Two women sit at a table looking at a laptop together Empower Work UX Lead Jessy Baer (left) and a peer counselor trainer collaborate on the design of its AI assistant.(Courtesy of Empower Work)

Nonprofits are racing to adopt AI, with a recent report showing 92 percent already using the technology to support their work. Many early adopters have focused on direct-to-user applications like AI tutors that help students learn, AI coaches that offer guidance at scale, and AI navigators that help people find benefits or services. These tools hold real promise, especially in a sector where demand often exceeds capacity. But for organizations whose work depends on empathy and human connection, automation can also feel at odds with care. The Center for Effective Philanthropy has noted that social change leaders are increasingly concerned about AI’s risks, including the loss of human connection when it matters most.

Empower Work feels that tension directly. While we use technology to scale, our core service, which offers free, confidential, text-based coaching for people navigating tough work challenges, relies on deep human connectivity. When someone texts into the line, they connect with a trained peer counselor ready to provide personalized support for difficult situations such as underemployment, poor management, and job loss.

Human relationship isn’t just at the center of our theory of change; it’s something our help seekers actively seek out. They frequently ask our peer counselors, “Are you human or AI?” because they want reassurance that there’s an actual person on the other end of the line. At the same time, demand for worker support is growing. Empower Work faced a challenging question: How could we maintain the human connection and empathy that make these conversations impactful, while incorporating AI in a way that helped us serve more people?

In this article, we outline some of the choices Empower Work made when integrating AI and what we learned, so other teams seeking to scale their people-centered services can adapt and integrate AI without sacrificing human empathy.

Start With a Problem Statement

Empower Work’s text-based coaching model supports help seekers (typically lower-wage workers and those from historically marginalized communities) gain clarity, confidence, and agency. Someone might text us after workplace bullying has affected their sense of safety or when they’ve lost confidence after a monthslong job search or when they know they need to leave a job but feel stuck by burnout and uncertainty. In a session with Empower Work, they receive emotional support from a trained peer counselor and leave with clear, immediate next steps.

With rising economic and political uncertainty, the demand for our text line was growing. To ensure our peer counselors could meet this increased need, we asked, “Where do we lose time or consistency that reduces the quality of human support we provide?”

This question is familiar to people working across service-based nonprofits. Staff need to respond quickly and consistently, but the work also requires judgment, trust, and care that cannot simply be automated. Through user experience research and workflow analysis, Empower Work found key constraints for our peer counselors:

  • First were manual, time-intensive but repetitive tasks: writing end-of-session summaries, catching up on prior messages from help seekers during handoffs, and finding resources in a complex template library.
  • Second was their cognitive load while having conversations with multiple help seekers at once. Uncertainty about how to best support the help seeker and respond in a timely manner increased feelings of stress.

The North Star became clear: preserve human empathy and judgment in interactions with help seekers and use AI to support our counselors behind the scenes so they can spend even more time deepening the connection that is at the heart of their job and our service. The choice reflected both what our help seekers directly told us they wanted and growing research on AI’s empathy gap showing AI can only perform one of the three components of empathy—it can understand another’s state, but it can’t actively feel what another is feeling or invest in them. Because of that, human responses in emotional contexts are often more likely to build trust, create positive immediate emotional impact, and provide the motivation to help people take action.

Building Trust

The next challenge was implementation: how to introduce AI in ways that strengthened interactions with help seekers rather than diluted them. We built an AI assistant plug-in that showed up in the peer counselor workflow environment where they were talking to help seekers. It could read the live text conversation of the counseling session and offer support in three areas: ideas for suggested next responses, relevant resources from Empower Work’s vetted library, and summaries for handoffs or end-of-session documentation. Building the assistant wasn’t the end, but the beginning of a learning process for integrating AI into our work. The steps below walk through our process and how service-centered organizations can test AI carefully, learn quickly, and set guardrails before expanding further.

Start with low-risk, high-friction work

As we explained above, a useful place to begin is with tasks that consume staff time but do not require AI to make consequential decisions. Across nonprofits, this often includes retrieving context, surfacing vetted information, and drafting internal documentation. When using Empower Work’s AI assistant, counselors always reviewed, edited, and decided what to share before any messages were sent. Starting with this kind of human-in-the-loop support allows organizations to reduce friction without handing off sensitive judgment to the model.

Embed AI into staff training at the right time

Training is where organizations can build comfort, reduce confusion, and make clear that AI is there to support human judgment rather than replace it. But timing matters. At Empower Work, we initially waited to introduce the AI assistant in the final week of peer counselor training through a homework reading. Counselors were told the tool existed, but they did not practice using it during live training sessions. When AI assistant adoption numbers were lower than expected, follow-up surveys revealed that the issue was not lack of interest, but rather that many counselors were simply not aware of the assistant, did not remember it was available during live shifts, or were unsure how to use it in the moment.

Based on this awareness gap, we decided to move the AI introduction from the end of training to the middle of training, after counselors had built a foundation in Empower Work’s core counseling skills and added live practice using the assistant with supervisors. After making this change, we saw a significant increase in usage and positive feedback from counselors.

