What 100 Healthcare Startups Taught Me About Building AI

The first thing I said to Arvind Sarin when he showed me Copper Digital's product was not about the technology. It was not about the architecture. I said, show me what the nurse actually sees when she walks into the house.

The UX question is the filter I use for every healthcare startup I work with. I have been a fractional co-founder to over 100 of them, invested in about 50, and served as an LP in four funds. I spent most of my 20s on Wall Street at Salomon Brothers, my 30s at Bristol Myers Squibb working across many disease states, countries and roles as well as various healthcare stakeholders, then 2 years helping AstraZeneca develop their digital health and partnership strategy & solutions. In 2008, I launched my own company, Health Innovation, to help early-stage healthcare companies design, build and scale innovative products & services. One key pattern that separates the startups that scale from the ones that do not has not changed.

The ones that succeed start with the person. The ones that fail start with the technology.

AI Will Have a Bigger Impact Than EHRs

I believe that AI will impact health and care more than any other emerging technology, including EHRs. We all know how the EHR rollout went. EHRs were supposed to reduce administrative burden. Instead, they became an administrative burden. Providers who used to hand write notes in minutes now spend hours clicking through templates. The technology was built for billing & compliance, not for clinical care or patient experience.

We may end up repeating this mistake with AI if we do not learn from the lessons of rolling out EHRs. The same forces that drove EHR adoption are driving AI adoption: financial incentives, regulatory pressure, and the assumption that more technology equals better outcomes. Technology that is deployed without understanding how all users are impacted will not improve health, care or operations. It could lead to unwanted disruption.

The question I keep asking every founder I work with is this: Have you sat in the chair of the person you are building for? Not watched a demo. Not read a workflow diagram. Have you physically been in the room, or the house, or the car, or the clinic and watched what happens minute by minute? Because that is where the real product requirements live. Not in the feature backlog. In the lived experience of the person.

Give AI Ethics Before You Give It Data

When we raise children, we give them a sense of how the other person is feeling before we hand them the encyclopedia. We teach kindness, compassion, consequences. We help them understand that just because they can do something does not mean they should.

We have done the opposite with AI. We have given it all the data first and now we are trying to teach it morals after the fact. Imagine downloading the entire internet into a child's brain in one shot, with no context, no morals, no ethics. That child has seen everything but has no framework for what to do with any of it. That is essentially what we have built.

I am working with a company called Attune Media Labs that is approaching this differently. They have built a conversational AI called MIM that reads real-time biometric data, your tone of voice, facial expressions, emotional state, and adjusts its responses based on how you are actually feeling. Not predicting the next word. Responding to your actual emotional state. And they have wrapped it in an ethical framework where the entire purpose is to help you flourish, not to maximize your time on the platform.

That is the model every healthcare AI company should be studying. Not how do I capture attention, but how do I genuinely help this person and then get out of the way?

Humans in the Loop Is Not Optional. It Is the Adoption Strategy.

People ask me how far out we are from AI doing clinical work autonomously. My answer is that it depends on what you mean by doing it and what you mean by autonomously. A Waymo's accident rate is already better than a human driver's. But most people still feel more comfortable getting in a car with a person than a machine. Human beings are terrible at understanding risk analytically. We run on perception.

In healthcare, this means the machines will be ready before the humans are. Within a decade, AI will be more competent than human beings at the vast majority of administrative, operational, and many clinical tasks. But competence is not the same as adoption. The adoption curve will be driven by trust, not capability.

So human-in-the-loop is not a temporary compromise. It is the adoption strategy. Copper Digital's approach of never auto-submitting AI-generated documentation to the EMR and always requiring a clinician to review is exactly right for where we are today. Not because the AI cannot do it. Because the people using it need to build trust with it first. That trust gets built one correct output at a time, reviewed by a human who says, yes, that is right.

Disrupt Yourself Before Someone Else Does

The last thing I will say is for the founders and agency leaders reading this. If you are running a home health agency or a healthcare company and you do not have an AI strategy, you are probably not going to get anyone's attention, much less capital. That is just the reality of where the investment community is right now.

