The arrival of artificial intelligence generally, and ChatGPT specifically, has generated a massive amount of discussion about both opportunities and potential threats. Within healthcare, the ChatGPT discussion has primarily focused on patient- and clinician-facing applications, with aims such as achieving scale in primary care, reducing the burden of clinical documentation, and facilitating clinical research.
Less has been written about the business side of healthcare, where presumably ChatGPT could be no less of an essential tool, as well as causing new problems. For a research/knowledge/insights-driven business such as Union, the opportunity would clearly be to drastically improve efficiency in the research process. The worries range from the immediate (what quality of the information does this tool generate?) to the existential (is human expert-led research about to be drowned out by a flood of AI-generated content?)
To gain some first-hand experience, I spent time putting the tool through its paces, quizzing it about healthcare strategy research-type topics such as the role of private equity in healthcare, Medicare program mechanics, players in the life science industry, the pharmacy value chain, and some hard-to-find info about provider contract rates—topics that that overlapped with live projects I am currently working on.
Here's what I learned.
Working with AI is like working with an early-career research analyst.
A core part of working in a mature information organization is working with early-career staff, in a role such as ‘analyst’—breaking projects into assignments they can do to help, editing what comes back, and coaching them so they can grow, thrive, and make ever-more insightful content in the future.
I found partnering with ChatGPT was eerily like that. I was asking it a wide range of real questions, which entailed carving out the subset that seemed “assignable” to me from larger-scale projects. The question-and-natural-language-answer interface made it feel like I was partnering with someone. And, just like working with an early-career analyst, I took every answer that came back with a ‘trust but verify’ attitude. ChatGPT was new to my team; I did not have a sense of what it could do well, vs. what its limitations might be.
ChatGPT vs. human analyst?
For many readers, AI is a threat to job security or growth: "Does this mean human analysts won’t be needed in the future?" My opinion is that you can dial down that immediate worry. While there were areas where ChatGPT grossly exceeded my expectations—in the big picture, the strong human analysts of my experience came out way ahead. And the whole experiment helps show how we should all be thinking about this kind of tool—not in replacing human researchers, but in helping them breeze through lower-value, time-consuming initial tasks, allowing them to free up time and energy where it is needed most in this complex world.
But I’m getting ahead of the story. Let’s start with how ChatGPT did vs. a hypothetical human analyst using real-world healthcare strategy research questions.
ChatGPT is the hands-down, no-question winner at speed in synthesizing what’s available from secondary sources. No real surprise there, but wow.
It was a foregone conclusion that ChatGPT would be tremendously faster. But seeing that time savings playing out in assignments I would have given a real human really brought home what a productivity enhancer this tool can be. For example, I asked it to write a general summary of the kinds of companies and major players that are found in the life sciences sector, and I got back a perfectly plausible answer in ten seconds.
The answer it produced wasn’t perfect (more on that in a moment). But 10 seconds vs. half a business day? That’s impressive. Especially because, just like working with a real analyst, working with ChatGPT ends up generating basically a first pass that needs to be followed up with a series of iterations—but doing so at a rate of a new draft every minute is quite a lot quicker than the time human teammates have traditionally needed to turn things around.
ChatGPT also wins on speaking specialized language from Day One.
Unlike any real-world human analyst I ever worked with, ChatGPT arrived ready to talk shop in specialized healthcare industry language. For example:
It succinctly explained, in natural language, the difference between Medicare part B and part D drugs. It defined the difference clinically (with part B being administered by a provider in a clinical setting such as chemotherapy, injectable medications, and immunosuppressive drugs, and Part D drugs being patient-administered, pharmacy-filled drugs)—and also the different cost-sharing arrangements that each entails for patients.
It recognized and correctly defined the term ‘drug channels.’ This was a particularly impressive moment for ChatGPT in my eyes, as not only did it correctly explain that the term refers to the entire ecosystem of drug manufacturers, wholesalers, distributors, PBMs, and so forth—it also volunteered something that felt pretty close to an insightful comment: “The term ‘drug channels’ highlights the fact that prescription drugs typically pass through multiple layers of intermediaries before they reach the patient, and that each layer of the distribution chain adds its own markups, fees, and rebates.” See what I mean?
Where ChatGPT would tie against a weak analyst—but not a strong one: ‘sometimes in error, never in doubt.’
