Rethinking Work With AI: What Stanford’s Groundbreaking Workforce Study Means For Healthcare’s Future
AI systems should reduce cognitive load, clarify ambiguity, and work alongside teams as intelligent collaborators, not as black-box disruptors.
What if the AI systems you’re building solve the wrong problems and alienate the workforce you’re trying to support?
That’s the uncomfortable reality laid bare by a new _research study_ from Stanford University. While AI pilots race ahead across administrative and clinical functions, most are still built on a flawed assumption: that automating tasks equals progress. But for the people doing the work—clinicians, care coordinators, billing specialists—that’s not what they asked for.
The report, _Future of Work with AI Agents_ , offers the most granular audit yet of how worker sentiment, task complexity, and technical feasibility collide in the age of artificial intelligence. Over 1,500 U.S. workers were surveyed across 104 occupations, producing the most detailed dataset yet on where AI could, and should, fit.
For healthcare, the findings could not come at a more urgent moment.
The healthcare sector faces a burned-out workforce, escalating administrative waste, and widespread dissatisfaction with digital tools that were meant to help. A majority of physicians _reported_ documentation burden is a leading cause of burnout, with recent studies showing U.S. physicians spend excessive time on documentation tasks. Nurse attrition has also remained a concern since the pandemic.
Meanwhile, AI adoption is surging, with the vast majority of health systems piloting or planning AI integration. However, there remains a lack of consistent frameworks to align these technologies with real-world clinical and operational dynamics. The result? Misplaced investment, fractured trust, and resistance from the very people AI is meant to assist.
The Stanford study confirms it: the majority of tasks that healthcare workers want automated—like documentation, claims rework, or prior auth form generation—are not where AI tools are being focused. In fact, less than 2% of those high-desire tasks are showing up in actual LLM usage today.
This is more than a technical oversight. It’s a strategic miscalculation.
This study clarifies that the future of AI in healthcare isn’t about replacing human judgment—it’s about protecting it. Leaders must pivot from automation-at-any-cost to augmentation-by-design. That means building AI systems that reduce cognitive load, clarify ambiguity, and work alongside teams as intelligent collaborators, not as black-box disruptors.
And, most critically, it means listening to the workforce before you deploy.
A New Lens on Work: Automation Desire vs. Technical Feasibility
Stanford’s framework introduces two powerful filters for every task: what workers want automated and what AI can do. This produces a four-quadrant map:
- Low Priority : Low desire + Low feasibility
This approach is especially revealing in healthcare, where:
- Judgment-heavy or interpersonal roles (triage, appeals, behavioral health) land in Red Light or R&D zones
Critically, 69% of workers said their top reason for wanting AI was to free up time for higher-value work. Only 12% wanted AI to fully take over a task. The takeaway? Augmentation, not replacement.
Where Healthcare Tasks Fall
**Green Light (Automate Now):**
- Document classification
**R &D Opportunity (Invest in Next-Gen AI):**
- Discharge coordination
**Red Light (Approach with Caution):**
- End-of-life care planning
_Y Combinator_ is one of the world’s most influential startup accelerators, known for launching and funding early-stage technology companies, including many that shape the future of artificial intelligence. Its relevance in this context comes from its outsized role in setting trends and priorities for the tech industry: the types of problems YC-backed startups pursue often signal where talent, investment, and innovation are headed.
The Stanford study highlights a striking disconnect between these startup priorities and actual workforce needs. Specifically, it found that 41% of Y Combinator-backed AI startups are developing solutions for tasks that workers have little interest in automating—referred to as “Red Light Zones” or low-priority areas. This reveals a substantial missed opportunity: if leading accelerators like Y Combinator better aligned their focus with the real needs and preferences of the workforce, AI innovation could deliver far greater value and acceptance in the workplace.
The Human Agency Scale: AI as a Teammate
To move beyond binary thinking (automate vs. don’t), the Stanford research team introduces a more nuanced framework: the _Human Agency Scale (HAS)._ This five-tier model offers a conceptual scaffold for evaluating how AI agents should integrate into human workflows. Rather than asking whether a task should be automated, the HAS asks to what extent the human remains in control, how decision-making is shared, and what level of oversight is required.
The scale ranges from H1 to H5, as follows:
- H5 : Fully autonomous AI — the task is entirely delegated to AI with little or no human intervention.
