Frequently asked questions

An AI agent is an application that pairs a language model (the ability to reason) with tools (the ability to act) to accomplish a goal: in essence, “a language model in a loop, using tools to complete a task.”

Unlike a model that only replies to a single prompt, an agent plans, takes actions, and observes the result, repeating the loop until the mission is done, with little or no human input at each step.

A traditional chatbot follows a single exchange: you ask, it answers from rules or text. An AI agent runs a five-step loop: get the mission, scan the context, reason and plan, act with tools, and observe the result to decide the next step.

That lets it handle multi-step tasks (like looking up an order, checking stock, and replying with a real answer) instead of returning pre-scripted text.

They ground their answers in authoritative knowledge before responding, a technique called RAG (retrieval-augmented generation). Instead of relying only on what the model “remembers,” the agent queries trusted documents, databases, or websites, like a library card, and bases the answer on facts.

This drastically reduces hallucinations and lets the agent cite the source of each answer.

Beyond answering, an agent can act in the real world through tools: send an email, schedule a meeting, update a record in a CRM, or query a database.

That is what turns an assistant from a simple source of information into one that completes tasks end to end.

Instead of a single all-powerful “super-agent,” a multi-agent system is a “team of specialists,” much like a human organization. A coordinator agent breaks the task down and routes each subtask to a specialized agent, then assembles the results into a final answer.

Dividing the work this way makes complex processes simpler to build, test, and maintain.

It can be, when built with the right practices. Strong platforms use defense in depth: deterministic guardrails (hard rules that limit what the agent can do), least-privilege permissions, a distinct identity per agent, and a human-in-the-loop checkpoint before sensitive actions.

That keeps the agent within predictable, auditable limits even against manipulation attempts such as prompt injection.

RAG, short for retrieval-augmented generation, is the technique of grounding a model's answers in an external knowledge base: before answering, the system retrieves the relevant passages from your documents and uses them as context.

That keeps the answer anchored in your own data, instead of relying only on what the model memorized during training: more accurate, updatable, and verifiable.

At Charla we use agentic retrieval: the assistant itself decides when to consult the knowledge base, calling search as a tool and on demand, rather than receiving passively injected passages.

MCP, the Model Context Protocol, is an open standard for connecting AI agents to external tools and data. It works like a “universal plug”: instead of a bespoke integration for every service, the agent speaks a common language with any compatible tool.

Charla connects to external systems as an MCP client, which lets you expand what an assistant can do by wiring it to applications and data sources in a standardized, reusable way.

Not in the ways that matter: Charla's posture is propose-then-approve. Knowledge and configuration changes land as reviewable proposals that need human approval before taking effect.

There is also an optional scope gate, a kind of semantic firewall, that refuses requests outside the assistant's purpose, and per-assistant budgets that cap spend.

In practice, humans stay in control of the more consequential actions, combining the assistant's speed with human judgment.

Charla is a platform for building agentic AI assistants that understand your business, answer in any language, and meet customers where they already are: your website, a public link, or WhatsApp.

Each Charla isn’t a single model with tools but a team of specialized agents working together (answering, critiquing, and even building new agents), each with its own knowledge, persona, and permissions.

Following Google’s five levels of agentic systems, Charla operates today at Level 3 (a collaborative multi-agent system) with a foot in Level 4. It goes well beyond a chatbot querying a knowledge base (Level 1) or a single agent that plans (Level 2).

Inside each Charla, a team of specialized agents works together: the answering agent runs alongside Steve, which critiques every answer against a gold standard (the “generator and critic” pattern), and Charlie, who builds new agents interactively, with a person in command.

No. You can build an assistant in minutes in the visual builder, no code required. But if your team does code, you can go further: Charla offers a REST API, SSE streaming, MCP connections, and custom skills.

No code if you want; with code if you need it.

Far more than a chatbot. Charla understands a vast base (documents, spreadsheets, and images), serves customers, integrates with your systems and APIs, and automates entire workflows, all within your brand.

It turns messy sources into structured output (a spreadsheet, a presentation, or a database), queries large databases in natural language to return analytical reports, and uses machine-learning models (forecasting, recommendation, optimization) as tools in service of your KPIs.

Charla is genuinely multilingual: it understands and answers in any language while keeping your brand’s tone, and the interface greets visitors in 13 languages.

Through agentic tools: a REST API and SSE streaming to integrate with your stack, MCP connections to plug in external tools, and versioned skills to build your own capabilities.

Every action is logged and auditable, and each change comes in as a proposal for you to review and approve: nothing happens in the dark.

Set a monthly credit cap per assistant; when it reaches the limit, it pauses on its own, with no surprises on the invoice. Each Charla is also fully isolated, with its own knowledge, persona, and permissions.

Just talk to our sales team on WhatsApp or by email (contato@charla.ia.br); we’ll help you set up and launch your first assistant in a few steps.

Usage is measured in credits, where 1 credit equals US$0.01 of AI provider cost. This makes consumption transparent and directly tied to what each assistant actually processes.

Spend can be capped per assistant with a monthly budget: a soft cap triggers alerts as you approach the limit, and a hard cap pauses the assistant once it is reached, avoiding surprises.

Pricing and quotas are arranged commercially with the Charla team, based on your use case. Talk to us to set up the right configuration.

Each assistant is isolated: it has its own knowledge base and its own access list, so one assistant's knowledge and permissions do not leak into another. Access is governed by role-based controls, with scoped permissions.

Privileged actions and tool calls are recorded in an append-only audit log. Secrets and credentials are encrypted and held in a secret manager, traffic is encrypted in transit (TLS), and connections to external tools are credential-sealed and egress-guarded.

Important: Charla does not currently hold a formal compliance certification. The points above describe the platform's security practices, not certifications.

A Charla assistant can serve on the authenticated web app, on WhatsApp, through a public, shareable chat URL, and through an embeddable website widget.

That way you can bring the same assistant to where your customers already are: internally, on a public page, or inside your own website.

Charla is model-agnostic by design: the model is a swappable layer, which avoids lock-in to a single provider and lets the platform keep pace with the technology.

The current generation runs on Google's Gemini via Vertex AI.