Technical architecture

The menu handles the usual, the AI handles what goes off script

In short: almost everything you get asked is the same twenty questions, and your button menu answers those instantly without spending AI. When a customer writes something that isn't on the menu, the AI steps in. And if a person is needed, the conversation reaches your team with the full history. Three steps, in that order.

How it decides who replies

The predictable always gets the same answer; the rest goes to AI

In short: the button menu handles most messages without spending a cent on AI. Only what doesn't fit any option goes to the AI, and whatever the AI can't close goes to a person.

1First, the menu (most messages)
  • “What time do you close?”
  • “How much is shipping?”
  • “Where is my order?”
  • Taking the order and confirming it
  • Asking for their name and address
What it costs
No AI spend
2Then, the AI
  • “Would that work for my case?”
  • Comparing two products
  • A complaint with several things mixed together
  • When they write the way they talk, with slang
  • When they ask two things at once
What it costs
~$39/month at 50.000 conversations
3And if needed, a person
  • “I want to talk to someone”
  • A complaint that needs to be resolved
  • A problem with a payment
  • Anything you'd rather decide yourself
What it costs
Your team, with the full history
See the technical details↓

The decision is exact, not probabilistic: there's no intent classifier and no confidence threshold to tune. The engine compares the message with the triggers of the active flows and with the options of the step the conversation is on. If it matches, it runs that step.

If it doesn't match, the AI rescue kicks in: it answers the question with your knowledge base and sends the conversation back to the step it was on. The rescue limit per session is two; on the third, the AI takes the lead in the conversation instead of resending the same menu.

Handoff to a person also comes from exact rules: it's detected in the text the customer writes (“I want to talk to someone”, “put me through to a person”), not by asking the AI. That's on purpose — if the AI misjudges, the customer gets stuck, and a “?” is a “?”.

A flow step can also call your system to fetch stock, an order status or an updated price, and use the response in the following steps. That still doesn't spend AI.

The problem: pure AI on WhatsApp

  • ✗Every message goes through the LLM, even "what are your hours?"
  • ✗The context grows with every turn: tokens = money × time
  • ✗A bot with 10.000 conversations/month can cost $35/month in AI alone
  • ✗1–3s latency for answers that could be instant
  • ✗And there's no way to pin it down: the AI can make up a price or a condition

The solution: FlujosChat hybrid architecture

  • ✓JSON logic handles 80% of the volume: instant speed, $0 in tokens
  • ✓The AI only steps in when the logic doesn't cover the case
  • ✓Flows guarantee correct, auditable answers for critical data
  • ✓78% lower API costs vs a pure-AI model
  • ✓Full control over the decision tree, with no room for hallucinations

The three steps, in order

Every incoming message goes through here. Most stay at the first step and never reach the AI.

01

The menu leads

As long as the customer's message matches a trigger or one of the options in the active flow, the flow replies. It's an exact match, so it answers instantly and spends no AI.

What it costs
No AI spend
02

The AI rescues

If the customer writes something that doesn't match any option, the AI answers that question with the information you loaded, and the flow goes back to the question it was on. The menu isn't lost.

What it costs
~$39 a month at 50.000 conversations
03

The AI takes over, or hands off to a person

By the second rescue in the same conversation it's clear that customer prefers to type, so the AI leads instead of leaving them going around the same menu. And if they ask to talk to someone —detected by text rules, not by the AI— the conversation reaches your team with the full history.

What it costs
Your team, only when needed

So how much does this cost per month?

The numbers by volume, with the calculation's assumptions in plain sight, live on a single page so there are never two versions of the same figure.

See the numbers →

Frequently asked questions

When does the AI step in, and when does the menu keep going?↓
There's nothing to tune: the rule is exact. As long as the customer's message matches a menu option, the menu replies. If they write something that doesn't match any, the AI answers that question and the flow goes back to the question it was on. If that happens twice in the same conversation, it's clear that customer prefers to type and the AI keeps the conversation. What you do choose is which questions the menu covers: anything there spends no AI.
How much does this save me compared to sending everything to the AI?↓
It depends on how many questions your menu covers. Assuming it covers 80% —typical when it has the usual ones: hours, prices, order status— AI spend drops by around 78%. The numbers by volume, with the calculation's assumptions in plain sight, are on the pricing page.
Can the menu look up data from my system in real time?↓
Yes. A flow step can call your system to check stock, an order status or an updated price, and use the response in the following steps. That spends no AI: it's still the menu.
What if the AI gets a price or a condition wrong?↓
That's why data that has to be exact goes in the menu, not in the AI. The menu always returns exactly what you wrote, so there's nothing to make up. We recommend always putting prices, returns and terms of sale there, and leaving the AI whatever needs interpreting.
Can I use a different AI instead of ChatGPT?↓
Yes. It works with any provider compatible with the OpenAI API, so you can connect Claude, Gemini or another one, including a cheaper one. Deciding who replies —menu, AI or person— doesn't change with the provider.