
Data Ready for Superintelligence: What to Standardize First
Which customer records, field names and data owners to standardize before any AI agent. A practical checklist that stays useful as models change.
Want to apply this in your company?
Book a call with Juan →To prepare company data for superintelligence in practice, standardize what any AI agent already needs today: complete records without duplicates, field names and statuses with one clear meaning, and a named owner for every data source. None of these depends on a specific model, so the work stays useful as models change. Superintelligence, as a concept, is still under debate and is not a product you can buy. Investment makes the most sense when it solves today's operation.
Key takeaways
- Clean records usually affect the quality of answers more than swapping the model, because duplicates and empty fields produce errors that look confident.
- Field names and statuses must have one meaning across the company, documented in a single place.
- Every data source needs a human owner who is accountable for accuracy, access and updates.
- The agent should read data through tools with limited permissions, never through unrestricted access to spreadsheets and systems.
- Standardizing is a small, repeatable process built from a real workflow, not a full reformulation of your data.
Why start with data and not the model
Models change often, but your operational data stays. An agent that looks up orders, contracts or schedules is only reliable if it finds the same record, with the same name, everywhere that record appears. Illustrative scenario: a customer appears as “Maria Souza” in one system, “M. Souza” in another and “Maria de Souza Lima” in a spreadsheet. No model can know on its own which record is the right one.
Anthropic's documentation on context engineering treats an agent's context as a limited resource that should be lean and informative. The same logic applies here. Before adjusting instructions, make sure the information you give the agent is clear, unique and up to date.
Record hygiene: the minimum every agent needs
Record hygiene does not require a complex data project. The items below solve most of the problems that show up in an agent's first tests:
- A unique identifier for each customer, product, location or order, which does not change when a name or phone number changes.
- Phone numbers and emails in a standard format, with area code and country code, and a flag showing which contacts are valid.
- A written rule for merging duplicates, stating which record prevails and what happens to the other record's data.
- Required fields defined for each record type, with the action to take when one of them is empty.
- A last-updated date visible on every record, so the agent knows when information may be outdated.
- Consent and contact restrictions recorded separately from the commercial record.
The last item deserves special attention. Personal data handled by a Brazilian company falls under the LGPD, and the record is where contact permission must be stored in a way that can be checked. When an agent sends outbound messages, this information determines whether a message may be sent, not just whether a number exists. On this topic, read WhatsApp Outbound Messages: Templates, Opt-In and Best Practices.
Field names and statuses: one shared vocabulary
For a person, “pending,” “waiting” and “open” can look like the same thing. For an agent, they are different states, and each can lead to a different action. Before connecting any tool, answer these questions for every status:
- Who changes this status, and which event triggers the change.
- Which action the agent may take while the record is in this status.
- Which status the system assumes when the information is missing.
The same care applies to field names and tool names. Anthropic's published guidance on writing tools for agents recommends clear names and descriptions, with parameters that leave no room for interpretation. In practice, a tool called “lookup_order” with an objective description works better than “get_data_v2.” Choose one language for technical names, keep the labels your team uses in the local language, and document how the two correspond.
Data owners: who is accountable for the information
A data owner is not the person who types the information in. It is the person who decides what it means. Each source needs a business owner, who defines what the record means and who may change it, and a technical owner, who handles access, integrations and backups. In companies that handle personal data at scale, the data protection officer appointed under the LGPD (the encarregado) usually takes part in these decisions, especially those involving new purposes.
Without a defined owner, the agent inherits the mess. When an error appears in a customer reply, someone must have the authority to fix the source, not just the copy the agent read. A backup owner is also necessary, because vacations and staff departures are when data sources stop being updated.
Checklist: standardize before the first agent
The safest path is to start with one workflow, not with every system in the company. A possible sequence:
- Choose one real workflow, such as checking order status, and list the data sources and fields it touches.
- For each source, name an owner and a backup with decision authority.
- Export a sample and note duplicates, empty fields and undefined statuses, without fixing anything blindly.
- Write the vocabulary for this workflow, with one sentence of meaning for each field and status.
- Fix the source data and record the rule for merging duplicates.
- Give the agent only the read access the workflow requires, with a log of every query it makes.
- Test with known records before opening the workflow to real customers.
For the testing stage, use simulated environments. The guide Mock Tools for Testing AI Agents: Safe Sandbox Testing Guide explains how to test without touching real data.
Illustrative scenario: one customer, three records
Illustrative scenario, not a real case: a service company keeps its schedule in one system, its customer relationships in a CRM and its follow-up reminders in a spreadsheet. A customer asks when their next visit is, and the agent finds two records with the same name and different phone numbers.
Without an identifier rule, the agent picks one of the records and answers confidently, even if it is the wrong one. With standardization, the agent confirms the full name and date of birth or the order number, queries only the schedule of the identified customer, and if doubt remains, hands the conversation to a person with a short summary of the history. The gain does not come from the model's intelligence. It comes from a unique identifier and a written handoff rule.
How Juan Carlo's method applies here
In our approach, standardization happens in the mapping and limiting stages. First we map the process and the data sources it uses, then we restrict which fields and actions the agent can access. In the pilot, one real workflow runs with a log of decisions and a metric agreed before launch. Only then is the operation handed over, with a decision to expand, adjust or stop. This does not guarantee results, but it grounds the decision in evidence.
Frequently asked questions
Do I need a data warehouse or a large project to start? Not necessarily. Start with the workflow you want to automate and the data sources it touches. Large data projects often delay operations, while standardizing by workflow produces rules you can reuse for the next ones.
Is standardizing data only for AI agents, or does it help the team too? It helps both, and that is a good reason to start. The same identifiers, statuses and fields the agent reads are the ones your team uses in reports and in human customer service. Once the rule is written, it also applies to people who join the team later.
Does personal data need extra care when feeding an agent? Yes. Personal data falls under the LGPD, so the purpose, the minimum access and the log of each query must be defined before the pilot. Before sending information to third-party services, check where that data is processed, as explained in the article on International Data Transfer Under LGPD When Using AI Providers.
What if the model changes later? Standardization still holds, because identifiers, owners, statuses and vocabulary do not depend on the model. Changing models may require instruction adjustments and new tests, but the work on the data stays. Before switching, run the same tests with known records and compare the results.