Responsible AI Policy
Last updated: July 2026 — how AgentCorp keeps AI agents overseen, bounded, and safe.
1. Our Approach
AgentCorp provides a workforce of specialized AI agents — Alex (executive assistant and orchestrator), Sam (sales), Riley (finance), and Taylor (marketing) — that businesses can 'hire' to perform real work. Because these agents take actions that affect your business, we design them to keep a human in control, to be transparent about their limits, and to fail safely. This policy describes the principles and safeguards that govern how our agents operate.
2. Human-in-the-Loop Approvals
Sensitive or irreversible actions require explicit human approval before an agent proceeds. This includes sending external communications, moving money, altering records in connected systems, and other high-impact operations. Agents surface a clear summary of what they intend to do so a person can approve, edit, or reject it. Human oversight is the default posture for consequential actions, not an optional add-on.
3. Agent Autonomy Boundaries
Each agent operates within a defined scope of tools, data, and permissions granted by your organization. Agents cannot exceed the integrations and approvals you have configured, cannot access another tenant's data, and cannot grant themselves new capabilities. Where a task falls outside an agent's boundaries or confidence, it stops and asks rather than improvising a workaround.
4. Model Providers
Our agents are powered primarily by Anthropic's Claude models, with xAI's Grok used as a secondary provider for specific workloads such as the marketing agent. We may add or change model providers over time to improve quality, safety, or reliability, and we maintain a current list of the providers that process data on our behalf at /legal/subprocessors. All providers are engaged under agreements that restrict how they may use data.
5. Training on Customer Data
We do not use your organizational data to train AI models without your explicit, written opt-in consent. Our model providers process your prompts and content to generate responses for you, but under our agreements they do not use your inputs or outputs to train their foundation models. If we ever offer features that would use your data to improve models, participation will be opt-in and clearly described.
6. Accuracy and Hallucination Disclaimer
AI agents can produce output that is inaccurate, incomplete, or fabricated ('hallucinations'), and they can misinterpret ambiguous instructions. Agent output is a tool to assist human judgment, not a substitute for it, and should not be treated as legal, financial, tax, or professional advice. You are responsible for reviewing agent output before relying on it, and human approval gates exist in part to give you that checkpoint on consequential actions.
7. Prompt-Injection and Adversarial Input Defenses
Agents that read external content — emails, documents, web pages, CRM records — can encounter instructions embedded by third parties that attempt to hijack the agent ('prompt injection'). We apply layered defenses, including separating untrusted content from trusted instructions, constraining the tools an agent may invoke, and requiring human approval before high-impact actions. No defense is perfect, so approval gates and least-privilege permissions remain an essential backstop.
8. Fairness and Bias
AI models can reflect biases present in their training data. We select providers that invest in safety and evaluation, and we encourage you to keep a human in the loop for decisions that affect individuals — for example, hiring, credit, or other consequential outcomes. Agents are not designed to make final decisions about people, and you should not configure them to do so without appropriate human review.
9. Transparency
Where an agent communicates on your behalf or produces work product, we aim to make its role clear to you within the platform. We surface which agent performed an action, what tools it used, and what it is about to do at an approval gate, so you can understand and audit agent behavior rather than trusting it blindly.
10. Escalation and Human Oversight
When an agent is uncertain, encounters an error, or reaches the edge of its authority, it escalates to a human rather than guessing. Orchestration is designed so that Alex can route work between agents and back to you, ensuring that ambiguous or high-stakes situations end up in front of a person who can decide.
11. Your Responsibilities
Responsible outcomes are a shared effort. You are responsible for configuring appropriate approval gates and permissions, reviewing agent output before acting on it, ensuring your use of agents complies with applicable law and our Acceptable Use Policy, and not deploying agents for decisions that require human judgment they cannot provide.
12. Reporting Concerns and Changes
If you observe agent behavior that seems unsafe, biased, or otherwise concerning, tell us so we can investigate and improve. We may update this policy as our agents, providers, and safeguards evolve, and we will note material changes with the date below.
Questions or concerns about our AI practices? Email privacy@agentcorp.work.