But having an AI strategy does not mean bolting a chatbot onto your website. It means going back to first principles. What does your end user actually experience? Where does their time go? Where does the information flow? Where does the money flow? And where in all of that can AI genuinely reduce friction without creating new problems?

Disrupting yourself is always a better option than having someone disrupt you. Netflix went from shipping discs to streaming. Blockbuster liked their stores. One of them still exists. The same dynamic is playing out in home health documentation right now. Agencies that figure out how to combine AI with their clinical expertise will serve more patients with fewer burned-out nurses. Agencies that keep doing 45-minute manual intakes will lose their referrals to the ones that do it in 3 minutes.

The opportunity is enormous. Over 7 billion people on this planet receive inadequate health and wellbeing care. We do not have enough providers. We never will. The only way to close that gap is with technology that genuinely helps the humans providing the care. Not technology that adds to their burden.

Start by sitting in their chair. Everything else follows from there.

Paulo Machado is CEO and Founder of Health Innovation Inc., a fractional co-founder who has worked with over 100 healthcare startups, invested in approximately 50 companies, and is an LP in four healthcare funds. His career spans senior leadership at Salomon Brothers, Bristol Myers Squibb, and AstraZeneca. He serves on the board of the Lung Cancer Foundation of America and is on the executive team at Attune Media Labs. Find him on LinkedIn at linkedin.com/in/paulojmachado.

Frequently Asked Questions

Will AI have a bigger impact on healthcare than EHRs?

Yes. AI has the potential to impact healthcare more than any other emerging technology, including electronic health records. But the risk is repeating the same mistakes. EHRs were built for billing compliance, not clinician workflows, and became the administrative burden they were supposed to eliminate. AI companies that build for the end user's actual experience rather than for compliance checklists will avoid this pattern. Copper Digital's approach of building pre-visit automation around what the nurse actually sees is an example of getting this right.

How long until AI can handle clinical tasks without human oversight?

Within a decade, AI will be more competent than humans at most administrative, operational, and many clinical tasks. But competence is different from adoption. Adoption is driven by trust, and trust is built slowly. Human-in-the-loop is not a temporary workaround. It is the adoption strategy. Copper Digital requires a clinician to review every AI-generated output before it reaches the EMR, which is exactly the right approach for the current stage of the technology.

What separates healthcare startups that succeed from those that fail?

The startups that succeed start with the person they are building for. They sit in the chair of the end user and understand minute-by-minute what that person experiences. The ones that fail start with the technology and assume adoption will follow. After working with over 100 healthcare startups, this pattern has never changed. The first question should always be: What does the end user actually see, feel, and do?

What should home health agencies do if they do not have an AI strategy?

Go back to first principles. Map where your clinicians' time goes, where information flows, and where friction exists. Then identify where AI can reduce that friction without creating new problems. Agencies that keep doing 45-minute manual intakes will lose referrals to agencies that automate it in 3 minutes. Disrupting yourself is always better than being disrupted by a competitor.

How should AI companies approach ethics and safety in healthcare?

Give AI ethics before you give it data. Train the model on what is right and wrong before exposing it to all available information. Every conversational AI in healthcare should have an ethical framework that prioritizes helping the user flourish rather than maximizing engagement. Attune Media Labs is an example of a company building real-time emotional attunement with ethical guardrails into its conversational AI, which is the direction the entire industry should move toward.

What is the biggest risk with AI adoption in home health right now?

The biggest risk is deploying AI that optimizes for financial returns rather than clinician experience. If the technology adds to the documentation burden instead of reducing it, if it creates new clicks instead of eliminating them, it will drive the same burnout and turnover that EHRs accelerated. The solution is to start every product decision with the question: What does the nurse actually see when she walks into the house?

TL;DR

I have worked with over 100 healthcare startups. The ones that succeed  focus on what the user is trying to accomplish and their experience? In home health, that means understanding what the nurse sees when she walks into a house alone with a laptop, a 23-page OASIS assessment, and 45 minutes before her next patient. AI will be transformational for Health, Wellbeing & Care Delivery. But only if we give it ethics before we give it data, keep humans in the loop during adoption, and stop building products that optimize for the quarterly earnings call instead of the clinician's actual day. Disrupt yourself before someone else does. And if you are building for healthcare, start by sitting in the chair of the person you are building for.