Slide
Perception — Tokenization
The model does not read words.
It reads tokens — integer IDs representing chunks of text.
“Is there any way to get a discount before I commit to the annual plan?”
16 tokens · 70 characters
Each word or word-fragment maps to one integer in the model’s vocabulary. The model never sees letters.
311 appears twice — once for “to get” and once for “to the.” Same ID. Completely different meaning. The model resolves that from surrounding tokens.
Framework — Perception row filled:
| Day 1 concept | This system |
|---|---|
| Perception | Token IDs from the input message |
| Representation | ? |
| Model | ? |
| Constraints | ? |
| Algorithm | ? |
| Action | ? |
Perception — Token IDs
The model does not read words.
It reads tokens — integer IDs representing chunks of text.
“Is there any way to get a discount before I commit to the annual plan?”
16 tokens · 70 characters
[3031, 1354, 1062, 2006, 316, 717, 261, 11522, 2254, 357, 8737, 316, 290, 12355, 3496, 30]
The model never sees the word “discount.” It sees 11522.
Every operation from here is linear algebra on these integers — tokenized, then converted to vectors.
Framework — Perception row filled:
| Day 1 concept | This system |
|---|---|
| Perception | Token IDs from the input message |
| Representation | ? |
| Model | ? |
| Constraints | ? |
| Algorithm | ? |
| Action | ? |
The Simple System in Action
Steps 1–3
1. Perception: Message + tool definition arrive. calculator(expression: string) → float
2. Representation: Context window assembled.
3. Model — LLM pass 1: Predicts a tool call:
{"tool": "calculator",
"input": "800 * 0.80 * 1.06"}Steps 4–6
4. Algorithm: Calculator runs: 800 × 0.80 × 1.06 = 678.4
5. Representation updated: [TOOL RESULT] calculator → 678.4
6. Model — LLM pass 2:
“With the 20% discount and DC’s 6% SaaS tax, your monthly cost would be $678.40.”
Correct because the calculator verified it.
Worked Example — The Full System (Steps 1–2)
Same question. Full system.
“Is there any way to get a discount before I commit to the annual plan?”
Step 1 — Retrieval (before LLM)
System detects discount + annual plan intent. Queries three sources:
Promotions DB → ANNUAL20: 20% off,
new customers, expires Apr 30
Pricing docs → Professional: $800/month
Policy doc → Up to 20% self-serve;
above 20% needs approval
CRM record → Lead score: 82
Plan: Professional
Status: new customerStep 2 — Context injection
All retrieved data injected into context window.
LLM now perceives: question + actual policy + active promotion + prospect’s account status.
Worked Example — The Full System (Steps 3–4)
Step 3 — LLM generates (grounded)
“Yes — we have a current promotion for new customers on annual billing: 20% off your first year. For the Professional plan, that’s $640/month — or $7,680/year. Would you like me to apply that to your account?”
Not because the LLM knew any of this. Because the system retrieved it and put it in context.
Step 4 — Tool calls (if prospect says yes)
apply_promotion(account_id, "ANNUAL20")
log_crm_note(lead_id, "Discount applied — 20%")Guardrail check: 20% within self-serve authority. ✓
What changed from the raw LLM:
- ANNUAL20 retrieved, not hallucinated
- $640 verified by calculator, not predicted
- Policy check confirmed self-serve authority
- CRM note actually logged
The LLM generated the words. The system provided the facts, enforced the policy, and took the action.
