Why Domain Experts Are Quietly Dropping Out of AI Innovation—And How to Keep Them

How Two Organizations Took Divergent Paths with Generative AI

Two U.S.-based organizations—a healthcare provider and a legal services firm—set out in early 2023 to harness generative AI for organization-wide solutions. At NE Health, doctors, nurses, and other domain experts helped develop tools like patient-friendly discharge summaries, and within two years the organization had built and deployed 141 such solutions, with a growing network of volunteers. At LegalCo, however, the effort to build a gen AI-powered legal research tool quietly collapsed: over 80% of the domain experts who had initially signed up gradually dropped out, and the company was left with only three operational solutions.

The difference, according to a two-year field study published by Harvard Business Review, was not in the quality of the people or the sophistication of the technology, but in the presence—or absence—of what the researchers call "scaffolding."

Scaffolding is the set of structures that make it feasible for busy domain experts to do the hidden, collective work required to move gen AI from a sandbox experiment to a live, value-creating tool. Without it, the study found, domain experts do not resist or lobby against AI; they simply run out of time and energy, and they stop participating.

The Hidden Work and the Scaffolds That Made the Difference

The Hidden Work of Collective Experimentation

While many leaders assume that rolling out an AI sandbox and a few training sessions is enough, the study uncovered three intensive modes of collective experimentation that domain experts must perform week after week. First, collective trial-and-error—painstakingly testing prompts, documenting what works, and troubleshooting erratic outputs with technical support. Second, collective review and revision—justifying approaches to colleagues with different priorities, debating metrics, and repeatedly reworking outputs. Third, collective alignment and integration—continuously re-evaluating use cases as models improve, swapping in new models, and adapting solutions to live workflows. This hidden workload intensifies an already demanding job and, left unsupported, leads to quiet withdrawal.

The Four Scaffolds That NE Health Built

NE Health succeeded because it erected four types of scaffolding that directly mitigated that hidden work. The first, trial-and-error scaffolding, included informal training events like "promptathons," a clear documentation method that fit existing routines, and a project triaging system that assigned a dedicated technical expert to high-priority projects—so that domain experts were not left alone with model hallucinations. LegalCo offered initial training but no ongoing support, unclear documentation, and only intermittent access to technical help.

The second, review and revision scaffolding, involved a shared evaluation rubric—with explicit weights for accuracy, completeness, readability, and patient-centeredness—that allowed cross-functional teams to align on what counted as a good solution. NE Health also set up knowledge forums where domain experts learned emerging best practices. LegalCo had no shared rubric and no knowledge-sharing forums, so teams wrestled with competing definitions of quality without resolution.

The third, alignment and integration scaffolding, gave domain experts a clear risk screen—technical, operational, compliance, and ROI risk—against which to evaluate potential use cases, and IT support to harden prototypes and integrate them into the electronic medical record system. At LegalCo, no such risk screen existed, and there was no integration support; domain experts built manual workarounds that consumed enormous time.

The fourth, and most systemic, was roles and rewards scaffolding. NE Health formally recognized gen AI innovation work in job descriptions and performance reviews, with material rewards—publications, promotions, bonuses—following engagement. At LegalCo, the work was invisible at review time, as one domain expert stated: "My annual review is still based on the same criteria as before gen AI existed."

What Leaders Must Build to Keep Domain Experts Engaged

For executives seeking to replicate NE Health’s results, the study points to concrete, evidence-based actions—all tied to specific scaffolding deficits observed at LegalCo.

• Provide ongoing, informal upskilling. NE Health’s "promptathons" gave domain experts a low-risk way to build skills continuously. Avoid one-off training and instead create recurring, peer-led sessions.
• Create a project triage process. Assign dedicated technical experts to high-priority projects so that domain experts are not expected to debug AI quirks alone, as LegalCo learned the hard way when intermittent support failed.
• Build a cross-functional evaluation rubric. NE Health’s shared scoring framework—weighting factors like accuracy, completeness, and patient-centeredness—prevented endless debate. Without it, review meetings generate waves of rework and disengagement.
• Give domain experts a clear risk screen. NE Health evaluated every potential AI use case against technical, operational, compliance, and ROI risk, channeling effort to viable projects. At LegalCo, the absence of such a screen led to wasted effort on solutions that did not meet management priorities.
• Integrate AI innovation work into HR systems. The single most powerful signal from the field study is that domain experts will stop experimenting if that work is invisible in performance reviews and uncoupled from promotions and raises. Formalize the role and reward it, or the initiative will fade.