
Questions Before Investing in Company AI Agents: Ten to Answer First
Ten questions to answer about process, data and owners before hiring an AI agent, with a practical checklist to help you decide and start a pilot.
Want to apply this in your company?
Book a call with Juan →Before hiring any AI agent, a manager needs answers about the process, the data and the people responsible, not about the model. "Superintelligence" is mostly a term for debate, not a product you buy. What a company actually runs are AI systems with a narrow scope, clear rules, controlled access and human oversight. When those answers don't exist, no vendor can deliver a result you can measure.
Key takeaways
- An agent only improves a process that is already described, with known inputs, outputs and exceptions.
- Data needs an owner, an authorized purpose and a minimum level of consistency before any integration.
- Every agent action needs an explicit limit, including what it must never promise or execute on its own.
- Without an agreed success metric and a record of decisions, there is no way to know whether the pilot worked.
- Who answers when the agent makes a mistake comes before the question of which tool to use.
Why the conversation starts with the process
Many purchases start with a demo. A demo shows the best case: a well-formed question, a customer who follows the script, a system that answers without errors. The person running the operation needs to know the real path, including the message that arrives on Friday night, the customer who changes topic halfway through, and the exception that only the manager knows how to resolve.
The questions below don't require technical knowledge. They require someone from operations to answer honestly. If the answer to most of them is "we don't know yet," the right step is to map the process, not to buy a solution.
About the process: what exists today
The first block covers what the company does now, before any automation. The answers show whether there is something stable enough to delegate.
- Is the process written down anywhere, or does it exist only in the heads of the people who perform it, and which steps does it really have?
- What triggers the service or task, and through which channel does it arrive?
- Which exceptions come up often, and who decides what to do with them today?
- At which points does the decision require human judgment, and at which points does it follow a clear rule?
The last question is the most useful for separating what a deterministic workflow can handle from what needs AI. Many steps work better with simple rules and checks, and AI is only needed where natural language or contextual interpretation is involved. Combining the two is what we call process automation with AI where judgment is needed.
About the data: what the agent can see and use
The second block covers information. An agent can only respond with what it can query, and a data error becomes a service error that still looks confident.
- Which information does the agent need to look up in order to answer or act, and in which system does each piece live?
- Can this data be used for this purpose, with a legal basis and handling that comply with the LGPD (Brazil's general data protection law)?
- Is the data consistent enough? For example, does the same status or the same amount appear differently in spreadsheets, systems and messages?
On purpose, remember that data collected for one function does not automatically become available for another. The Brazilian data protection authority (ANPD) publishes the main guidance on this topic, and the company should document what it decided. When the agent makes decisions that affect the individual, Article 20 of the LGPD covers the right to request a review, and that right needs to be built into the workflow design. The details are in LGPD Automated Decision-Making: The Right to Review (Art. 20).
If the database is disorganized, the first task is to standardize it, not to hire an agent. This is covered in Data Ready for Superintelligence: What to Standardize First.
About the owners: who decides and who answers
The third block is the most overlooked, and it is often what stalls a project after the first week. An agent without an owner becomes a shared problem, and nobody fixes a shared problem.
- Who owns the process and has the authority to change its rules when the agent fails?
- Who approves sensitive actions such as refunds, discounts, cancellations or final quotes?
- Who reads the conversation logs and responds to the customer when there is a complaint?
The answer to the last question defines the handoff policy to humans. On channels like WhatsApp, the company also has to follow platform rules about messages sent outside the customer service window, which directly shapes how the agent is designed.
Turning the answers into a pilot
Once the questions are answered, the next step is to choose one real, small workflow to test. The sequence below works as a guide and can be adjusted to your case.
- Choose a workflow with enough volume to observe and with limited impact if something goes wrong.
- Write down what the agent may answer, what it may execute, and what always goes to a person.
- Give the agent the minimum access it needs, preferring read access before write access.
- Define a success metric and the criteria for adjusting or stopping before you start.
- Log every conversation and every action, with the reason, the source consulted and the result.
- Review samples regularly with the process owner, not only with the people who built the solution.
- At the end of the period, decide whether the operation will be expanded, adjusted or shut down.
The NIST AI Risk Management Framework organizes AI risk management into functions such as govern, map, measure and manage, which fits well with this controlled pilot logic. For security of instructions and tools, the OWASP Top 10 for LLM Applications lists common risks, such as prompt injection through external text.
Illustrative scenario: the status check that became three answers
Illustrative scenario: a repair shop receives many WhatsApp questions every day about the progress of devices under repair. The director imagines an agent that answers everything on its own. While mapping the process, the team finds that the status lives in three places: the work order system, a spreadsheet at the bench and the memory of one technician. Warranty decisions also depend on the manager, who currently makes them by phone.
With that map, the decision changes. The first pilot becomes only the status check, reading from the work order system, with no permission to change anything. Warranty, quotes and exceptions stay with people, and the agent passes the case along with the full history. Before expanding, the shop standardizes the status in one single place. The agent did not become smarter. The operation became clearer, and the agent started working on top of it.
When it is not yet time to hire
The readiness question itself is a useful filter. Some signs suggest the investment should wait.
- There is no agreement on what counts as a correct answer for the most common cases.
- The data the agent would need sits in personal channels, such as private chats and notebooks.
- Nobody accepts responsibility for sensitive actions.
- The company cannot describe what it would do if the agent made a mistake with an important customer.
- Nobody has defined how to tell whether the pilot succeeded.
In these cases, the useful work is diagnosis and design, not implementation. This is the stage where a manager gains the most control over the investment.
How Juan Carlo's method organizes these questions
In how we work, these questions map to four stages: map the process and the current baseline, limit data and actions, pilot a real workflow with logs and an agreed metric, and then hand the operation over and decide whether it will be expanded, adjusted or stopped. The first stage usually takes the longest, and it is the one that prevents the most rework. To understand what a company actually operates when it talks about advanced AI, read Difference Between AGI, ASI and AI Agents: What You Can Run.
Frequently asked questions
Can a company hire superintelligence today?
In practice, companies buy AI systems with a defined scope, such as agents for customer service, triage or information lookup. These systems work well when the process, the data and the owners are clear. Broader terms are useful for discussion, but they do not replace a specification of what the agent must do.
Do I need perfect data to start a pilot?
You do not need perfect data, but you do need to know which data will be used and where it comes from. A pilot can start with one reliable source and a small scope, leaving the rest of the standardization for later. What matters is that known inconsistencies are recorded, not discovered by the customer.
What if nobody wants to own the AI process?
That is a sign the decision to invest is not yet mature. Before any hiring, leadership needs to name a person with the authority to change rules and approve exceptions. Without that name, the agent has nobody to report errors to, and the errors pile up without correction.
Can the agent make decisions with customers on its own?
It depends on the type of decision and its impact on the person. Sensitive actions, such as denying a refund or rejecting a request, should require human approval and a review channel, in line with data protection and risk management guidance. For informational, low-impact answers, the agent can act on its own, as long as there are limits, logs and a handoff to a person when needed.