Use these 8 tips to build an AI agent responsibly

Start small, test, then iterate.

The Scoop: Edelman Trust Barometer shows a slip into ‘insular’ mode

Wellness Travel & Spa PR Strategies That Convert

AI agents can analyze data and generate insights faster than ever.

But without careful structure, these tools can produce errors, risk data privacy or damage brand reputation.

Martin Waxman, adjunct professor at York University’s Schulich School of Business and advisor at Ragan’s Center for AI Strategy, shared a framework for building agentic AI responsibly.

“Agentic AI isn’t a robot stomping down the street. It’s a system that can work semi- or fully autonomously, and now is the best time to start figuring out how to build it responsibly,” said Waxman, who speak at AI Horizons Conference next month.

Here’s how PR and comms pros can get started.

Identify agent-ready tasks
Look for repetitive, structured tasks that can be partially or fully automated. AI performs best with predictable, step-by-step work. “You could create an agent that pulls signals from sources like Meltwater, analyzes patterns based on your instructions or produces a report or summary,” Waxman said. Make a list of daily tasks, like social media monitoring, basic reporting or emails, that could be handled in part by AI agents.
Define scope and data sources
Clarify exactly what the AI can do and what it cannot do. Limit its data sources to trusted information, Waxman said. This prevents AI hallucinations or off-topic outputs. Waxman built an agent he called “Future Marketing Bot” to support his class of the same name. Its function was to help answer student questions, particularly when a new semester began. Waxman gave the agent specific instructions on what to reference while answering their questions. “I told it, ‘You must consult the course syllabus. That’s it.’ Then I gave it links to credible sites for additional context.” Be clear and specify which internal databases, websites or resources an AI agent can access.
Break it down by step
Map out every step in a task before trying to automate it. Machines don’t understand subtext so clear instructions are critical, Waxman said. “You have to define every step you do in your job. Then you can see which steps can be automated and how an AI system could perform them in a repeatable, structured way,” he said. Document workflows step by step, noting decision points and dependencies, Waxman said.
Collaborate with IT
Work with your tech team to ensure safe, effective implementation, he said. Misaligned or unauthorized AI tools risk security breaches. Schedule planning sessions with IT before testing AI agents or connecting multiple systems. “Build relationships with IT. Tell them what you’re thinking, how you’ve divided tasks into distinct steps and ask if it’s feasible. Trying to do it on your own, or using shadow AI tools, creates serious risks,” Waxman said.
Develop guardrails
Set rules around ethical use, data safety, privacy and approved tools. Policies ensure that AI supports organizational goals and reputation, Waxman said. Work with cross-functional teams to draft an AI use policy before experimenting with agentic systems. “Communications, marketing, legal, IT, operations, finance — all of these groups need to be involved. Policies give you the foundation to experiment safely and strategically,” he said.
Start small and test
Pilot AI systems in controlled environments before full-scale deployment., Waxman said. This will make it so much easier to identify errors, refine prompts and measure impact. “Test it in a controlled environment, play around with the data and then refine it,” Waxman said. Launch a single AI agent on a limited dataset, then evaluate results and expand gradually, he said.
Include human review
Ensure outputs are checked by a human before action is taken. This can help maintain quality control and prevent mistakes. Waxman described using one AI agent to analyze data, then passing it to a human analyst for final recommendations and testing.
Continuously update
Regularly revisit AI prompts and guardrails. AI effectiveness and safety will inevitably evolve with data, tools and a business’ overall needs, Waxman said. “Every time you build something, you’ll understand better how to structure instructions and prompts logically, step by step,” he said. Schedule routine reviews of AI performance and update instructions as needed.