Going back to the life science sector question, I felt that here is where AI showed me that troubling shortcoming sometimes found in human analyst: ‘sometimes in error, never in doubt.’ It told me confidently that the life science sector was composed of pharmaceutical and biotech companies. When I asked it, hey wait a minute, “Aren’t device makers a part of the life science industry?” it shamelessly answered: “Yes, in fact, device makers are, an essential part of the life science sector.”
Here is where I started to lose trust in the ‘partnership’. I began to get the uneasy feeling that this analyst was working hard to tell me what I wanted to hear—and also that it was more interested in providing a simple, confident answer than in getting things exactly right. Red flag.
What ChatGPT cannot do, that a strong analyst can do: divine and reflect the ‘why’ behind questions.
Spending time with ChatGPT helped me see the ways in which, despite its interactive format, ChatGPT is a tool for research, and not a replacement for human researchers–largely because it didn’t understand the ‘why’ behind questions as a human being would. And this inability to channel the audience (in this case, me), showed up in many of its answers. A few examples:
It doesn’t volunteer the descriptive data one needs to truly understand a landscape
When I asked chat GPT “What healthcare areas is private equity investing in?" it gave me a non-quantified list. With a distinct feeling of early-career-analyst-coaching deja-vous, I had to directly ask it, “could you get me some data to describe how much PE is investing in different parts of the industry?” ChatGPT didn’t take my follow-up as a criticism of its work, or as a hint about how to improve. It confined itself to informing me that, circa 2021, 46% of PE deals were made in services (Pitchbook), 24% in tech (Bain and Company), 13% in pharmacy and biotech (PitchBook), and 5% in real estate (JLL).
Since it does produce data when asked, maybe it’s nitpicky to fault ChatGPT for not including descriptive data. But, to me this seemingly minor shortcoming is a signal of not getting the ‘why’ behind questions; a strong human analyst like the ones I’ve had the privilege of working with would know, or at least quickly learn, that the ‘why’ behind questions like that one is a desire to understand the lay of the land—and the relative magnitude of different elements is an important feature in that landscape.
It's inconsistent on including a heads-up about need-to-know controversies.
In the case of 340b, it told me both how the program entitles some healthcare entities to discounts on outpatient prescription drugs—and how there is a push-and-pull among ‘some stakeholders’ who feel this is a critical revenue pillar for under-reimbursed hospitals vs. those who feel the discount is being inappropriately used. In the case of PBMs, ChatGPT told me both what they are supposed to do (act as an intermediary, secure lower drug prices), and also what they are under scrutiny for (lack of transparency, accused of contributing to higher drug prices).
But it didn’t always point out market sensitivities. For example, it told me all kinds of things about what PE is doing in healthcare when I asked—but kept mum on any issues or problems until I directly asked it, “Are there any controversies involving PE and healthcare services?” Only then did it explain what it obviously ‘knew’: that while ‘some people’ feel that it brings needed resources and efficiencies, others think PE can drive healthcare service cost up, quality down, and restrict the autonomy of physicians.
Similar to my issue with ChatGPT not volunteering descriptive data, in my view a human analyst would proactively flag controversy and issues. The human analyst knows both that this tends to be where things get interesting, and also, that failing to warn of sensitivities in any given topic leads to trouble down the line.
ChatGPT only knows things people will say “in public” and in print.
I asked ChatGPT several variations on, “how do negotiated provider rates in Medicare Advantage compare to fee-for-service Medicare, or to negotiated provider rates in commercial contracts?” and its struggles to answer this question were enlightening.
To its credit, it generally appeared to understand my question, another impressive display of coming in knowing specialized healthcare language. But despite seeming to get my drift, ChatGPT could not provide a satisfactory answer. It variously told me that the answer depended on a lot of different factors, could not be known because it was contained in confidential contracts, and might be available somewhere on the CMS website. Which brings me to my next point.
ChatGPT can’t do off-the-record primary research.
ChatGPT’s shortfall on negotiated rates (above) shows a couple of different problems. It’s actually a bit of a secondary-research fail; it didn’t spot a 2020 article by CBO researcher Dana Pelach, “Prices for Physicians' Services in Medicare Advantage and Commercial Plans” in the journal “Medical Research and Review” (available at least in abstract form online). That study used 2014 claims data to reveal that–at least as of 2014, and in that sample—MA prices for physician services were “close to Medicare FFS prices, varied minimally, and were similar in and out of network”–in contrast to commercial contracts, in which rates were substantially higher and more variable.
