
What Is a Foundation Model? Why Your Company Doesn't Need to Train One
A foundation model is AI trained on broad data for many tasks. Learn why companies use one without training it, and what changes when picking a vendor.
Want to apply this in your company?
Book a call with Juan →A foundation model is an AI model trained on large volumes of general data, able to carry out many different tasks such as summarizing, classifying, drafting and answering questions, based on instructions written in plain language. Companies usually use one of these models off the shelf, without training it from scratch, and adapt its behavior with instructions, the company's own documents and usage rules. As a result, the decision is no longer only about which model answers best. It also involves data handling, control, records and how easily you can switch providers.
Key takeaways
- A foundation model is a general, pre-trained model that serves as the base for many different applications.
- Training a model from scratch requires data, infrastructure and a team that most companies do not need for common support and administrative processes.
- Clear instructions, controlled access to information and integrations with your company's systems usually solve the problem before any change to the model itself.
- When choosing a vendor, the data policy, where processing happens, usage logs, permission controls and the technical cost of switching models all matter.
- Superintelligence is a hypothetical concept. What companies buy today is a tool built on a foundation model, and the decision should rest on what it does in your workflow.
What a foundation model is, without the jargon
A foundation model comes out of broad training, in which the system learns patterns of language and other kinds of content from a very large set of texts and examples. After that training, it can be used for tasks that no one programmed one by one. The same base can answer a customer question on WhatsApp, summarize a meeting, extract data from an order or suggest a reply to a complaint.
The name reflects the role the model plays: it is the foundation on which other applications are built. Someone using a ready-made chat is using a product that runs on a model. Someone who connects a vendor's API to their own business workflow is building an application on the same base. This distinction matters because it changes what you control and what you delegate. To separate product, model and concept, read ChatGPT superintelligence? What it is and what it is not.
Why companies use a model without training it
Training a model requires large volumes of data, significant computing capacity, specialized staff and a long evaluation process. For most routines in support, quoting, triage and administration, the gain comes from defining the task, the context and the limits well, not from building a proprietary model.
There are three practical reasons for this choice. The first is focus: time that would go into training usually pays off more when it is spent mapping the process, organizing information and measuring results. The second is testing speed, because changing an instruction or trying another model takes hours, while your own training takes much longer and requires rerunning evaluations. The third is the starting point: ready-made models have already gone through general evaluations by vendors and the research community, a baseline a company would not have on its own.
Before thinking about any adjustment, check what your company has organized. Good answers depend on standardized data, consistent product and service names and written rules, as explained in Data ready for superintelligence: what to standardize first.
Training, tuning or giving context: how to decide
The useful question is not whether the company should train a model, but what the smallest resource is that solves the observed problem. Use this guide as a starting point:
- If the task is general and the knowledge it needs fits into instructions, start with a ready-made model and a well-defined prompt, with examples of inputs and outputs.
- If the answer depends on documents that change often, such as price lists, policies and manuals, use context retrieval, where the system searches for the right passage before answering, instead of embedding that content in the model.
- If the format or style must be very specific and repetitive, evaluate fine-tuning only after you have reviewed examples from people and an agreed success metric.
- If the company is considering training a model from scratch, the conversation has moved from process automation into research, which is usually not the first step.
- If the workflow involves sensitive personal data or confidential information, the processing location and access controls matter more than how sophisticated the model is.
AWS documentation on context retrieval and fine-tuning compares these approaches clearly and is a useful read before this decision.
What changes when choosing a vendor
When a company uses a ready-made model, choosing a vendor stops being only a question of answer quality. It also involves the contract, the architecture and the ability to control what happens to your data. Some questions should be answered in writing before any pilot:
- Where data is processed and stored, and whether there is international transfer that requires attention under Brazil's LGPD, as detailed in International data transfer under LGPD when using AI providers.
- Whether the vendor uses the information you send to improve its models, and how that use can be disabled or contracted differently.
- Which usage records, logs and identifiers the system generates, and whether your company can export them for audit.
- Which actions the agent can take on its own and which require human approval, including handoff to a person when the conversation goes off script.
- How dependent the operation becomes on a single model, and how much technical effort it takes to switch vendors without rebuilding the entire workflow.
For simple, high-volume tasks, such as classifying messages or standardizing a customer record, smaller models or models running on your own infrastructure may be enough. That path deserves testing before any decision, as discussed in Small language models for business tasks: when they are enough.
How to apply the choice in a pilot
The working method we follow has four steps: map the process and the baseline, limit data and actions, pilot a real workflow with records and an agreed metric, and then hand over the operation and decide whether to expand, adjust or stop. Applied to model selection, this means testing more than one option in the same workflow, with the same examples, before signing any contract.
Illustrative scenario: a building materials store wants to use AI to sort quote requests arriving on WhatsApp into ready stock, custom orders and technical questions. The team follows a simple path:
- Gathers a representative batch of real messages, already anonymized, and classifies them manually with help from someone in sales.
- Defines what counts as a correct answer for each category and which cases must go straight to human review.
- Runs the same set through two ready-made models and records the responses, response times and the cost of each call.
- Compares the types of errors, not just the overall accuracy, separating mistakes that cost a sale from those that are only an annoyance.
- Chooses the model that meets the metric with the required controls and keeps a future switch possible.
The result of this pilot tells you which model serves this workflow, with this data and these limits. It does not tell you which is the best model on the market, and that mistake is common in decisions made only by comparing product demos.
Risks a better model does not solve
Switching to a more capable model does not remove operational problems. Customer messages can contain hidden instructions that try to steer the agent's behavior, a practice known as prompt injection. The OWASP list of risks for applications built on language models is a practical reference for including this type of threat in your assessment.
Vendors also update their models, and behavior can change without the team noticing right away. For that reason, keep a fixed set of test cases, log the responses and reevaluate the workflow whenever the model version changes. For broader governance, the NIST AI risk management framework offers useful vocabulary for discussing the topic with leadership, legal and technology teams in the same language.
Frequently asked questions
Is a foundation model the same thing as ChatGPT? Not exactly. ChatGPT is a product built on foundation models, with its own interface, usage rules and features. The model is the technical base, and the product is the application that delivers that base to the user.
Do I need to train a model to have an AI agent in my company? In most cases, no. An agent usually works with a ready-made model, clear instructions, controlled access to information and rules for handing off to people. Your own training only makes sense if measurements show a need that context and information retrieval cannot solve.
Can a foundation model be considered superintelligence? Superintelligence generally describes a hypothetical system with capabilities far beyond human ones in nearly every area, not a product available today. The models companies contract are useful but limited tools that make mistakes and require control. Use that term with caution in internal communication.
How do I choose between model vendors? Start with the workflow, not the brand. Test at least two options with the same real examples, check how data is handled, the records generated and how easy a switch would be, and agree on the success metric before the pilot begins.