Outcomes Articles

Can You Prove Your AI Gives Clinicians Time Back?

October 14, 2026

About This Report

Featuring insights from Mary Varghese Presti, Richa Gupta, and Jeroen Tas hosted by our CEO, Saul Marquez.

While a nurse records fluids, vital signs, and observations, a patient may need attention. Mary Varghese Presti described that tension after her team spoke with nursing executives, managers, and bedside nurses about the work filling a shift. Documentation had to get done. So did patient care.

If you're presenting clinical AI to a hospital buyer, a faster documentation task is a useful result. It leaves an unanswered question: did clinicians gain usable time once review, corrections, and coordination were counted? That is the question an expansion or renewal should help answer.

Before making a workload claim, agree with the hospital on whose work you will measure, what they did before deployment, and what they still do afterward. Those distinctions matter when saved minutes begin shaping expectations about staffing, retention, or financial return. A task-speed result cannot support all of those promises on its own.

Start With the Work Clinicians Actually Do

A physician documenting a patient visit and a nurse recording observations throughout a shift both work in an electronic health record, but they do different jobs. Mary made that distinction when discussing Microsoft's Dragon Copilot nursing work. Physicians generally create narrative notes covering an assessment and treatment plan. Nurses record structured information such as fluid intake and output, vital signs, and pain scores. A solution designed for a physician's note won't necessarily fit nursing documentation.

The difference extends beyond the software. Mary described nurses noticing several things as they move through a patient's room, including whether a meal was eaten and whether a patient has gotten out of bed. They don't necessarily speak every observation aloud. Capturing those details with AI can require nurses to verbalize what they're seeing and then verify that the information lands in the correct fields. That changes their work.

That makes the intended outcome more specific. Is the goal to reduce manual entry, make observations available sooner, or prevent a second round of documentation after patient care? Each deserves a baseline and a way to examine the resulting work. Nurses may need to check generated entries or adjust how they communicate in the room. Measure those requirements.

Mary also described a longer-term vision for AI agents handling predictable coordination work. Nurses may arrange meals for new admissions, communicate allergy precautions, or coordinate with ancillary departments when something is needed. These jobs still take time. Her examples identify potential work to examine, rather than proving that autonomous agents have already taken it over.

For a hospital buyer, the first practical move is to name the targeted tasks, identify who currently performs them, and define what completion requires. Measure the nursing workflow on its own terms. Once the work is defined, the next question is whether clinicians actually use the technology and experience the intended improvement.

Measure Use, Burden, and Financial Return Separately

Richa Gupta, Vice President of Care Transformation at Ambience Healthcare in her May 2026 conversation, described the physician encounter as a sequence of connected responsibilities. Before a patient arrives, the physician reviews the chart. During the visit, the encounter must be documented. Afterward, appropriate diagnoses and services need to be captured for coding and billing. Reducing work at one stage doesn't automatically show what happened across the entire encounter.

Richa begins her value discussion with utilization. She reported that Ambience was used in 80% of visits at participating client organizations. That figure concerns visits, not clinician adoption. Its population and period require verification. Still, the underlying buying question is sensible: a tool that isn't consistently used has limited opportunity to change the work it's intended to improve.

The buyer needs to see the denominator behind that result. Were all eligible encounters counted, or only visits involving an enrolled group? Did clinicians use the product across specialties and over time? Then evaluate what they did before and after use, including chart preparation, generated-note review, documentation completion, and work done outside scheduled hours. Actual use and measured burden belong in different columns of the evaluation.

Richa also described a reported $13,000 per physician per year ROI associated with a St. Luke's study and CFO sign-off. The calculation and costs aren't established in the interview. Some financial benefit may come from appropriately recognizing services clinicians already delivered. That can matter to a hospital, yet it doesn't establish that physicians have less work.

Coding deserves its own test as well. Richa described external audits of coding accuracy, but the methods and reported performance were not independently established in these interviews. A coding audit should examine whether the documentation supports the services actually provided. It shouldn't be treated as a proxy for reduced documentation burden, staff retention, or clinical safety.

Keep utilization, workload, coding, and financial evidence separate. Even strong physician-level results leave another question unresolved: what happened to the surrounding care team and patient experience?

Look Beyond the Individual Task

Jeroen Tas raised that issue from a broader perspective in his 2020 Outcomes Rocket conversation about innovation at Philips. He described evaluating technology through four dimensions: clinical outcomes, patient experience, care efficiency and cost, and the experience of healthcare professionals. He also emphasized stating the intended value, collecting evidence to support it, and continuing to measure after a product is installed.

A physician may finish documentation sooner without anyone having measured whether a nurse, medical assistant, or another team member has taken on additional checking. A system may make more information available without showing whether clinicians can find the relevant details when they need them. These are possibilities to test, not outcomes established by the three interviews. Task timing doesn't answer those questions.

Jeroen explained that medical technology operates within information flows, clinical work practices, and sometimes reimbursement requirements. His discussion of multidisciplinary remote care illustrated how a useful technology can depend on a team organized to respond to its information. That account was historical and wasn't a study of the current nursing or physician AI tools. Its criteria remain useful for evaluation.

Mary's private-preview experience offers a concrete illustration of the distinction between a promising observation and a demonstrated outcome. She reported that information captured across nursing shifts helped a neurologist recognize a pattern and take action. That's worth examining. Yet it doesn't by itself prove improved patient outcomes, nor does recording more context establish that nurses spent less time working. Accuracy, review effort, and clinical usefulness would still need evaluation.

The same caution applies to retention. Richa described physicians reconsidering retirement and a substantial turnover decline at one client organization. She expressly presented Ambience as one enabler alongside other organizational initiatives. Without accounting for those parallel efforts, the company's role cannot be converted into a measured product-only retention effect.

Measure effects on users, teammates, patients, and operations. Continuing measurement after installation makes it easier to see which effects persist and which remain uncertain before assigning broader credit to the technology.

Agree on the Evidence Before the Next Value Review

At the next AI expansion, renewal, or value review, sit down with clinical users and nursing or physician leaders, implementation and commercial teams, and hospital operations, finance, and evidence staff. Agree on what workload relief would mean in this deployment. Put the plan in a short record everyone can inspect:

Users and workflow: who the tool is intended to help and which work is being evaluated.
Baseline and comparison: the work before deployment, an appropriate comparator, and the follow-up period.
Use: eligible encounters and the denominator behind any utilization claim.
Remaining work: who reviews, corrects, or coordinates afterward, including work passed to teammates.
Effects: staff experience, patient-care and efficiency measures, and a separate coding audit when coding is part of the offer.
Accountability: a data source and a named owner responsible for presenting each measure.

At the review, show what improved, what work remained, and what can reasonably be attributed to the product. A documented task benefit may support a narrower claim than a staffing, retention, or financial promise. If the comparison or attribution is incomplete, qualify the claim and keep collecting evidence; withhold broader assertions the findings do not justify.

Mary's nursing account brings the decision back to patient care: did clinicians gain time they could use, or did necessary work move to someone else? For your next review, bring one workflow's before-and-after record and ask the clinicians that question before promising them their time back.

About the Conversations

These are three separate Outcomes Rocket conversations, not a shared case study. During the time of recording, the guest names and titles were as follows: Mary Varghese Presti, Corporate Vice President of Microsoft Health and Life Sciences. Dr. Richa Gupta, Vice President of Care Transformation at Ambience Healthcare. Jeroen Tas, Chief Innovation & Strategy Officer at Philips.