Meeting clinicians where they are through signal-guided AI innovation Authors: Benjamin Rader, Ryan D Nagy, Heather L Nelson, Peter C Laussen, John S Brownstein Published: 2026 The recent launches of ChatGPT Health and Claude for Healthcare represent exciting offerings, but they do not mark the introduction of large language model (LLM) chatbots in clinical settings. Clinicians, like patients, have already started incorporating these tools into daily cognitive workflows, often outside formal institutional channels.1 2 This is not a governance issue alone; the tools clinicians independently adopt can reveal where existing systems are failing to deliver adequate clinical workflow support. We argue for signal-guided innovation: treating organic clinician behaviour as a leading indicator of where artificial intelligence (AI) can have impact. In our experience, a core challenge for hospital innovation teams is not the sourcing of quality models or understanding conceptual needs (eg, reducing documentation time); it is knowing which tools truly alleviate burden.3 The distinction between efficacy and effectiveness appreciated in clinical trial frameworks needs to be incorporated in AI evaluation.4 A model trained in ideal conditions may demonstrate a high area-under-the-curve yet still fail to integrate into existing workflows. In a PubMed search of 455 articles describing ‘artificial intelligence’ and ‘clinical evaluation,’ only two referenced ‘intention-to-treat’ principles. Evidence suggests AI tools succeed based on usability, trust and institutional readiness, with algorithmic performance often a secondary consideration.5 6 Clinicians are trying to optimise cognitive performance, adopting tools that protect limited bandwidth amid difficult workloads. As domain experts, they are uniquely positioned to be both test subjects and outcome evaluators of clinical AI tools—yet they often lack the time and institutional mandate to do so. In parallel, administrators face an increasing volume of AI vendor pitches that are difficult to validate or scale. At Boston Children’s Hospital (BCH), this experience has guided our approach to clinical innovation broadly and AI integration specifically. We invest in both bottom-up staff enablement and in top-down tool creation. At BCH, innovation often begins with someone stating, “I’m already using this tool, can we make it work here?” This signal suggests to us that innovation teams can be impactful as implementation enablers, assisting with governance and capacity building alongside sourcing novel tools. Consider LLM-driven evidence synthesis platforms. Clinicians are drawn to tools that rapidly summarise complex evidence, and the broad adoption of platforms like OpenEvidence supports this. The demand exists whether or not health systems formally sanction it, and prohibition is rarely productive. With OpenEvidence, our organisation signed a Health Insurance Portability and Accountability Act (HIPAA) Business Associate Agreement to support patient data privacy while providing clinicians clear use boundaries that preserve utility. This approach enabled rapid institutional adoption while preserving clinician trust, transforming a publicly available tool into a governed resource. The same logic applies to general-purpose LLMs. An internal BCH survey found that over one third of all providers were using ChatGPT for non-clinical work-related tasks.1 This adoption occurred organically because these tools reduce friction in ways that matter to clinicians. Rather than attempt to curtail use that clinicians have already identified as helpful, we elected to formalise access by becoming launch partners of ChatGPT Health, OpenAI’s HIPAA-compliant enterprise offering. This approach was designed to make it easy for clinicians to use the tools they were already using, while ensuring appropriate data governance. Our goal is not only to introduce new technology, but to bring existing behaviour into an environment that protects patient data, offers clinicians trustworthy tools and meets institutional security standards. Clinicians, especially paediatricians, often have needs too niche or unprofitable to attract vendor attention. This is heightened at institutions like BCH, where many patients have rare diseases or high medical complexity. When staff signal needs that the market has ignored, innovation teams must construct solutions. At BCH, multiple departments independently requested custom AI tools tailored to their workflows. LLMs increasingly make building these applications feasible without requiring extensive middleware. In response, we developed InternalGPT, a Health Insurance Portability and Accountability Act (HIPAA)-compliant tool where hospital employees can leverage LLM capabilities via a graphical user interface and application programming interface (API) with appropriate guardrails (eg, behind our institution’s firewall, auditability). Rather than delivering AI as a finished product, InternalGPT offers infrastructure for converting medical specialty-agnostic and context-agnostic applications into deployable prototypes. To date, over 4000 BCH staff members have signed up for access. Like any attempt to implement an institutional framework, operationalising signal-guided innovation is not risk-free. Formalising these tools implies institutional endorsement, which clinicians may misinterpret as comprehensive validation rather than facilitated access, or as diminishing the need for critical appraisal of AI-generated outputs. But large language model (LLM) chatbots hallucinate, reflect training biases and can produce outputs that, if accepted without scrutiny, affect clinical decisions. These problems are amplified in paediatrics, where training data are sparse and adult-derived models may not generalise to paediatric physiology.7 8 At BCH, we mitigate these risks through a multilayered governance process with representation from scientific review, information security, privacy and clinical leadership, alongside case-by-case validation efforts. These safeguards, however, are not a substitute for clinician education on appropriate use boundaries and limitations of each tool. We also recognise that any comprehensive approach to implementing innovation depends on infrastructure (eg, dedicated innovation teams, legal capacity) that many community hospitals and safety-net organisations lack. However, signal-guided innovation may itself lower this barrier by prioritising tools clinicians have already adopted, reducing upfront discovery costs. Importantly, though, any approach to AI innovation must integrate patient perspectives, as emerging evidence suggests patients expect transparency around AI use in their care and may not yet trust health systems to deploy it responsibly.9 10 The utility of clinical AI will not solely be determined by the latest well-funded startup. It will be shaped by whether health systems can create environments where innovation is both ambitious and usable. Clinicians are necessary partners for this, not just passive recipients. Governance that discourages rather than channels organic adoption misses valuable signals about where AI efforts should be directed. At BCH, we have created a partnership of information technology, informatics, clinical leaders and an innovation team to navigate this tension and ensure alignment between operational requirements and institutional reality. We expect similar dynamics exist across other health systems and encourage others to adopt signal-guided approaches suited to local institutional governance and contexts.