AI applied to the business, running in production.

The most common mistake is choosing the model before the problem. The result is an interesting pilot with no return and no place in the process. CCX reverses the order: we map three to five workflows where human time is most expensive and repetitive, such as customer service, document analysis, reconciliation, contract review and ticket triage.

For each workflow, we build a prototype with the least infrastructure possible and measure three things: quality against the human baseline, cost per task and response time. Only what passes all three criteria goes to production, with guardrails, telemetry and a feedback loop. This avoids AI that only produces good-looking reports, and premature fine-tuning before RAG and prompt and context engineering have been exhausted.

OpenAI stands out in complex reasoning and in agents with heavy tool use. Anthropic, in long context, dense document analysis and legal and technical reasoning. Google's Gemini, in multimodality and integration with Google Cloud. Open models, such as Llama, Qwen and Mistral, meet requirements for on-premises data, very low latency or massive volume at a controlled cost.

In practice, the stack is usually hybrid: one model for heavy text analysis, another for integration agents, a fast one for high-volume classification and an open one, hosted on your own network, for data that cannot leave it.

In e-commerce, AI enriches the catalog from the product spec sheet, classifies reviews and returns and serves customers integrated with the platform, escalating to a human when needed. In the supply chain, it forecasts stockouts by SKU, channel and region, plans routes in the TMS and checks documents such as invoices (nota fiscal), bills of lading and packing lists.

In the back office, it analyzes contracts in batches, reconciles ERP and bank, and triages tickets. In sales and marketing, it enriches lead scoring, personalizes emails and proposals and monitors the market.

There are four layers: data (ERP, e-commerce, CRM and BI in a warehouse, with a vector database), application (a model with RAG over internal data and prompts versioned in Git), agents (when the model needs to call APIs and tools) and evaluation (LLM-as-judge, human review and telemetry).

Governance is part of the architecture: logging of every prompt and response with personal data masking, human review for low-confidence cases, kill switches and usage limits, an immutable audit trail and LGPD compliance with the DPO involved from the scoping stage. On security, protection against prompt injection, data leakage and jailbreaks, following the OWASP Top 10 for LLMs and the NIST AI RMF.

Start small, but complete: one high-volume workflow with a clear return, delivered end to end, from raw data to automated action. A pilot with measured results becomes the argument for the next projects; a broad strategy program with no concrete delivery loses sponsorship at the first budget squeeze.

A typical pilot, lasting 8 to 12 weeks, covers discovery, a prototype in production, evaluation and handover, with a lean squad: a senior AI engineer, a data engineer, a product lead and an advisory DPO.