Praxian AISolution Center
Trustworthy AI agents

Risks of poorly built agents and the controls adopted by Praxian AI

Artificial intelligence agents can profoundly transform the way an organization works. They can analyze large volumes of information, interact with different systems, identify opportunities, prepare recommendations and execute process steps quickly and at scale.

Broad capability, controlled authority
Agent capability Broad
Authority granted in stages
  1. Observes
  2. Recommends
  3. Prepares
  4. Executes

Autonomy is expanded by case type and always based on evidence.

What problems a poorly built agent can present

  1. Plausible but incorrect answers

    Language models can produce very well-formulated answers even when the information is incomplete or their interpretation is wrong. A convincing answer, therefore, should not automatically be considered true.

  2. Misinterpretation of the request

    When instructions are ambiguous or context is insufficient, the agent may misinterpret the user's intent and follow a path different from the one expected.

  3. Disregard of rules and exceptions

    An agent may know a rule and still fail to apply it in a given situation. The risk increases when mandatory rules exist only in documents or written instructions, without blocks implemented in the system.

  4. Excessive autonomy

    Agents with excessive permissions can change records, send messages, access data, move resources or trigger other systems beyond what the task requires.

  5. Improper use of sensitive information

    Without an appropriate architecture, the agent may receive more information than it needs, retain data for too long, present content to unauthorized people or use technology services not approved for a given type of information.

  6. Execution based on incomplete data

    The agent may try to complete an analysis even when documents are missing, sources disagree or the case falls outside the situations it was prepared for.

  7. False confirmation of completion

    An agent may state that an action was performed without checking the target system. For example, it may report that a record was updated or a document was sent without independent confirmation of that operation.

  8. Different behavior after changes

    Replacing the model, changing an instruction, adding a new source or updating an integration can change the agent's behavior. For this reason, the performance achieved during implementation cannot be considered permanent.

How Praxian AI reduces these risks

  1. Precise definition of each agent's authority

    Before implementation, we establish what the agent can consult, recommend, prepare and execute. We also define prohibited actions, operational limits and situations in which the activity must be handed over to a person. This way, the agent's intelligence can be broad, but its authority remains controlled.

  2. Separation of analysis, decision and execution

    In relevant processes, we avoid having a single component interpret the information, make the decision, execute the action and confirm on its own that everything was done correctly. The agent can analyze and recommend; independent rules check the conditions; a person approves when necessary; a controlled component executes; and the system verifies the result.

  3. Mandatory rules implemented outside the LLM

    Financial limits, required documents, approval authority levels, access restrictions and other critical conditions are not merely described in the instructions given to the agent. Whenever possible, they are turned into system rules that block improper operations. As a result, security does not depend solely on the LLM's ability to interpret and follow an instruction.

  4. Least necessary access

    Each agent receives only the data, systems and tools essential to its function. Credentials are specific, permissions are limited and operations can be interrupted or revoked when necessary.

  5. Protection of sensitive information

    Projects can incorporate data minimization, masking, environment segregation, role-based access control, encryption, retention policies and traceability of queries performed. We also assess which information can be processed by each provider and when private or higher-control alternatives are needed.

  6. Trusted sources and verifiable evidence

    Whenever the activity requires factual accuracy, the agent is instructed to consult previously approved sources, such as official documents, corporate databases and client systems. In addition, its relevant conclusions must be accompanied by the evidence used. The system distinguishes between information obtained from a source, the agent's interpretation, a recommendation and an action actually executed.

  7. Stop and escalation criteria

    The agent is prepared to recognize situations in which it should not decide or act on its own. In these cases, the correct response is not to improvise a solution, but to stop the flow and escalate the situation to the appropriate person in charge.

  8. Human oversight proportional to risk

    Human involvement is defined according to the impact of each activity. Standardized, reversible, low-risk tasks can receive greater automation. Critical decisions remain subject to review or approval by accountable professionals. This makes it possible to achieve significant productivity gains without transferring to AI responsibilities that must remain under human control.

  9. Gradual rollout of autonomy

    Agents can begin operating by only observing the process, without interfering in it. Next, their recommendations are compared with the professionals' decisions. Only after demonstrating adequate performance does the agent begin to prepare or execute certain actions. Autonomy is expanded by case type and always based on evidence. If performance deteriorates, it can be reduced immediately.

  10. Continuous monitoring

    After implementation, we track indicators such as identified errors, human interventions, exceptions, limit violations, discrepancies between sources, unauthorized access attempts, recommendation quality and results actually achieved. Monitoring makes it possible to identify behavior changes, correct problems and drive the agent's controlled evolution.

  11. Change control and rollback capability

    Changes to models, instructions, tools, information sources or business rules go through versioning, testing and acceptance proportional to risk. Containment mechanisms, operation suspension and change rollback can also be adopted, preventing unexpected behavior from spreading through the process.

Situations that require escalation

The agent must stop the flow and escalate the situation to the appropriate person in charge when it identifies:

  • Missing mandatory information
  • Conflicting sources
  • Unforeseen situation or request outside its scope
  • Low confidence
  • Possible legal, clinical, financial or reputational consequences
  • Unauthorized access attempt
Trustworthy AI agents

Sophistication with control

Praxian AI does not believe that security means restricting artificial intelligence to simple tasks. An agent can analyze thousands of documents, integrate different systems, identify inconsistencies, generate complex recommendations and coordinate multiple steps of a process without receiving unrestricted authority to make critical decisions.

Our approach combines:

  • Generative intelligence and specialized models
  • Deterministic rules and controlled integrations
  • Information security and independent validation
  • Human oversight, traceability and continuous monitoring

No complex system is entirely free of risk. The professional commitment lies in identifying these risks, reducing their likelihood, limiting their consequences and creating mechanisms to quickly detect and correct any deviation.

It is this architecture of trust that makes it possible to use AI agents in relevant and sophisticated processes, while preserving the organization's control and the accountability of its managers.

At Praxian AI, the agent's capability can be broad. Its authority, however, is granted in a controlled, progressive manner, proportional to risk.

Schedule a Strategic Diagnosis