July 8, 2026 at 12:00 PM ET
How research administration teams move from breakdowns to breakthroughs, with Matt and Gigi Smith of Smith-Atlas.
Please fill out this form to access the recorded webinar.
Most research offices already know AI is coming for their workflows. The harder question is why so many AI rollouts stall, and what to do differently. In this session, Matt Smith and Gigi Smith of Smith Atlas came at that question from two directions: the human side of change, and the practical tools a research office can build for itself. Gigi drew on rhetoric and emotional intelligence to explain why adoption breaks down. Matt gave a live demo of an AI assistant that builds other AI assistants. Together they made the case that making AI work is less about the technology than about the people using it.
Industry analyses put overall AI project failure rates well above 80%. Of the projects that fail, only about 23% trace back to the technology itself. The rest come down to governance, change management, and a striking gap in planning: 73% of failed projects never had an agreed definition of success before they started. In research administration specifically, the stakes are sharper. Teams are protecting confidential data and intellectual property while using public AI tools. The work is compliance-critical, so a hallucinated citation or a misread regulation carries real consequences. That is why human review has to stay non-negotiable. Gigi highlights: before you can solve any of those technical problems, you have to solve an engagement problem. If your people aren't with you, the best tool in the world won't land.
Gigi walked through five of the most common barriers to AI engagement, and noted that three of them share a single through line: emotions.
Even for those who consider AI a threat, the reality is it's already within the fence. There's no such thing as keeping it out."
The takeaway isn't that resistance is irrational. It's that the resentment is often aimed at the wrong target. The cause of most problems isn't the technology. It's the gap in supporting the humans who have to use it. The world launched a world-changing tool without getting everyone ready for it, and it was humans who did the launching. Leaders owe their teams clarity on where the real problem sits.
One of the clearest ideas of the session was the difference between AI innovation and AI literacy. Innovation gets the attention, but literacy is the keystone that everything else builds on. Gigi compared it to learning to read. You are allowed to dislike reading, but no one would argue you shouldn't have been taught. AI is the same. Not everyone needs to be an innovator or a fan, but everyone needs to evolve their core competencies, especially at work. When that clicks, people move from "Why should I?" to "How can I not? I don't want to fall behind." The practical warning: employees are being pushed to work with AI without a solid understanding of what it is and, just as importantly, what it is not. Literacy opens the door not just to access, but to the comprehension that quiets fear.
Gigi's recommendations for actually leading the change:
Matt took the second half to the practical level: instead of large, expensive SaaS platforms, he focused on the lightweight tools a research office can spin up for its own everyday work, solving a specific problem quickly inside your own environment. His demo was a meta tool: an AI assistant whose job is to build other AI assistants. Rather than writing a system prompt from a blank page, the tool runs a structured interview, which Matt likened to the old "21 questions" game. It walks you from a broad starting point down to specifics: the assistant's role and persona, its audience and their skill level, the tone, the background knowledge, the guardrails, and the desired output. The result is a clear, ready-to-test system prompt you can paste into whatever platform you use, whether that's Copilot, Gemini, or Claude. A pro tip he shared: once one model has drafted a system prompt, hand it to a different model and ask what could be improved or what best practices are missing. Feed those suggestions back into the original, then make the final call yourself. He also builds his assistants to be multilingual, so principal investigators and staff who are English-language learners can work in whatever language they prefer.
Matt then showed what that meta tool produced: a Policy Evaluation and Drafting Assistant, built to handle a task every research office knows well, drafting a new university policy. Here's the flow he demonstrated:
"How do you protect sensitive data when using AI?"
Matt's answer: use sanctioned tools, and work with your Chief Information Security Officer to map which tools are approved for which use cases. At Boise State that lives in published data-use guidelines, backed by a homegrown, heavily logged tool on Amazon Bedrock and by Gemini for Workspace. Gigi added that threats will always outrun our defenses, which is why knowing what not to do is a core pillar of AI literacy.
"Should we bring in an outside consultant or train internally?"
Both landed on a hybrid. The learning has to stay loyal to your environment, so internal people who know your real use cases are essential. Start with any AI committee or task force you already have; if that infrastructure doesn't exist, bring in an external partner to work alongside your leadership.
"How did you build the assistant that builds assistants?"
No template. Matt built it iteratively, after making many single-purpose tools himself, by encoding the process he already knew he repeated each time. He version-controls his assistants because they keep improving as new ideas come.
Matt and Gigi closed by acknowledging how much ground they covered, from a very granular tool walkthrough to high-level, existential change. Their intention wasn't to overwhelm, but to offer fundamentals for finding your own footing. They left attendees with a resource kit built specifically for this Atom Grants webinar, including The Research Administrator's Mixtape: A Guide to Durable AI Implementation, a reference guide that turns the session's imperatives and takeaways into something you can use to move your team from breakdown to breakthrough. The kit also includes an open invitation to send in any question you didn't get to ask, and an option to go deeper on anything covered. The through line of the hour: making AI work isn't only about the tech. It's about thoughtfully merging it into the workflow, and giving people the knowledge and support to operationalize it. Prioritize your people as people, and the tools follow.
Matt Smith is a co-founder of Smith Atlas. He brings roughly 18 years in research administration across three vantage points: the U.S. government, private practice as an attorney and outside counsel, and higher education, where he has spent the last 12 years at Boise State University. He began working with AI on research contracting before COVID and now focuses on building practical AI assistants for everyday research-administration work.
Gigi Smith is a co-founder of Smith Atlas and the Founding Director of AI and Literacy Services at the College of Western Idaho. She is an educator across audiences ranging from college students to business professionals, on topics including AI literacy, research writing, communications, leadership, and systems thinking. Her background in literature and rhetoric shapes the lens she brings to change management and the emotional intelligence behind successful AI adoption.