Build feedback loops into the technology

Staff and volunteers closest to the work need simple ways to flag when an AI feature is helpful, off base, or creating friction, and teams need a clear process for turning that feedback into improvements. At Empower Work, that meant creating a volunteer AI advisory council of peer counselors who could respond quickly to product changes and help test prompt variations. The team also built feedback directly into the assistant through five-star ratings and open-response comments on outputs. Each week, the UX lead reviewed that feedback, identified recurring themes, translated them into prompt refinements, and tested those changes again with the advisory council.

Because the advisory council could directly see how their feedback shaped the assistant, they developed a sense of ownership over the tool rather than feeling that new technology was being imposed on them. We also did not make the assistant mandatory. Counselors were invited to try it at their own pace, with light nudges and social proof showing examples of how other counselors were benefiting from it.

Scale with values-aligned partnerships

The strongest AI partnerships do more than add funding or technical skills; they reinforce the values of the core service itself. For nonprofits, that can mean saying no to attractive AI use cases that are more appealing to funders than they are helpful to constituents. At Empower Work, some donors were most interested in direct-to-help-seeker AI counseling, such as a chatbot that could respond to workers without a human counselor involved. That kind of tool might have produced higher immediate projections for “total reach” numbers in grant proposals, but we knew the community wanted human support in high-emotion moments. PagerDuty, one of our philanthropic and technology partners, proved to be the right collaborator because it aligned with our philosophy of starting with internal AI tools behind the scenes that support scale.

Design Lessons

In high-trust services, small design choices can have outsized effects on trust, usability, and adoption. Imagine a counselor is having a conversation with a help seeker who feels safe and understood. The counselor wants help thinking through the next response, so they write a prompt to AI such as, “Should my response focus more on emotional support or practical next steps?" If that prompt were accidentally sent to the help seeker, the trust and rapport built in the conversation could be damaged. Or imagine a newer counselor who is still building confidence. If the AI suggests a single definitive-sounding response option, they may assume that that response is the answer and copy and paste it without fully considering whether it fits the moment or their own voice/instinct. Risks like these helped guide our design priorities and safeguard strategies:

1. In fast-paced settings, more flexibility is not always better.

Nonprofits may assume that giving staff an open-ended AI interface will create flexibility, but in live service environments it can also increase cognitive burden and introduce avoidable risks. In Empower Work’s early prototypes, counselors had both an open-response chat box and prompt buttons for next response guidance, relevant resources, and summary. In practice, the open field proved less helpful: It slowed replies, increased mental load, and created the possibility that an AI prompt could accidentally be sent to a help seeker. Removing it made the tool simpler, faster, and safer to use.

2. AI should support judgment, not replace it.

In human-centered services, the goal is rarely to produce a single, best answer. More often, staff need support that sharpens their discernment. At Empower Work, that meant offering counselors a small set of suggested next responses rather than one definitive recommendation. Over time, the team found that a limited range of plausible options preserved human choice while avoiding the overload that comes with too many possibilities.

We also collaborated with our top AI assistant peer counselor users to create best practices for using the assistant thoughtfully. We gave counselors sample scenarios, AI-generated responses, and asked them to rewrite the responses as they would actually send them and then organized these into principles with example scenarios showing the side-by-side AI response before and after a counselor made changes.

One thing that stood out was how useful the AI could be when the counselor needed help finding the next general direction in a complicated conversation, but it also highlighted how necessary it was for the counselor to adjust the response in their own voice. In one example, a help seeker was trying to decide whether to stay in a difficult job for five more years to keep retirement medical benefits, or risk moving to another job with the potential to be just as bad. The assistant offered a focusing question, but our counselor did the human part—slowing it down, making it collaborative, and writing it in her own tone and voice.

3. Model choice is an ongoing design decision.

In social sector settings, performance must be judged in context. A slightly stronger model on paper may be the worse choice if it adds latency to real-time interactions. When we initially launched the AI assistant the larger models produced modest gains in output quality but introduced delays of up to a minute per prompt, which was too costly in a service where counselors aim to respond to people quickly.

Six months later, we revisited the comparison and found that newer models had significantly improved the quality of more emotionally complex outputs, especially next response guidance, without introducing the same level of latency. We shared side-by-side examples with counselors, and they strongly preferred the newer model even if it meant a bit more latency, so we updated.

The lesson is that model selection is not a one-time decision. As models, costs, latency, and use cases change, organizations need to regularly re-examine the tradeoffs and choose the model that best balances quality, speed, and cost within the realities of live support.

Time Back Where It Matters

After six months, roughly 65 percent of Empower Work counselors used the assistant at least once per shift, and a year later, usage had grown to nearly all active counselors. The assistant helped counselors manage more simultaneous conversations, from about two to nearly three, and counselors saved time and responded to help seekers faster: The average time to share resources dropped by 41 percent, and the average time to write session summaries dropped by 60 percent. And throughout, we maintained trust with our help-seeker community by ensuring that a human counselor was always present in moments of need, supported by AI but still using their own judgment, empathy, and voice.

When someone is anxious about losing their job or facing a hard conversation with a manager, they do not want an AI chatbot; they want a person with shared experience who can listen and feel with them. That belief shaped our guiding principle when integrating AI: AI does not counsel help seekers directly in highly emotional moments. Instead, it supports the people providing care so they can stay more present with the people seeking help. The lesson for peers and philanthropy is simple: For deeply human, empathy-based services, use AI to strengthen human support rather than replace it. That is where trust is built, and where deeper impact happens.

Read more stories by Jaime-Alexis Fowler & Taymar Quezada.