Problem: a booking bot in Instagram Direct can write a fluent reply with a price that is not in the price list or a time that is already taken. Result: a split where code returns exact prices and free slots, and a checker rejects any drafted reply that contains a number code did not return or touches a forbidden topic. Stack: Python 3.10+ standard library (tested on 3.15.2). Boundaries: the model that drafts replies is not included; the example uses hand-written drafts to show what the checker accepts and rejects
The design
In Direct, people write in their own words, send voice notes, ask several things at once and come back a day later. A language model handles that conversational part well. Prices, dates and appointment times are a different matter: they must be exact, so they come from code that reads the price list and the calendar
On top of that, a second check looks at every drafted reply before it is sent. In production I use a second model for the rule part, for example "do not discuss contraindications", and code for the numeric part. The example below implements both as simple code so the idea is testable in isolation
I built an AI receptionist for a clinic in Instagram Direct on this pattern. It answers only from the clinic's documents, about 114 price items and 48 services, takes prices from the price list rather than the model's memory, and sends a card to the doctor in Telegram; the booking appears only after the doctor confirms. This article is a teaching version of the idea, not that bot's code
Input, processing, output
Input: a requested service and day; a drafted reply from the conversation model
Processing: code looks up the price and free slots, or the nearest day with slots; the checker extracts every number-like token from the draft and rejects it unless code returned it, and rejects forbidden topics
Output: the facts for the model to use, and a list of problems for each draft (empty means it may be sent)
The example
Prices, dates and drafts are invented for illustration
PRICES = {"consultation": 50, "cleaning": 80}
CALENDAR = {"2026-11-14": ["11:00", "14:00"], "2026-11-15": []}
FORBIDDEN = ("contraindication",)
def facts(service, day):
"""Code side: exact data only, never guessed."""
price = PRICES.get(service)
slots = CALENDAR.get(day)
if slots == []:
nearest = next(((d, s) for d, s in sorted(CALENDAR.items()) if s), None)
return {"price": price, "slots": [], "nearest": nearest}
return {"price": price, "slots": slots or [], "nearest": None}
def check_reply(reply: str, data: dict) -> list[str]:
"""Checker side: the drafted reply may only use numbers code returned and must respect rules."""
problems = [f"forbidden topic: {w}" for w in FORBIDDEN if w in reply.lower()]
allowed = {str(data["price"])} | set(data["slots"]) | ({data["nearest"][1][0]} if data["nearest"] else set())
for token in reply.replace(",", " ").split():
if any(ch.isdigit() for ch in token) and token.strip(".") not in allowed:
problems.append(f"number not from code: {token}")
return problems
if __name__ == "__main__":
d = facts("cleaning", "2026-11-14")
print("facts:", d)
print("draft 1:", check_reply("Cleaning costs 80, free at 11:00 or 14:00", d) or "ok")
print("draft 2:", check_reply("Cleaning costs 75, free at 11:00", d) or "ok")
print("draft 3:", check_reply("No contraindication for you, 80 at 14:00", d) or "ok")
print("sunday:", facts("cleaning", "2026-11-15"))
Output, exactly as printed when I ran it:
facts: {'price': 80, 'slots': ['11:00', '14:00'], 'nearest': None}
draft 1: ok
draft 2: ['number not from code: 75']
draft 3: ['forbidden topic: contraindication']
sunday: {'price': 80, 'slots': [], 'nearest': ('2026-11-14', ['11:00', '14:00'])}
Draft 2 is the failure this design exists for: a plausible price that appears nowhere. Draft 3 is the other one: a correct number wrapped in a topic the bot must not discuss
Design notes
Code returns, the model phrases. The model never decides a price or a slot; it receives them and turns them into a sentence. If code returns nothing, the model says it will check with the receptionist
Nearest options instead of invention. When a day is full, code returns the nearest day with slots, and the model offers that. It never proposes a time code did not return
Check what is sent, not what was intended. The checker reads the final draft. A model can receive the right price and still type a different one; only a check on the outgoing text catches that
Separate checkers for separate jobs. Rules about topics need judgement, which is why production uses a second model for them; numbers need exactness, which is why code checks them. Neither check relies on the model that wrote the reply
Errors and limits
Token matching is crude. "80.00" versus "80", or a time written as "2 pm", will be flagged. Normalise formats in the drafting prompt or in the checker
Forbidden topics by keyword are only a demo. Real rules need a model check, because customers ask cleverly and in many languages
Test on real messages. Voice notes with noise, three questions in one message and a return to the price a day later are where both the model and the checker are tested properly
Reproduce it
- Save the code as
booking.pyand runpython3 booking.py - Compare with the output above
- Add a slot for 2026-11-15 and confirm the "sunday" line returns it with no nearest suggestion
- Change draft 2 to 80 and confirm it passes
A bot like this, with one core for Direct, WhatsApp and website chat, takes me about 2–3 weeks including testing. The non-technical version: Instagram DM automation: booking with checks
This article was created with the help of AI.



Top comments (0)