Whether or not you believe that AI agents are digital (co-)workers, there are profound ethical questions to consider when your software becomes more autonomous. Generative AI models acting on your company’s behalf widen ethical risks. Instead of publishing customer-facing AI assistants and agents, start within your company first, and require traceability, oversight, and low-stakes testing before handing agents any real power.
Rebecca Bultsma, AI Ethics & Responsible AI Consultant, joined me on “What’s the BUZZ?” to discuss how to teach your AI agents ethical behavior. This is especially relevant when agents have their own optimization functions, which might diverge at times. Here’s what we’ve talked about…
Why AI Agents Raise the Stakes for Leaders
Generative AI models behave in human-like ways, so the ethical questions that used to be niche are now front-page concerns. For leaders, that means bias, hidden assumptions, and opaque decision paths move from academic problems to practical business risks.
Agentic AI systems are trained on large amounts of internet data, which reflects certain cultural frames and gaps. That can skew outputs in ways that harm customers or employees. When you give an agent the power to act (e.g., to purchase, submit an application response, or change a setting), you multiply the consequences. The model won’t “intend” harm, but it may optimize for a goal set by vendor defaults or the wording of the prompt instead of your brand values.
Think in terms of incentives: is the agent optimizing for the platform’s goals or for your customer’s best outcome? Without visibility into how a choice was made, accountability lands squarely on you and your organization. That’s why understanding model limitations, monitoring behavior, and treating agent deployments as a governance issue are practical necessities for leaders.
ONLINE COURSE — Mitigate AI Business Risk
Business leaders are under pressure from their boards and competitors to innovate and boost outcomes using AI. But this can quickly lead to starting AI projects without clearly defined, measurable objectives or exit criteria.
Learn how to implement proven risk mitigation strategies for starting, measuring, and managing AI projects. Along the way, get tips and techniques to optimize resourcing for projects that are more likely to succeed.
When everyone at a vendor’s demo table says “agents,” take a breath. Don’t expose customers or the public until your people have hands-on experience and you’ve built guardrails. Begin with controlled, internal scenarios on low-stakes tasks so your team can observe failures and patterns without risking reputation or legal exposure.
Set up experiments using personal test accounts first. Script typical use cases and edge cases. Log inputs and outputs. Ask the team: where did the agent fail? What assumptions surfaced? Pair these sessions with plain-language training so non-technical staff know agent limits. For example, running an internal agent for meeting scheduling or basic HR FAQs is a far safer place to learn than letting it handle payroll changes or applicant screening externally.
Also, build basic constraints into the agent from day one: time limits, spending thresholds, and approval gates for actions beyond a certain risk level. These small controls convert a theoretical worry into observable behavior you can iterate on. When you’re confident internally, consider a staged external rollout with monitoring in place.
Thanks for reading The AI MEMO! Share this thought-provoking post with your network.
Require Traceability, Oversight, and Legal Awareness
If an agent is doing anything that could impact people, such as money, access, hiring outcomes, or emotional support, you need audit trails and clear human accountability. Too often, we assume the tech will explain itself; it won’t. Require logging of prompts, responses, and agent decisions. That lets you trace back if something goes wrong and supports remediation.
Designate human owners for each agent workflow. The owner isn’t the agent; it’s a person who reviews logs, checks for bias, and can shut the agent down. Use approval layers for higher-risk actions (finance, identity, applicant data). If you allow agents to buy products or transfer funds, cap limits, and require human confirmation beyond a low threshold.
Keep an eye on emerging laws. New rules in some places already target chat interfaces and how they interact with minors or give advice. That legal landscape will keep evolving, so involve legal and compliance early. Finally, plan communication: if a customer or employee is interacting with an agent, be clear they are talking to an automated system and what safeguards exist. Transparency reduces confusion and forms part of your duty of care.
Summary
You don’t need to sprint to external agent rollouts. Three things to take away: generative models amplify ethical risk and hidden bias; learn inside your walls first — run low-stakes experiments and limit autonomy; and require traceability, human oversight, and legal review before agents act on behalf of people or your brand.
Now that you’ve read this, do three practical things this week:
Run one internal test using a personal account and document where the agent fails.
Draft a simple approval flow with caps for any agent that can take actions (spending, access, hiring).
Assign a human owner and set up logging for any agent in the pilot.
If you treat agents as governance problems first and tech opportunities second, you’ll avoid being the organization that learns the hard lesson in public. Take the time to learn, test, and protect — then scale.
Equip your team with the knowledge and skills to leverage Agenti AI effectively. Book a consultation or workshop to accelerate your company’s AI adoption.
Join my podcast for leaders and hands-on practitioners. Each episode features a different guest who shares their AI journey and actionable insights. Learn from your peers how you can lead artificial intelligence, generative AI, agentic AI, and automation in business with confidence.