The Stanford study reveals a clear pattern across occupations: the majority of workers—particularly in healthcare—prefer H2 or H3. Specifically, 45.2% of tasks analyzed across all industries favor an H3 arrangement, in which AI acts as a collaborative peer. In healthcare contexts—where judgment, empathy, and contextual nuance are foundational—H3 is even more critical.
In roles such as care coordination, utilization review, and social work, tasks often require a mix of real-time decision-making, human empathy, and risk stratification. A system built for full automation (H5) in these contexts would not only be resisted—it would likely produce unsafe or ethically problematic outcomes. Instead, what’s required are AI agents that can surface relevant information, adapt to the evolving contours of a task, and remain responsive to human steering.
John Halamka, President of Mayo Clinic Platform, _reinforced this collaborative mindset_ in February 2025: “We have to use AI,” he said, noting that ambient listening tools represent “the thing that will solve many business problems” with relatively low risk. He cited Mayo’s inpatient ambient nursing solutions, which handle “100% of the nursing charting without the nurse having to touch a keyboard,” but was clear that these tools are “all augmenting human behavior and not replacing the human.”
These insights echo a broader workforce trend: automation without agency is unlikely to succeed. Clinical leaders don’t want AI to dictate care pathways or handle nuanced appeals independently. They want AI that reduces friction, illuminates blind spots, and extends their cognitive reach, without erasing professional identity or judgment.
As such, designing for HAS Level 3 (equal partnership) is emerging as the gold standard for intelligent systems in healthcare. This model balances speed and efficiency with explainability and oversight. It also offers a governance and performance evaluation framework that prioritizes human trust.
Building AI for HAS Level 3 requires features that go beyond prediction accuracy. Systems must be architected with:
- Adaptability.
Healthcare doesn’t need one-size-fits-all automation. It requires collaboration at scale, grounded in transparency and guided by human expertise.
These perspectives align perfectly with the Stanford findings: workers don’t fear AI—they fear being sidelined by it. The solution isn’t to slow down AI development. It’s to direct it with clarity, co-design it with the people who rely on it, and evaluate it not just by outputs but also by the experience and empowerment it delivers to human professionals.
1. **Reskill for Collaboration, Not Just Code** The skills rising in value are not just technical. Communication, systems reasoning, and ethical decision-making are essential for teams working alongside AI. Training must evolve accordingly.
2. **Audit Tasks by Workforce Sentiment** Before deploying AI, survey the teams you intend to support. Expect resistance if they ask for relief from documentation but get tools that attempt to guide clinical conversations.
3. **Design for HAS Level 3** Building for H3 (equal partnership) requires agentic AI capabilities: persistent memory, task decomposition, explainability, and real-time adaptability. LLMs alone aren’t enough.
4. **Governance Beyond Compliance** Governance must go beyond regulatory checklists. As agentic AI systems grow more adaptive, oversight must ensure real-time accountability, explainability, and alignment with clinical goals. Dynamic validation, not static rules, will determine long-term trust and safe deployment.
5. **Measure What Matters** The true ROI of AI is trust, relief, and time reclaimed. Outcomes like “claims processed” or “notes generated” aren’t enough. Metrics should track cognitive load reduced, time returned to patient care, and worker trust in AI recommendations.
While throughput remains a necessary benchmark, these human-centered outcomes provide the clearest signal of whether AI improves the healthcare experience. Measurement frameworks must be longitudinal, capturing not just initial productivity but long-term operational resilience, clinician satisfaction, and sustainable value. Only then can we ensure that AI fulfills its promise to elevate both performance and purpose in healthcare.
Dr. Rohit Chandra, Chief Digital Officer at Cleveland Clinic, _gave voice_ in June 2025: “It’s made their jobs a ton easier. Patient interactions are a lot better because now patients actually engage with the doctor,” he said, referring to 4,000 physicians now using AI scribes. “I’m hoping that we can keep building on the success that we’ve had so far to literally drive the documentation burden to zero.”
## Build With, Not For This moment is too important for misalignment. The Stanford study offers a blueprint.
For healthcare leaders, the message is clear: If you want AI to scale, build with the workforce in mind. Prioritize the Green Light Zones. Invest in agentic systems that enhance, not override. Govern AI like a trusted partner, not a productivity engine.
The future of AI in healthcare won’t be determined by the size of your model. It will be defined by the quality of your teaming.