
How to Prepare a Company for More Capable AI: Design Principles
Design principles so your operation can benefit from better AI models without rewriting processes: organized data, permissions and evaluations.
Want to apply this in your company?
Book a call with Juan →Preparing your company for more capable AI models means designing operations so a better model can replace the current one without rewriting processes, integrations and controls. This comes down to four decisions: data that is organized and has an owner, minimum permissions for the agent, limits and human approval for high-impact actions, and evaluations that measure business results, not just the quality of the text. Teams that do this before they need to switch models do not have to rebuild the whole system each time.
Key takeaways
- Keep business rules out of the prompt. Prices, deadlines, eligibility and policies should live in documented sources and, where possible, in deterministic systems.
- Organize data with an owner, a format, a last-updated date and an official channel before you connect any agent to it.
- Give the agent only the tools and permissions the task requires, with human approval for sending messages, charging customers, canceling services and changing records.
- Build evaluations from real cases in your operation, with success criteria defined before you test a new model.
- Log inputs, decisions, handoffs to humans and errors, so any improvement is measured rather than just felt.
What does "without rewriting" mean in practice?
A model is one component inside a workflow. When a company treats the model as a replaceable part, it needs clear contracts around it: what information goes in, what output format is expected, which tools can be called, and what happens when the answer is not reliable. If those contracts exist, switching models changes the quality of interpretation and writing, but it does not change how the system looks up a customer record, logs a conversation or routes a case to a person.
The most common mistake is putting everything into the prompt: exception rules, product lists, deadline tables and tone instructions in one block of text. That text may seem to work with one specific model, but it becomes fragile when the model changes, when a rule changes or when someone needs to audit an answer. For a closer look at the difference between what AI does and what traditional systems should do, compare this with the article on what superintelligence is and what ChatGPT is not.
How do you organize data so a better model can use it well?
More capable models reason better about the information they receive, but they cannot compensate for contradictory data. If the catalog says one thing, the sales spreadsheet says another and the front desk uses a third criterion, the agent will reproduce that confusion more fluently. Before connecting an agent, define one official source for each type of information and record who is responsible for it.
A few minimum fields help a lot: a unique identifier for the customer or order, a status with closed values, the date of the last update, the person responsible for the information, and a marker showing what is public and what is restricted. This topic is directly tied to Brazil's data protection law, the LGPD. Before sending personal data to an external provider, you need to know the purpose, the legal basis and the conditions for the transfer. For that, read about international data transfer with AI and what the LGPD requires. Organizing data also means applying data minimization: the agent should receive only what the task needs. A practical deep dive is in data ready for superintelligence: what to standardize first.
What permissions should the agent have?
The principle that guides this part is least privilege: each agent, integration and token should have only the access needed for its role. In practice, this means separating read from write, limiting each tool to a specific set of actions, and preventing a single agent from querying, changing and sending messages across all systems. Authorization should be enforced in the target system, not just stated in the prompt. The OWASP guidance on authorization reinforces this point.
There is also a risk specific to language models: external content, such as a customer message, an email or a web page, can contain instructions that try to change the agent's behavior. OWASP describes this problem as prompt injection and recommends treating any external text as data, not as a command, and limiting what the agent can do even when it is manipulated.
Illustrative scenario: a service company uses an agent on WhatsApp to check the status of a work order. The agent can read the status and the expected date. It cannot reschedule visits, change prices or grant warranty exceptions. When a customer asks for an exception, the agent records the request with its context and hands it to a staff member who has the authority to decide. If a more capable model is adopted tomorrow, the design still holds, because the sensitive decision never depended on the model's capability.
How do you build evaluations that survive model changes?
Without evaluations, switching models is a gamble. The most useful way to start is to gather 30 to 50 real cases from your operation, including common questions, ambiguous requests, out-of-scope requests and attempts to get the flow to break. For each case, define the expected behavior: the correct answer, the action that should be taken, what must never be said, and when a handoff to a human is mandatory. Run this set whenever the model, the prompt or a tool changes.
