Skip to content
ChatterfyHelp Center
HomeGeneralCRMAI ManagerTrackerPlatformsFAQUpdate log
Основы и возможности
Подключение AI
How to connect AI to a bot
How to set up AI after connecting
How to improve AI quality
What to do if AI makes mistakes
Внутреннее устройство
Промпты и переменные
  1. AI Manager
  2. Подключение AI
  3. What to do if AI makes mistakes

What to do if AI makes mistakes

What to do if AI makes mistakes Typical symptoms of AI issues in Chatterfy and what to check first: hallucinations, the wrong language, errors , высокий расход, проблемы Recognition, Variables, задерж

Updated June 21, 20262 min


What to do if AI makes mistakes

Typical symptoms of AI issues in Chatterfy and what to check first: hallucinations, the wrong language, errors AI Routings — Deep Dive, высокий расход, проблемы Recognition, Variables, задержки и зацикливание flow., high usage, Recognition and Variables issues, delays, and flow looping.
Systematic AI improvement — iterations and metrics: How to improve AI quality. Layer architecture and module interconnections: How AI works in Chatterfy.


Catalog of typical issues



AI makes up facts, prices, and features

Check
• Restrictions — what the bot must not do — whether there is an explicit instruction not to make things up.
• Global Legend — who is the bot — whether there are any vague wordings.


What to do

  1. Add the facts to the Reply Addition Info of the AI Reply step.
  1. Lower creativity.
  1. Change the AI Reply model if necessary.


AI replies in the wrong language

Check
• AI Settings — detailed breakdown — the language, tone, and strictness of the wording.


What to do

  1. Explicitly specify in the prompt: ALWAYS answer only in … (the required language).


AI Routings — Deep Dive leads to the wrong branch

Check
Whether the conditions of different Routings overlap.
The order of routes in the list.
Whether the wording of the conditions is too similar.


What to do

  1. Narrow the conditions.
  1. Reduce the number of branches.
Learn more AI Routings — Deep Dive.


AI is too expensive

Check
How many consecutive AI Reply steps there are in the flow.
The length and complexity of prompts.


What to do

  1. Use a lighter model for AI Reply.
  1. Shorten and simplify prompts.
  1. Reduce the number of AI Reply steps inside the flow.
Learn more AI Models — model selection.


AI ignores the step context

Check
Whether the step Goal is filled in.
Whether key data is present in Reply Addition Info.
Whether Global Legend conflicts with local instructions.


What to do

  1. Duplicate the critical context locally inside the step.
Learn more AI Reply — detailed breakdown of the step.


Recognition does not read photos or documents

Check
Photo mode = Recognition.
Recognition Tasks and Functions.


What to do

  1. Clarify the recognition task.
  1. For complex cases, switch to Manual.
Learn more AI Recognition — deep dive.


Voice messages are not recognized

Check
Voice = Recognition.
Audio quality and speech language.
What format the user sends the voice message in.


What to do

  1. Temporarily enable Manual for voice messages.


Flow moves on too early

Check
Whether the Requirements are too broad.
Whether there are any false Routings triggers.


What to do

  1. Tighten the conditions.
  1. Separate similar routes.


Tone of voice breaks

Check
• AI Settings — detailed breakdown, creativity, and examples in the prompt.
Whether there are contradictions between the Global Legend and the examples.


What to do

  1. Reduce creativity.
  1. Add an example of the desired message to the AI Reply step.
  1. Use a local Legend on the problematic step.


What to check for any issue

Most often, the cause is in one of these places:
  • What broke
  • Where to check
  • Response text
  • AI Settings / AI Reply settings
  • Flow branch
  • AI Routings — Deep Dive
  • Photos / voice messages
  • AI Recognition — deep dive
  • Variables
  • Bot Variables in AI Prompts


Related

How to improve AI quality
How to improve AI quality Iterative tuning: prompts, A/B, model selection, measurements. The “observe → hypothesis → adjustment → measurement” cycle is the only way to avoid turning tuning into random

Was this article helpful?

BackHow to improve AI quality

On this page

  • What to do if AI makes mistakes
  • Catalog of typical issues
  • AI makes up facts, prices, and features
  • What to do
  • AI replies in the wrong language
  • What to do
  • leads to the wrong branch
  • What to do
  • AI is too expensive
  • What to do
  • AI ignores the step context
  • What to do
  • Recognition does not read photos or documents
  • What to do
  • Voice messages are not recognized
  • What to do
  • Flow moves on too early
  • What to do
  • Tone of voice breaks
  • What to do
  • What to check for any issue
  • Related