Healthcare AI should give doctors attention back, not shorter visits

At healthcare conferences, the AI conversation tends to end up at the same question: will AI replace doctors? Christine Landry, global healthcare partner at Kyndryl, a technology services company that helps health organizations modernize and manage mission-critical systems and workflows, hears it often. Sometimes it comes as a question and sometimes as a comment, followed by a long argument. She thinks the more useful debate is somewhere else. The question she cares about is whether AI can make it easier for a doctor to give great care.

That changes how health systems should judge AI. If the main promise is saved clinician time, the obvious next step is to fit more patients into the day. Landry argues the real return is attention: more mental room for clinicians to think, use their judgment and focus on the patient in front of them. Building for that goal changes where AI goes, how it is governed and how success gets measured.

Why time saved is the wrong success metric

Landry starts from a familiar problem. Clinicians spend a lot of time finding information, chasing documentation, coordinating care and moving data between systems. None of that, she says, is why the physicians she knows went to medical school. AI can clearly take on some of the load: documentation, assembling patient history, spotting patterns, flagging risks and surfacing information a clinician would otherwise have to go and find.

Where she breaks from the usual pitch is what the freed-up time is for. She questions whether speed should be how healthcare measures success at all. Some organizations already put limits on how much time they allot to appointments, she notes, and tools now exist to do that. But a three-minute visit is probably not the innovation any patient is hoping for. "I'm not sure that I want my doctor optimizing for speed," she said.

Her alternative is to give clinicians time they can turn into capacity. "It's not about shortening the cycle," she said. "It's about actually doing something different with the time that they have available to them." In practice, that means a physician looking at the patient instead of a screen, noticing anxiety and picking up the visual cues a patient gives about why they really came in.

Too much data, not enough information

Wearables, care at home and more connected systems have sharply increased the data reaching clinicians. Landry says the real problem is less the volume than the fragmentation. In her words, "health care has somehow managed to have both too much data and not enough information exactly the same time."

Her example is a deteriorating patient. The clinician does not need another dashboard or data feed to say the patient is getting worse. They need the relevant observations, the patient's history and the risk signals, pulled together at the moment they sit with the patient and decide what to do. There is a human side to that moment too: the patient is watching the clinician and needs to see confidence in what comes next.

That leads to a design rule that cuts against how many products get sold. Landry says the best AI will be AI the clinician barely notices: no new screen, no new button, just the right insight appearing at the right time inside the workflow they already use. Seen that way, adoption is a design problem more than a training problem. If the tool is one more system to learn, clinicians push back. If it fits the steps they already take, the adoption question mostly goes away.

Timing matters as much as placement. Landry estimates a clinician has, on average, maybe eight minutes in the room with a patient, and patients rarely describe symptoms in tidy clinical terms. She argues that when an insight appears is almost more important than what it says. Getting that right is a workflow design job, not a question of model capability.

Deciding where automation stops before deployment

If AI is going to work inside clinical workflows, someone has to decide what it may do on its own. Landry describes those decisions as a spectrum rather than an on-off switch, and says she discusses them with clinicians and administrators every day:

• Automate: work that is routine, low risk and reversible, where you don't have to worry much about what happens if something goes wrong. Scheduling, retrieving information and administrative tasks fit here.
• Recommend, with the clinician deciding: AI says what it is seeing and what the clinician might consider, but the clinician makes the decision.
• Step aside: high-consequence decisions, where clinical judgment and human accountability have to be explicit.

The rule she stresses most is about timing: set those boundaries before deploying any technology. "The most expensive way to develop your governance model is to do it the way I just described, which is to do it as you go along," she said. She also swaps out the industry's usual phrase. "People say human in the loop. Prefer to say human in control." In her framing, AI-native healthcare does not mean autonomous healthcare. Better outcomes are the goal, and the higher the stakes, the more deliberate an organization has to be about where human judgment applies.

