Tom Hittinger

The through-line

Hittinger's recurring argument is that the constraint on AI in healthcare is no longer the technology, it is the leap from working demonstration to enterprise scale, and the way to make that leap is to stop framing AI as a better way to do a task and start framing it as a way to transform a process. He describes a market that has already proved the business case, with over 75% of healthcare and life sciences organisations saying AI has met or exceeded their ROI goals, but where only "10% to 25% are actually scaling fully into production and thinking we are enterprise-ready in AI" [1]. Two things are true at once in his telling: "people are getting a lot of value already from AI, and two, there's more value to be gained" [1].

The shift he tracks runs from the inflection point of late 2022 to the agentic era of 2025, and he frames it as a change in what technology is for rather than what it can do. He notes Google has been working on the transformer architecture since 2018, and that the underlying stack was always advanced, but that "it's different when you're talking to a client and a customer about an API versus a vision" [1]. Throughout, he anchors back to the same discipline: "the conversation has fundamentally always been around what are the problems that customers are trying to solve? What are the areas of greatest cost, greatest pain?" [1]

On the difference between AI as assistant and AI as collaborator

His sharpest distinction is between the first wave of AI and the agentic one. First-wave AI is "AI as an assistant, you prompt it, it responds," which helps a person do "task A, task B, task C more effectively, more efficiently" [1]. The real change is "going from assistant to true collaborator," using agentic platforms "not just do a single task, but to actually transform an entire process" [1]. He grounds this in what buyers actually ask about: not how to do one action better, but "How do I serve my patients better?", "How do I get drugs to market faster? How do I better identify the next candidate?" Those, he argues, are not task problems: "That's a process problem, and hundreds and hundreds of tasks" [1].

He offers a working definition that deliberately trades elegance for clarity: "Agentic AI is the best parts of robotic process automation combined with Generative AI" [1]. RPA automated a process end to end through predefined rules; agentic AI adds the ability "to reason, plan, and act end to end in a way that the doctor doesn't have to click the button every single time" [1].

On trust, and why usability comes before it

Asked what it will take to earn the confidence of patients, providers and regulators, Hittinger answers with usability rather than governance language: "The first thing for me is you have to be able to understand what you're building and actually how to use it" [1]. He is candid that agents are hard to build, walking through the reality that "we have to understand source system A, source system B, how they interact, how they exchange information, who needs permissions, what task needs to be performed. It's a lot of steps and it gets very, very technical very, very quickly" [1]. He points to agent space as interesting on exactly these grounds, because building agents should not be reserved for "a really, really technical data scientist or AI engineer" but open to "anyone who has the functional expertise and the goal in mind" [1]. Underneath this sits a strong statement about where the real problem lies: "The fundamental problem is how do I create the value for my customers and for my employees, for my clinicians, for my staff" [1]. Building an agent is a technology problem; step one for trust "is making it usable, so people can see it in action" [1].

On which use cases have actually gained traction

He is impatient with what he calls "the use case bingo game that we've been playing for the last few years," insisting instead on "anchoring to what is the value we going to create" [1]. He separates what already works from what could. Customer service leads, driven by a specific insight about health insurance: only 5% can explain basic terms like deductibles and out of pocket, and "that's fundamentally why people are calling in to call centers is I don't understand the product that I'm using" [1], so the win is helping the person answering the phone respond more easily. Second is code generation, where enterprise architects and developers have seen "a rapid uptick" in autocompletion that helps them code faster [1]. Third is knowledge management, searching protocols, handbooks and guides so that "nurses able to start their job more easily" [1]. Fourth, spanning pharma, biotech, health insurance and providers, is marketing campaign creation [1]. The common thread he draws is that "any area or any problem or pocket where there's a lot of paperwork or administrative paperwork or toil, that's where we can find efficiencies and value to be had" [1].

He also frames the whole field by the underlying problems it must solve: bringing the next drug to market "faster, better, cheaper", processing claims and servicing members in insurance, and among providers delivering excellent care "in a way that drives health literacy for patients, so they can actually follow the treatment protocol and get the outcomes and live the healthiest life that they want to" [1].

On what agents change at the bedside

His signature clinical example is clinical search and summarisation. Today one can build AI that lets a doctor see a patient summary and query a search bar for something like the last A1C measures [1]. The agentic version replaces the clicking with a goal: a doctor saying "I actually want to understand what this patient has been diagnosed for over the last 10 years", which then informs medical coders submitting paperwork and claims and informs the care plan, with "those 20 to 30 steps of them previously combing through every medical record, every chart, every image, every encounter" happening in the background [1]. The upshot he claims is "making information and insights more readily accessible", given that medical records function as "virtual filing cabinets in many ways" [1].

On discovery, access and the informed patient

Two horizons excite him most. The first is discovery, both scientific and clinical. He returns to literature search as the canonical case, noting there is more information coming from the medical and scientific community than any one person "at their full-time job is just to read all the new studies" could keep up with [1]. Making insight discoverable changes the care an individual provider gives, replacing "a legacy recommendation and evidence-based guidelines from 10 years ago" with something "fresh on the bleeding edge" and specific to a patient's demographics, location, medications and prior procedures [1]. The second is access: it is hard to see a doctor, visits are short, and personalisation is limited, so enabling patients to do self-education "with authoritative information before and after a visit that helps them have a better authentic conversation with their provider" is, in his view, "already happening with Generative AI today" [1].

On scaling now rather than later

His single piece of advice to healthcare executives is blunt: think "about scaling now, not later" [1]. He has "seen too many proofs of concept and even pilots reach a dead end for many reasons that could have been preempted" earlier in the process [1]. The example he uses is an after-visit summary or a nurse handoff report, outputs that depend entirely on the data underneath: "Your data has to be accurate. It has to be accessible, it has to be usable in the right standard and format" [1]. The gap between demo and deployment is where projects die, because "you can easily make something work for 10 users, but making it work for 10,000 users really requires a lot of elbow grease and leg work to establish that data foundation, to establish that right provisioning, to establish the right security and infrastructure" [1]. His conclusion is that infrastructure work is never wasted: "it's always time well spent investing in infrastructure for the sake of outcomes and for the sake of being able to build differentiated next generation AI applications" [1].

Takeaways

• The ROI question is settled and the scaling question is not: over 75% of healthcare and life sciences organisations report AI met or exceeded ROI goals, while only 10% to 25% have scaled fully into production [1].
• Value comes from redesigning processes, not accelerating tasks: serving patients better or getting drugs to market faster is "a process problem, and hundreds and hundreds of tasks" [1].
• His working definition of the new wave: "Agentic AI is the best parts of robotic process automation combined with Generative AI," able to "reason, plan, and act end to end" without a human clicking every button [1].
• Trust starts with usability, so agent building must be accessible to people with functional expertise rather than confined to data scientists and AI engineers [1].
• The proven use cases cluster around administrative toil: customer service, code generation, knowledge management and marketing [1].
• In insurance, the call centre problem is a comprehension problem, since only 5% can explain basic terms such as deductibles and out of pocket [1].
• Design for scale from the start, because moving from 10 users to 10,000 demands data accuracy, accessibility, standards, provisioning, security and infrastructure [1].
• Discovery is the horizon he finds most exciting: surfacing current evidence tailored to a patient's demographics, medications and procedures instead of guidelines from a decade ago [1].