But there’s a more important issue here. Even if ChatGPT had found this result, in my view, this is an instance where human researchers would shine by switching gears to off-the-record primary research. A real-world provider executive would give a trusted human research team, in a confidential interview, up-to-date information on their organization’s MA negotiated rates and how those compare to the rest of the book of business. A high-quality relationship with well-placed plan executives would garner a similarly candid and much more wide-ranging description of that same landscape. But even with providers alone, researchers could use multiple interviews with different organizations, and a lot of follow-up questions about different specific market dynamics in each case, to qualitatively extract the moving pieces to advance industry knowledge about the state of negotiated MA base rates and what drives them higher or lower.
ChatGPT won’t give me feedback on my questions.
Here’s maybe where I missed working with a human analyst the most. Unlike the human, ChatGPT won’t tell me what was easy vs. difficult to find among the results it came up with, which is important information for judging where to spend research time and effort. It won’t tell me I’m asking the wrong question. It won’t point out that there’s a more interesting and important adjacent question we should be focusing on instead.
7 takeaways about what all this means for the healthcare insight business
Taking a step back, here are my thoughts about the immediate-term role of AI technologies like ChatGPT for research/knowledge/insights-driven businesses in healthcare.
#1 Rather than replacing early-career human analysts, reporters, etc., this technology should be used to help make their lives easier and progress faster so they can shine in their fields. I have seen many early-twentysomething humans completely flummoxed by the task of composing a passable interview outreach email template; ChatGPT does this for them in five seconds. It will also help human analysts get rapidly up to speed on the general outline of unfamiliar, specialized industry topics, who the major players are, where to find general statistics, and so on.
#2 Identifying and disseminating internal best practices on effective use of this tool will be an operational differentiator. Every analyst I ever worked with knew how to use a search engine—but some used it much more effectively than others. The potential productivity and efficiency gains for an organization if all staff know how to use AI wisely and well are huge. The faster knowledge organizations can learn how to apply these technologies—the faster they can tap into those gains.
The competitive differentiation of an insight business will hinge on its ability to:
#3 Know what questions to ask. AI aside, this is the standard problem of big data; we routinely talk with healthcare organizations that are sitting on reams and reams of amazing proprietary information, but who need expert help making productive use of it. The successful insight business of the future will be the one that asks better questions—the ones that channel what the intended audience does not know, and needs to.
#4 Spot limitations. AI (as currently deployed) is awesome at taking a first swing at things. Sending that first swing to press opens up the possibility of putting low-value, wrong, incomplete, or skewed information out under one’s precious information business brand. It will behoove information businesses that want to keep their credibility to use this tool—like any web search—as raw material—only the first step in a long process involving skeptical, knowledgeable humans at every step of the way.
#5 Do confidential primary research. The easier it is to scrape up and synthesize secondary public-domain research, the greater the need for primary, off-the-record research with stakeholders. This will be the way any expert or insight business will advance (and stay ahead of) the frontier of what is known.
#6 Look beyond conventional wisdom. AI is a wonderful tool for figuring out what is generally known, and generally thought, about a given topic. That means what customers will need from insight businesses will be the idea to see the negative space—areas that are under-discussed, incorrectly perceived, or in need of a wholesale re-framing.
A bottom-line prediction about the future for organizations like Union
#7 The ultimate coin of the future insight business will continue to be human capital. Insights businesses will need to deliver differentiated value to a market looking for real insight and answers it cannot find elsewhere. To do this, it will need staff with certain make-or-break characteristics, such as: experience, creativity, skepticism, audience empathy, and the ability to forge high-trust relationships with others in the industry.
About Union
Union is a researcher-founded insight company. We understand the history, incentives, and forces that drive the healthcare industry. We also have a passion for learning and teaching others about how it all works. Want to chat with a human? Reach out at info@unionhealthcareinsight.com
About Amanda
Amanda has worked for 20+ years in information organizations – first, a short stay in print media and then a long (and ongoing) one in healthcare strategy and best practice research. She holds a relevant-to-this-post graduate degree in Communication, Culture, and Technology, She finds it helpful to remind herself something learned in that program: Most information technologies are initially viewed as a threat to culture, but over time, important benefits also end up being found.
Comments