Function calling: why it changed the way to build agents

The LLM no longer just answers: it executes actions in your system. This is what you need to know.

The real problem

Before talking about AI, technologies or providers, it's worth naming the real problem: customers write at 10:35 PM on a Sunday and expect a response in 5 minutes. If they don't get it, they go to the competitor who does respond.

A well-built chatbot solves 70-80% of repetitive queries without human intervention. The other 20-30% are escalated with context.

1) What function calling is and how it works

This is the technical starting point. It's not complex, but there are decisions that lock you in long-term if you get them wrong.

• Define the main intents (3-7 total)
• Audit the current knowledge base (FAQs, policies, catalog)
• Choose the primary channel (WhatsApp API is usually the winner for LATAM)

2) Tool design: what to expose and what not to

This is where conversational design comes in, which is more art than science. The goal is not to imitate a human, it's to solve the problem fast.

• Short, on-point greetings, not introduction monologues
• Always keep a visible "human exit"
• Micro-confirmations: repeat what was understood before executing actions

3) Security: the LLM is not your guardian

This is the difference between a bot that annoys and one that converts. Most people don't think about it and then wonder why users hate it.

• Measure "time to resolution" by type of query
• Log every fallback to iterate on the prompt or intents
• Review 20 real conversations per week for the first 8 weeks

Recommended architecture

Most chatbots in production today combine:

• An intent classification layer (rules or embeddings)
• A hybrid response engine (templates + LLM when appropriate)
• Integration with the business system (CRM, ERP, calendar)
• A human accessible with one button tap

Realistic costs

• Initial setup: USD 500 to USD 3,000 depending on complexity
• Monthly maintenance: USD 100 to USD 500 (infra + API + tweaks)
• Cost per conversation: USD 0.01 to USD 0.08 with modern models

Note: a human conversation in a contact center costs between USD 2 and USD 7. The math speaks for itself.

Common mistakes we see

Thinking the bot will replace humans No. It frees them from repetitive tasks. Humans focus on the complex and relational.

Too many options at the start 7 options in the first message = lost user. Less is more.

Ignoring monitoring A bot without metrics is worse than no bot. You'll find out when it's too late.

Next step

If you want to see a demo applied to your industry, at StriqTech we build a working pilot in 7-14 days.