
How to Prepare a Company for Superintelligence: A Maturity Checklist
A maturity checklist for preparing your company for more capable AI models: data, processes, permissions and governance, step by step, without hype.
Want to apply this in your company?
Book a call with Juan →Preparing a company for superintelligence, in practice, means getting ready what any more capable model would need to find inside the business: organized data, mapped processes, minimal permissions, and one person responsible for decisions. No product can be bought to make a company ready. What helps is a set of conditions that apply to any scenario involving more capable models, including the ones already arriving today that change customer service, sales, and operations.
Key takeaways
- Maturity depends on data, processes, permissions, and governance that the company controls, not on the model you choose.
- Before any agent, map the real process and record a performance baseline.
- Give the system only the data and actions the task requires, and log every use.
- Treat governance as a routine: who approves, who reviews, and who has the authority to pause the system.
- Run a pilot with an agreed metric before deciding whether to expand, adjust, or stop.
What being ready when the model changes means
Models change often, so the useful question for leadership is not which model will arrive next, but whether the company can swap, measure, and review what it already uses. An operation that depends on tacit knowledge held by one person is fragile with any technology. An operation with documented processes, reliable data, and clear responsibilities can absorb a model change without stopping.
It also helps to separate conceptual debate from operational decisions. The superintelligence glossary for executives helps align vocabulary before internal conversations, and the difference between AGI, ASI and AI agents shows why planning should start from what the company actually runs today.
Data checklist: what needs to be organized
A model is only as good as the information it receives, and bad information that looks reliable is the worst case. Before any automation, check the points below.
- You know which data sources exist, who owns each one, and how often they are updated.
- Customers, products, prices, and order statuses follow one standard, not three versions in different spreadsheets.
- There is a written list of what must never enter any AI system, including personal data without a defined purpose and legal basis.
- You can say where each piece of information comes from when the system uses it to answer a customer or suggest an action.
- Old data has a rule for deletion or updating, so the system does not repeat what has already changed.
If you answer no to two or more of these points, preparation starts with organizing the data, not with choosing a tool. The article on data that is ready for superintelligence covers what to standardize first.
Process checklist: what you need to know before delegating
A process that exists only in someone's head cannot be delegated to any system, because nobody can verify whether the delegation is correct. Mapping means writing the flow as it happens today, including the exceptions that people rarely like to mention.
- The flow is written down with inputs, outputs, owners, and deadlines.
- The most common exceptions are listed, along with how they are resolved today.
- There is a current performance record, such as response time, rework rate, or number of complaints, to compare against later.
- It is clear which decisions require human judgment, such as offering a discount outside the price list or accepting a deadline exception.
- You know which steps are deterministic, meaning they follow a fixed rule, and which ones truly require interpretation.
This last distinction matters a great deal. Steps with fixed rules are usually better handled by traditional automation, and AI is reserved for the part where language or context is ambiguous. Mixing the two without a clear criterion is one of the most common sources of problems.
Permissions checklist: the system should only be able to do what it needs
Permissions are where maturity shows most clearly, because that is where an error stops being a bad answer and becomes an action in the real world. The general security recommendation is to give each system the smallest set of access rights the task requires, a basic access control principle described in the OWASP Authorization Cheat Sheet.
- Each agent or integration has its own identity, not a personal employee login.
- Read and write access are separate permissions, and most cases should start with read-only access.
- Actions that move money, change customer records, or send bulk messages require recorded human approval.
- Credentials stay out of prompts and shared spreadsheets, with rotation and revocation rules defined.
- There is one central place where you can see which tools the system may call and which are turned off.
When the system uses external tools, also review the risk of instruction injection in text it reads, such as customer emails or web pages. The OWASP Top 10 for LLM Applications lists these risks and can serve as a review guide for whoever operates the application.
Governance checklist: who decides, reviews, and stops
Governance sounds bureaucratic, but in practice it answers three questions: who is responsible for the system, how you know it is working, and what happens when it is not. The NIST AI Risk Management Framework organizes this work into functions of governing, mapping, measuring, and managing risk, and it can be used as a reference in conversations with the board.
- Each system has a named owner with the authority to pause it.
- A written policy defines what staff may and may not put into AI tools, as the company policy for employee ChatGPT use proposes in general terms.
- Decisions that affect people, such as refusing credit, cancelling a contract, or classifying a customer, have human review and a channel for appeal.
- The system's decision logs are stored and can be consulted by anyone who needs to audit them.
- Personal data handling follows Brazil's LGPD (the general data protection law), with a clear purpose, data minimization, and a defined owner for incident response.
- There are written criteria for stopping, such as an error rate above the agreed limit or a recurring complaint about the same issue.
The last item is the one most often forgotten. Stopping a system must be a planned procedure with a clear owner and clear steps, not an improvised decision in the middle of a crisis.
A four-step plan to start without overbuilding
You do not need to complete the whole checklist before starting. The safest path is to move in short cycles, one real process at a time.
- Choose a process with meaningful volume, a clear rule in most cases, and an identified owner.
- Map the current flow and record the performance baseline over a few weeks.
- Define what the system can read, what it can write, and what always requires human approval.
- Run a pilot with a record of every decision, comparing results against the baseline and the metric agreed before the start.
- At the end of the pilot, decide whether to expand, adjust, or stop, and document the decision and the reason.
After the first cycle, the checklist for data, processes, permissions, and governance becomes a review tool rather than a one-time project.
Illustrative scenario: an electrical materials distributor
Illustrative scenario: a distributor wants to use an agent on WhatsApp to answer questions about stock availability and delivery times. During the review, the team finds that inventory sits in two spreadsheets that use different product names, and that the delivery rule changes by region, but nobody has written that rule down. The pilot does not start with the agent. It starts with standardizing the product catalog and documenting the delivery rule. Only after that does the agent get read access to stock, without the power to reserve items, and any exception request goes to a human salesperson with the full conversation history. If the rate of correct answers does not improve against the baseline, the decision will be to fix the data source before any expansion.
How Juan Carlo's method applies to this preparation
The method we use in consulting follows four moves: map the process and its baseline, limit data and actions, pilot a real flow with records and an agreed metric, and hand over to the team with a decision to expand, adjust, or stop. It does not depend on predicting which model will arrive. It depends on leaving the operation in a condition to make good use of any model that makes sense for the business. Companies that want to start with a diagnosis find the same starting point in AI for construction and real estate developers, covering buyer support and after-sales, where organizing information comes before any agent.
Frequently asked questions
Do we need to buy a specific tool to prepare a company for superintelligence? No. Preparation is done through data organization, process mapping, restricted permissions, and governance, which apply to any model. Tool selection comes later, once the process and success criteria are clear.
Where should a company start if it does not use AI in any process yet? Start with a process that has volume and a clear rule, preferably one with low risk to customers. Map the current flow, record performance for a few weeks, and only then define the pilot. That starting point creates the baseline you need to judge any result later.
Do we have to review everything the system does by hand? Not necessarily, but decisions that affect people, move money, or change records should have recorded human approval or review. The rest can be monitored through sampling and agreed indicators. The level of review should be set by the risk of each action.
How do we know the company is mature enough to expand? The company is ready to expand when the pilot reached the agreed metric, the decision logs are complete, a named owner exists, and there are written criteria for pausing. If any of these fails, the right move is to adjust before growing.