She does not think vendors, Kyndryl included, should set an organization's risk appetite, and she does not think the decision belongs to IT alone. She recommends bringing together clinical leaders, operations, technology, security, risk and governance and, above all, frontline teams. That group can answer the practical questions: what AI is allowed to do, what evidence it needs, what gets escalated, when a human must step in and who owns the decision.

She says this matters more as AI moves from providing information to taking action. "If AI is gonna act, at machine speed, our governance can't operate at manual speed," she said. Governance kept in a binder on a shelf cannot keep up with a system acting in real time.

She ties that preparation to how much risk organizations will take on. In her experience, risk tolerance varies widely. Organizations that designed the interaction between people and technology into their workflows from the start tend to get more comfortable with automation. Those that have not mapped their workflows still have, as she puts it, a bit of a fear factor, which she says she understands.

Trust is designed, not announced

Clinician caution about AI makes sense in a field where decisions can be life or death. Landry argues the industry sometimes makes trust more technical than it needs to be. Physicians do need confidence in the quality and provenance of the data behind a recommendation. They also need answers to plain human questions: Why am I seeing this recommendation now? What information did it use? What is the AI allowed to do? What am I expected to do with this? What happens if it's wrong?

The part teams most often underestimate, she says, is organizational change management. Good technology and strong governance will still struggle if the people expected to use them were not brought along. That means involving clinicians early, letting them help shape the workflow, explaining why the technology is coming in and what will and will not change for them, and asking what they actually need. You cannot, she says, drop a tool into a hospital on Monday morning and announce that everyone can trust it.

Trust, by her account, builds through experience. People need to see that decisions are traceable, understand the boundaries, know who is accountable and feel they had a say in how the workflow was designed. She treats those as conditions to build, not boxes to tick.

Start with one workflow and measure what changed

Asked what sets apart organizations that are really changing care with AI, Landry narrowed it to one trait: they choose deliberately where to start. Instead of taking a long list of use cases and counting how many reach production, they pick a few workflows where they want to change an outcome. Her criteria for that first pick: the workflow is reasonably well understood, it matters to patients, clinicians or ideally both, and its outcome can be measured.

Those early workflows also build trust, she argues, because people can see what changed and why. That evidence is what makes scaling possible. The organizations doing this well review the workflow end to end, redesign it around people, AI and systems together, and handle the less glamorous groundwork: data, integration, security, governance and workforce readiness. They don't measure how much AI has been deployed. They measure what changed for patients, clinicians and the organization.

Her advice to leaders is to map the friction before choosing any technology. Candidate workflows include scheduling, clinical documentation and aftercare coordination. Within the chosen one, she suggests asking where the delays are, where people hunt for information, where handoffs are unnecessary, where human judgment really matters and where risk gets introduced. Only then should the question become how people, AI and technology can work together to improve the outcome. Skip that step, she warns, and you end up with a solution looking for a problem. It is a little "garbage in, garbage out," she said: an agentic framework placed on top of a poorly understood process will not fix it.

The measure that matters is attention

Landry's closing argument flips the common assumption that more technology means a less human experience. Patients do not arrive with a neatly structured history. They hesitate, remember things halfway through, tell stories about their families and sometimes leave out the most important thing. If AI can quietly handle listening, retrieval and routine coordination, she said, "what we're really giving clinicians back is attention and that presence. And I don't think there's anything more human than that."

The ultimate measure of healthcare AI, in Landry's framing, is not how much it automates but how much humanity it gives back. Patients, she notes, don't leave a clinic admiring the tech stack. They remember how they were treated and how they felt.

She puts some of the work on patients as well: insist on answers and push clinicians to make time for them as people, because patient demand is part of how the system changes. For health system leaders, the practical takeaway is three questions to settle before buying anything. Which workflow are we redesigning? Where on the automation spectrum does each step belong, and is that rule enforced in the workflow? And will we measure whether clinicians got their attention back, or only whether visits got shorter?