OpenAI's documentation on evals describes this kind of testing as a way to measure system behavior against explicit criteria. Anthropic has also published material on evaluating agents, which stresses measuring complete tasks, not just isolated answers. To test without exposing real customers, simulation environments are useful, as described in the article on testing AI agents with mocks and a sandbox.
The success criteria must come from the business. Metrics such as time to first response, unnecessary handoff rate, number of information errors and cases resolved without rework tell you more than a text quality score. A good evaluation set also keeps the cases that failed, because they become the next round of tests.
What records and observability should you keep?
After any incident, the operation should be able to answer three questions: what was requested, what the system did, and why it did it. This requires logging the input, the model and prompt versions, the tools called, the response delivered and the reason for any handoff to a human. Observability concepts such as traces, metrics and events are described in the OpenTelemetry documentation, which also has conventions specific to generative AI systems.
Logs must respect privacy. You do not need to store everything in full text for every purpose. In many cases, identifiers, decisions and minimal excerpts are enough, with a defined retention period. This balance between traceability and minimization should be decided with your legal advisor or your company's data protection officer.
A six-step plan to get started
The sequence below works as a roadmap for any team that wants to prepare without stopping operations:
- List the three workflows with the highest volume or the highest cost of error, and describe how each one works today, including the exceptions people resolve from memory.
- For each workflow, identify the information sources, the owner of each one, and the fields that are outdated or in conflict.
- Split actions into three groups: look up, suggest and execute. Decide which the agent can do alone and which require approval.
- Write the deterministic rules outside the prompt, in tables, systems or official documents the agent consults.
- Build a set of 30 to 50 real cases with expected behavior, and run that set before any model change.
- Define a minimum decision log and a person responsible for weekly review of handed-off cases and errors.
After this roadmap, a model change stops being a rewrite project. It becomes a configured change, tested against the evaluation set and released with a rollback plan.
How we work follows four stages: map the current process and its baseline, limit data and actions, pilot a real workflow with logs and an agreed success metric, and then hand over the operation and decide whether it will be expanded, adjusted or stopped. This is the same logic that makes a company ready for better models: when the pilot has limits and metrics, replacing the model is a data-based decision, not a leap of faith.
Which decisions remain human?
More capable models expand what can be automated, but they do not remove the company's responsibility. Exception policies, damage repair, decisions that affect customers' or employees' rights, and the choice of when to stop an agent remain decisions for people with authority. The LGPD addresses automated decisions and the right to review them, and the topic is covered in the article on automated decisions and Article 20 of the LGPD.
It is also worth following Brazil's regulatory environment for AI, which may bring new obligations for companies that use automated systems. A practical summary is in Brazil's AI legal framework: what your company should track. Preparing your operation now, with limits and records, reduces the work of future compliance, whichever model you choose.
Frequently asked questions
What does it mean to prepare a company for more capable models?
It means designing the operation so that switching or upgrading the model does not require rewriting processes, integrations and controls. That covers organized data, minimum permissions, rules kept outside the prompt and evaluations based on real cases. When these elements exist, the model is treated as a replaceable component.
Do I need to wait for superintelligence before I start?
No. The design decisions that make an operation ready, such as organized data, least privilege and decision logs, are useful with today's models and remain useful as models evolve. Waiting for an uncertain future only delays the organization that is already needed. Starting small, with one real workflow and an agreed metric, is the most practical way forward.
How do I know if a new model is better for my operation?
The answer depends on an evaluation set built from your real cases, with success criteria defined by the business. Run the same set with the current model and the new one, compare the results and look closely at the cases that failed. Public benchmarks help compare general trends, but they do not replace testing in your own operation.
Can I use a better model without sending personal data outside Brazil?
It depends on the architecture, the provider and the contractual terms, and each case needs to be assessed with legal support. A useful practice is to minimize the data sent, mask identifiers when possible, and evaluate whether a locally hosted model meets the need. The LGPD requires that the purpose, the legal basis and the transfer conditions be clear before the use begins.