- Home
- AI for business
- LLMs for enterprises
The right language model for each task.
We compare models by task, volume, response time and data rules, and keep switching models simple for when the picture changes.
The useful question is not which model is best, but which one solves this task with the quality, cost and time it calls for. The answer is usually more than one, and it changes with each new version.
How an LLM enters the operation
Between the task and the application sit the data rules, the model choice, the safeguards and the evaluation.
- TaskRead, extract, classify
- Data rulesWhat can leave the network
- Task targetsQuality, cost and time
- GatewayCalls the right model
- Long contextDense documents
- Fast modelHigh-volume classification
- Open modelOn the company network
- GuardrailsPersonal data and injection
- EvaluationReal reference cases
- ApplicationAgent, search or service
Reference architecture. Models and layers change with each company's task and data rules.
How we choose the model
Four criteria, measured on your task and with your data.
Quality on the task
A set of real cases compares the models on your task. A public benchmark is a starting point, not a decision.
Cost per task
It adds up tokens, cache, search and human review. Price per token alone does not tell you what it costs to solve a task.
Response time
Live customer service calls for a fast answer. A batch analysis can use a slower, cheaper model.
Data rules
LGPD, contracts and internal policy define what can leave the network and where the model has to run.
When to use each model family
A starting point, which evaluation confirms or changes.
What goes around the model
The model is the part that changes most. The layer around it is what stays.
Gateway
A single point for calling models, with usage limits, fallback and model switching without rewriting the application.
Guardrails
Personal data filtering, protection against prompt injection and limits on what the model can do, following OWASP's recommendations for LLMs.
Versioned prompts
Prompts in Git, tested with every change, like any other code.
Telemetry
A record of cost, time and quality per task, with personal data masked.
How we start
- 1
Define the task
What goes in, what comes out and how to tell whether the answer is good.
- 2
Compare models
We test a few models on your real cases and measure quality, cost and time.
- 3
Integrate
We put the chosen model behind the gateway, with safeguards and telemetry.
- 4
Track
Each new model version goes through the same evaluation before it enters.
Frequently asked questions
What is an LLM?
It is a large language model, trained to predict text. In practice, it reads, summarizes, classifies, extracts data and writes. On its own, it does not know your company's data or act in your systems: that is where RAG over company documents and AI agents connected to the ERP come in.
What is the best LLM for enterprises?
There is no single one. The best is the one that solves your task with the quality, cost and time needed, within your data rules. It is common to combine one model for heavy analysis, a fast one for classification and an open one for data that cannot leave the network.
Is it safe to use an LLM with company data?
It can be, with the right design: API contracts that do not use your data for training, personal data masking, encryption, operation logs and, when needed, an open model running on your own network. The decision involves the DPO from the scoping stage.
Is it worth running your own model?
Only when there is a clear reason: data that cannot leave the network, very low latency or a volume that makes the API more expensive than running it. Your own model brings infrastructure and team costs, which have to be counted.
Do I need fine-tuning?
Rarely at the start. First, exhaust prompt, context and RAG. Fine-tuning makes sense when there are many good examples and a well-defined task those paths have not solved.
How do you avoid being locked in to one model vendor?
With a gateway between the application and the models, versioned prompts and your own evaluation set. That way you can test and switch models while measuring the effect, without rewriting the application.
Talk to an AI specialist.
Thirty minutes with a CCX engineer, at no cost, to look at your task and what it asks of a model.