To Adopt AI at Scale, Employees Need to Trust Agents

Imagine this scenario: You’re a manager testing out a new AI agent “assistant” for your team, and as you set it up it asks for permissions. Three options appear:

“Access, create, and delete all files.” That feels uncomfortable.

“Read and send emails.” Sounds risky. What if it sends something weird to our clients?

“View and make payments.” I don’t trust this. These are some of the most critical aspects of our work. Permissions denied.

You try to run the AI agent anyway, but it constantly gets stuck. You feel as though you’re struggling with an early AI chatbot. The agent designed to take work off your plate is instead wasting your time. You close it, more disillusioned than ever with AI agents. I could never make my team use this, you think.

Some version of that scenario is playing out for thousands of people every day, and the cost of distrust shows up clearly in the numbers. Fifty-nine percent of enterprise organizations say they are using agentic AI, but only 9% have successfully turned that into autonomous workflows—the very purpose of AI agents. Only 27% of middle managers see real ROI from their AI deployments—a shocking waste of investment when you consider that, on average, companies will be investing $202 million in AI during the next 12 months.

Zoom in and the role that trust plays becomes clear. A study with over 3,000 people on the adoption of an AI financial adviser found that most concerns were trust-related. Technological performance came second. For example, privacy weighed 31% on the decision to adopt the adviser, 23% was the fear that it might take actions they wouldn’t want it to, and 11% was the worry that they wouldn’t understand the AI’s actions.

To understand how to overcome this trust problem, we performed an in-depth review of the latest scientific evidence on human-AI interactions. Then, we asked leading AI researchers at Wharton and business leaders at the largest organizations that are providing and deploying AI agents (among them Google, ServiceNow, and Zapier) how they are facing and solving this issue.

We recently published the combined findings in the Wharton Blueprint for AI Agent Adoption, where we identified practical techniques that help managers improve AI agent adoption in their organization.

Here are five of the most effective ones:

Be upfront about the agent’s limitations. Counterintuitively, telling people where an AI is likely to fail made them perceive the agent as up to 14.6% more transparent, which directly increased trust. Across two experiments, people who were clearly informed about an AI’s known weaknesses trusted it more and worked with it up to 7.2% more effectively than those who weren’t told, even though the AI itself was identical.

When you give your team a new tool, state clearly where your agent is likely to struggle and when people should double-check its output. Think of the agent as a new colleague who’s open about which parts of the job they’re still learning. That signals self-awareness, not weakness.

Most AI tools don’t make this easy. For example, Google’s Gemini or Anthropic’s Claude offer only a blanket disclaimer about possible errors, not the specific tasks where errors are most likely to occur. That means that, in most cases, you need to build your own internal guidelines for your team.

Make sure the agent feels competent and not overly friendly. You’d struggle to trust a new lawyer who showed up cracking jokes in flip-flops, or who was overly agreeable to even your most dubious requests, no matter how sharp they actually were. The same logic applies to AI agents. Across multiple experiments, people were less willing to use an AI that came across as friendly and warm than one that came across as competent.

This research finding is consistent with what Chris Caldwell, the CEO of Concentrix, a Fortune 500 company that deploys AI-integrated business processes in large organizations, told us he and his team see. “Customers tend to get frustrated with overly polite and obedient technology that isn’t accomplishing things at speed,” he said.

To help your agent project competence, make sure it’s designed to explain what it did in carrying out its task (e.g., “I used these three criteria”) and why it made the choices it did (e.g., “This task is more urgent, so I prioritized it”).

Have it show that it understands employees’ long-term goals. People accepted an AI’s recommendations 54% more often when the AI demonstrated a clear understanding of their broader goal (e.g., a budget that supports next year’s headcount needs). This is one of the most powerful insights in our findings, but one of the most underused.

Cleo, a personal-finance assistant, is one of the few AI systems that does this very well. Rather than surfacing isolated recommendations, it ties every suggestion to a daily goal and long-term financial roadmap. Apply the same principle by having your AI agents explicitly connect their actions to the user’s stated goals (e.g., “I’m reconciling this with the HR budget because it may clash with the headcount plan you outlined”) rather than delivering a recommendation without context.

Frame it as a helper, not a powerful entity. Teams prefer to work with AI agents that feel subordinate, not equal or—worse—more powerful. Across multiple experiments, people initially perceived 13.8% lower privacy risks with an AI agent compared to a person doing the same task. That changed completely when they were reminded the agent had real power (e.g., making decisions on their behalf). At that point, they started to see the agent as riskier than the person.

To avoid this, make sure your team knows your agent is just a tool. Frame your agent as “assisting” or “supporting,” not “deciding” or “evaluating,” especially around sensitive information. Take inspiration from Microsoft, which used this idea when it came up with the name Copilot. Skip language like “AI can take this over for you.”

Make clear that people are still in control when it matters. Agents should have a moderate level of autonomy, not minimal or total. We found multiple studies showing that the level of control is fundamental to successful adoption.

ServiceNow does this well by designing into AI agent workflows a “control tower” to set guardrails in advance, monitor activity, and even override decisions when necessary. That kind of implementation addresses the fear of losing control and of making irreversible mistakes, while still letting the agent handle routine tasks on its own.

Adam Seligman, the CTO of Workato, shared with us how he and his team increase trust in their AI agents when deploying them to enterprise customers. “Companies,” he told us, “need to be able to say, ‘You can draft this email, but you can’t send it,’ or ‘You can recommend inventory moves, but you can’t execute them.’ Without clear controls, agents stay stuck on trivial tasks, because nobody trusts them with anything important.” Set your agents to ask for confirmation before any consequential action.

. . .

The goal in all of this is not to make your team trust AI agents blindly. They shouldn’t. Scientific research shows that trust is a critical condition for adoption and must be earned and appropriately calibrated. AI is prone to mistakes, and people need to stay in the driver’s seat for decisions that matter. But without enough trust, employees will treat AI agents as glorified chatbots, or actively avoid them, and organizations will keep struggling to see ROI from their agentic AI investments.

Copyright 2026 Harvard Business School Publishing Corporation. Distributed by The New York Times Syndicate.

Explore AAPL Membership benefits.