Advance care planning is broken

Here's what I'm building to fix it

Photo by Juri Gianfrancesco

Hello friends! My how the time flies. Before I get to the article, I’m pleased to announce that the third annual Toronto Bioethics Workshop will be taking place on June 12th and 13th at the Jackman Humanities Building at the University of Toronto. The keynote speaker will be Nir Eyal from Rutgers University. If you’re interested in submitting a paper for consideration, the deadline is March 31st. If you wish to attend the workshop, registration is free but required. There’s more info here: https://philevents.org/event/show/145018

---

Last year, I wrote an article for this newsletter making the case for supply-side bioethics. The idea is that bioethics has historically accepted scarcity as a given, then focused on what to do in light of that scarcity. For example, there aren’t enough organs to give to everyone who needs one, so we need to figure out how to distribute the ones that become available. Bioethics is inextricably in the scarcity business.

But while the distribution of scarce resources is an important issue that isn’t going away, I argued that, for each issue, bioethicists should also attend to how we can create more supply. Research ethics is a classic example of ignoring this: research ethics boards enact well-meaning requirements that, in aggregate, significantly constrain the amount of research that gets conducted. Examples abound. IRBs/REBs are tasked with preventing harmful research without having to consider the harms of research not being conducted (by preventing new drugs and treatments from being developed). The natural consequence is that it becomes more difficult to conduct research. If you’re thinking that this is analogous to the Nuclear Regulatory Commission’s nearly fifty-year de facto ban on nuclear reactors, you’re not wrong! I’ll put the details in this footnote.

Some people reached out after I wrote that article and I had some nice conversations about its ideas. But then I had the bioethicist’s equivalent of a road-to-Damascus moment. The problem, I realized, is that I’m a bioethicist, so I can be doing something about supply-side issues. There are many problems out there, and the thing I realized is that I can build solutions to these problems instead of only talking about them.

Advance care planning is at the top of my list. Nearly everyone agrees that advance care planning is important. Every province has a legal process for people to make advance directives, health systems and hospitals have full-time teams who educate people on advance care planning, polls consistently find that Canadians believe ACP is important, and various groups, including Dying With Dignity Canada and Advance Care Planning Canada (ACPC), have resources for making directives and having conversations.

These are all valuable contributions. Still, we’re a long way off from solving ACP. There are non-ideal aspects of the process from beginning to end.

The first is that most people don’t make a directive or discuss their healthcare wishes with loved ones. ACPC’s 2024 national poll found that only seventeen percent of Canadians have an advance care plan, which is the same as it was in 2019 and lower than in 2021 (19 percent). The same polling finds consistent reasons why people don’t do it: it’s hard to think about, people already trust their loved ones to make the right decisions, people don’t feel confident communicating their wishes and values, people don’t have someone they can trust to make decisions, and people don’t feel supported to have these conversations. There are other structural factors not mentioned in the poll, including that physicians are usually unable to have these conversations due to time constraints.

The second problem is that most of these directives are written on paper and physically stored in the person’s residence, which means they have to be known about and then found when the person loses decision-making capacity. Data on this is sparse, but, anecdotally, I’ve had multiple clinical consults where the patient’s family says that they know their loved one made a directive, but they can’t find it.

The third problem is that directives become out of date. People have kids, get married, get divorced, get a diagnosis, or in some other way experience something that changes their priorities. If someone made a directive a decade ago, there’s good reason to wonder if the directive accurately reflects what he would want now.

The final problem is that many directives don’t help the SDM make decisions with confidence. In some cases, it’s because the directive is too vague. It will say “I want no heroic measures” or “don’t hook me up to machines”, leaving the SDM and clinical team to determine what constitutes ‘heroic’ or what counts as a machine. (If the latter seems clear, consider whether an IV drip counts.)

In other cases, it’s because the directive is too specific. If someone watched her father go through dementia, she might have clear wishes about dementia but leave out other scenarios. If she gets in a car crash, her SDM must then determine how her wishes about dementia apply to this very different scenario.

Of the seventeen percent of people who make an advance care plan, let’s generously suppose that seventy-five percent of the directives are available when needed, seventy-five percent are up to date, and seventy-five percent of those provide clear guidance for the situation the patient is in. This means that only 7.17 percent of people have an available, up-to-date, action-guiding directive.

This is bad. It means there’s a good chance that, at some point, over ninety percent of people receive treatment that doesn’t align with their wishes. It means that most SDMs are making decisions under uncertainty, which causes distress. This has further negative effects, including time off work, mental health claims, and presenteeism. Physicians, nurses, and other hospital staff have to navigate these conversations and experience distress of their own. It means that healthcare systems, which are already strained, are providing unwanted treatment. Advance care planning scarcity is bad for everyone.

Last year, I started thinking about this problem more seriously. (This is the explanation for my long absence from Value Judgments.) I wondered how, without changing the law or how healthcare is practiced, can we use technology to address as many of the problems with advance care planning as possible?

The answer has to address all four problems simultaneously. A solution that gets more people to make directives but doesn’t improve the quality of those directives just creates more documents that fail under pressure. A solution that produces better documentation but doesn’t reach more people is an improvement at the margins.

So here’s where I ended up. The core of the problem isn’t the document, it’s the conversation. Or, more precisely, the absence of it.

When I think about the best advance care planning I’ve seen, it wasn’t someone filling out a form. It was a skilled clinician or social worker sitting with a patient and asking good questions. Not “do you want CPR?” but “tell me about a time you saw someone go through a serious illness—what worried you about that?” Not “check the box for life-sustaining treatment” but “if you couldn’t do the things that make your life meaningful, how much would you be willing to go through for the chance of getting them back?”

