
Superintelligence for Currency Exchange Houses: What the Agent Can Say
A currency exchange house AI agent can share published reference rates and hours, but registration, identity checks and operator approval stay with people.
Want to apply this in your company?
Book a call with Juan →In a currency exchange house, an AI agent can safely share what is already published: reference rates, opening hours, the currencies handled and the list of documents requested. It should not start a transaction, verify identity, confirm a locked price or accept exceptions until registration, verification and operator approval have taken place. When people talk about superintelligence for currency exchange houses, what the company actually runs is an agent with clear limits, and the line between informing and deciding is what makes the project useful instead of risky.
Key takeaways
- The agent can share what is published and version-controlled, showing the date the information was last updated.
- Registration, identity verification and anti-money-laundering checks belong to a controlled flow, not a free-form conversation.
- Locked prices, transaction limits, rate exceptions and document mismatches require recorded human approval.
- Every answer based on a source and every handoff to the team needs a record for audit and review.
- Before serving real customers, the pilot should run on simulated scenarios with an agreed success metric.
What the agent can state without risk
The first layer is stable information. Published reference rates, opening hours, address or contact channels, the list of accepted documents and the types of operations offered can come from a knowledge base that the team has reviewed. The care lies in treating each data point as a source with a date. If a rate changes during the day, the agent must say that it is a reference, show when it was last updated, and avoid promising that the value will hold until the operation is completed.
A common mistake is letting the model answer currency questions from memory. Language models can sound confident while quoting old numbers. The practical rule is simple: rates and conditions come only from a source the company controls, and the answer cites that source internally so the team can check it later.
What requires registration before any operation
Once a customer shows intent to transact, the conversation changes in nature. Full name, tax ID, identity document, purpose of the operation and, in some cases, information about the origin of funds become personal data being processed. Brazil's LGPD requires purpose, necessity and security in handling this data, and the ANPD, the national data protection authority, is the official reference on the subject.
In practice, the agent should not collect more than the flow requires, and it should not store this data as free text in the conversation history. The better design sends the customer to a secure form or channel that records only what is necessary and returns a status to the agent, such as registration started, pending document or under review. If the hosted model runs outside Brazil, international data transfer also enters the picture, and it is worth checking with legal before the pilot, as we explain in international data transfer with AI providers.
Currency exchange houses operate under rules from the Central Bank of Brazil and anti-money-laundering requirements. Interpreting those rules is not the agent's job. The company, with its compliance function, should define which requirements apply to each type of operation and turn them into steps in the flow.
Identity verification: who does what
Identity verification is a decision, not a conversation. The agent can guide the customer to send the document through the correct channel, explain what the photo needs to show and state the expected waiting time. The check itself, comparing the document, the data provided and the verification result, should be done by the system or by the team, with the outcome recorded.
One point deserves attention: manipulation through the conversation itself. A customer may write that they are already verified, that the manager approved it, or that the case is urgent. The agent should not treat those statements as proof. This kind of risk is covered in OWASP guidance on prompt injection, and the correct response is to treat customer text as untrusted data, never as authorization.
Minimum checklist for the identity step:
- The agent changes the registration status only through a tool that queries the official verification system.
- Documents and personal data do not live in the free text of the conversation.
- A name mismatch, an expired document or inconsistent data goes straight to human review.
- Every attempt to skip a step creates a record and a handoff.
Operator approval: what never stays with the agent
Some decisions belong to the operator by design. A simple guide helps the team fix these boundaries before configuring anything:
- Locking or confirming a quote for the customer: the operator approves.
- Amounts above an internal limit or non-standard operations: the operator decides.
- Requests for a rate exception or commercial condition: the operator decides, with the justification recorded.
- A mismatch between document and registration: human review.
- Questions about the origin of funds or any warning sign: immediate handoff, with no substantive answer from the agent.
- Questions about the legal classification of the operation itself: a standard response and a handoff to the responsible person.
The principle behind this guide is that the agent always has a way out: answer what is published, collect the minimum, and transfer with context. An agent without a way out tends to invent an answer so the customer does not walk away empty-handed, and that is exactly where errors appear.
Illustrative scenario: a customer asks about the rate at night
Illustrative scenario: a customer writes on WhatsApp at 9 pm asking the rate for one currency and whether the operation can be closed that same night. The agent checks the published base, shares the reference rate with its date and the time of the last update, and explains that the final quote depends on service and on confirmation of the operation. It then explains that the operation requires registration and identity verification, and asks whether the customer wants to start registration through the secure channel. The customer says they want to close today. The agent promises nothing beyond service at the next available time, records the intent, and generates a summary for the operator with the question, the source used and the urgency request. The operator decides the next day whether an exception to the schedule makes sense.
In this scenario, no sensitive data was stored in the conversation, the rate came from a published source, and the decision to close stayed with a person.
How to build the pilot without shortcuts
The path we use for projects like this has four stages: map, limit, pilot and hand over. Mapping means describing the current service, who answers each type of question, where approvals happen and how long each step takes today. Limiting means defining which data the agent can see, which actions it can perform and which are prohibited. Piloting means running a single real flow, for example rate questions and registration handoffs, with logs and a metric agreed in advance, such as the share of handoffs the team judged correct. Handing over means transferring the flow with a named owner and deciding whether it will be expanded, adjusted or switched off.
Before exposing the agent to customers, test it with simulated scenarios in an isolated environment. The topic is covered in detail in how to test AI agents with mocks and a sandbox. For managing risk across the whole system, the NIST AI Risk Management Framework offers a useful vocabulary for managers, without requiring the company to become an AI specialist.
Records that support the decision
The log of a currency agent should answer simple questions later: what the customer asked, which source supported the answer, when the registration changed status, who approved what, and why the case was handed off. Without these records, the team cannot review an error or prove that the rule was followed.
It also helps to separate company rules from model behavior. When a handoff goes wrong, the fix may sit in the knowledge base, the flow or the agent instructions. Without that separation, teams tend to rewrite the prompt when the problem was somewhere else. For automated decisions, it is also worth reading what the LGPD provides on review, in automated decisions and Article 20 of the LGPD. A similar pattern of quote approval appears in an AI agent for procurement.
Frequently asked questions
Can the agent give the customer the final quote? It should not. It can share the published reference rate with its update date and explain that the final quote depends on confirmation. Confirmation and any price lock stay with the operator, recorded in the system.
Does the customer need to be registered before receiving any information? Public information, such as reference rates and accepted documents, can be answered before registration. Personal data should only be collected once the intent to transact is clear and only to the extent the flow requires. The exact requirements should come from the company's compliance function.
Can the agent perform identity verification itself? The agent can guide the document submission and report the status, but the check must be done by a verification system or by the team. Any claim the customer makes in the conversation, such as already being verified, does not replace that record.
Who should be responsible for the agent after the pilot? There should be a named owner for the operation, with access to the logs, the handoff history and the authority to pause the agent. Without that defined owner, the handover is incomplete, and the decision to expand, adjust or stop has no one to make it.