These conversations identify what actually matters. They catch the tensions—the person who says she wants everything done but also says she’d never want to live in a nursing home—and work through them in real time. The product of a good ACP conversation isn’t just a set of preferences. It’s a document that explains why someone wants what they want, in enough detail that an SDM facing a situation the directive didn’t anticipate can reason from the person’s values rather than guess.

The problem is that these conversations require time and skill that the healthcare system can’t provide at scale. Physicians don’t have thirty minutes per patient to do this well. And the structural factors in the polling data—people finding it hard to think about, not feeling confident, not feeling supported—aren’t going to be solved by better pamphlets or even digitizing forms.

This is what I’m building. Tiller Labs is a company I co-founded to improve high-stakes healthcare decision-making. Our first product, Tiller, is an AI-guided advance care planning platform that walks people through a structured conversation that does what a good clinician would do. It asks plain-language questions about values and tradeoffs, catches vague or conflicting preferences, works through them interactively, and produces documentation that connects a person’s stated wishes to the reasoning behind them.

I want to be specific about what “AI-guided” means here, because I know it raises reasonable questions. One is about scope. Tiller is not making clinical decisions or telling people what they should want. It’s guiding a structured elicitation process that helps people identify and articulate their preferences. Another is about safety. The conversation is designed by a bioethicist and built with domain-specific constraints. It’s not a general-purpose chatbot that happens to ask about healthcare. Every question, follow-up, and validation layer exists because we know what can go wrong when these conversations are done poorly, and we’ve spent months on testing to prevent them from going wrong.

People also might wonder whether people will actually talk to an AI about something this personal. Our early testing is finding that many people find it easier, not harder. There’s no fear of judgment, no pressure, no time constraints, and no worry about burdening someone. The barriers the polling data identifies—such as not feeling confident or not feeling supported—turn out to be partly social barriers. A well-designed guided conversation can lower them in ways that a pamphlet or even a willing family member sometimes can’t.

Instead of a checkbox form, the output has three components. The first is a directive. The second is the transcript of the conversation, so the SDM can look at the details. Finally, the SDM and clinician can access a decision support tool, which also uses conversational AI, to get more clarity about the person’s wishes.

Suppose that Sharon is Zach’s SDM. Yesterday, Zach had a stroke and is unconscious in the hospital. Zach’s clinical team wants to know Zach’s wishes. Sharon can use the directive, but she wants to make sure she’s understanding it correctly. To do this, she uses the decision support tool. She tells it that Zach had a stroke and asks if he and Tiller discussed strokes specifically. If they did, the tool gives the specific sections of the conversation. If not, it finds the parts of the conversation that best apply to Zach’s current situation—for example, what he would want if he were unconscious and his recovery was uncertain—to help Sharon think through the decision with confidence. With this tool, Sharon doesn’t have to guess or carry the weight of wondering whether she got it right.

That addresses the quality problem. The accessibility problem is addressed by the fact that the documentation is digital, shareable, and flows directly to the people who need it: family members, substitute decision makers (SDMs), and, eventually, clinical teams. It doesn’t live in a box or filing cabinet.

Once a person has worked through the conversation process and made a directive, Tiller will send regular review reminders. But instead of prompting the person to change her directive only if her wishes have changed, it asks her to confirm that her directive still accurately reflects her wishes if she doesn’t want to change anything. That stops directives from becoming out of date.

The uptake problem—getting more people to actually do this—is where I’ve had to think most differently. The standard approach has been to focus solely on integrating ACP into clinical care: train physicians, build it into annual checkups, fund hospital-based programs. These are worthwhile efforts, but they’re slow, resource-constrained, and they reach people when they’re already in the healthcare system, often when a crisis is already developing.

What I’ve come to believe is that advance care planning also needs to move upstream of medicine. The place where most working-age adults can be reached is the workplace, not the doctor’s office. Employers and benefits providers already pay for the downstream consequences of unpreparedness through caregiver leave, mental health claims, stress-related productivity loss, and employees making impossible decisions for loved ones without support. Advance care planning is upstream prevention.

So Tiller distributes through employers, employee assistance programs, benefits carriers, and financial advisors, which are the organizations that have existing relationships with the people who need this and existing reasons to care about it. This isn’t the approach I would have predicted a few years ago, but it’s where the evidence points. Clinical channels are important, and we’ll get there, but non-clinical channels move faster and reach people before the crisis rather than during it.

I’ve been having conversations with benefits carriers and EAP providers across Canada over the last few months, and the response has been positive. The gap in current offerings is clear. What has been missing is a tool that’s built with enough domain expertise to handle the sensitivity and complexity of these conversations well.

I’m writing this partly to share what I’ve been working on and why, and partly because I think the supply-side framing from last year’s article applies directly here. The advance care planning problem isn’t going to be solved by distributing the existing supply of good conversations more efficiently. We need to create dramatically more supply. That’s what we’re trying to build.

If you work in employee benefits, workplace wellness, or financial advisory and any of this resonates, I’d welcome a conversation. And if you know someone in those spaces who has been thinking about caregiver support or end-of-life planning, an introduction would mean a great deal. We’re entering through workplace channels first, but we’re also setting up clinical pilots, so if you work in either space, get in touch.

Finally, if you’d like to try Tiller and tell us your thoughts, we’d love to hear from you.

I’m at eric@tillerlabs.com, or you can simply reply to this email.

There’s lots more to come.