Insights from Marvin Pascua, VP, Solution Consulting at b.well
As AI moves from answering health questions to acting on them, the riskiest failures are the ones nobody sees. In this article, we explore where the gaps in your health record come from, why AI assistants sound just as confident when they’re working from half of it, and what it takes to close those gaps before a proactive agent acts on them.
Key Takeaways
- AI health assistants can only reason over the records they can see. When records are missing, stale or contradictory, the answer still sounds confident.
- The three gaps: records that were never found, connections that silently stop updating and sources that disagree with each other.
- Proactive AI agents raise the stakes because they act on whatever record they have, and an approval prompt doesn’t show what that record is missing.
- Much of the fix starts upstream: be able to actually locate records, check daily for updates, flag lapsed connections and reconcile medications before the model sees them.
A Confident Answer, Built on What It Can See
You’ve felt mildly lightheaded for a couple of days. It’s Saturday and your doctor’s office is closed, so you ask your AI assistant: Could this be the new medication I started, or is it the same thing that happened last spring?
The answer comes back in seconds, calm and specific. Whether it’s right depends on something you can’t see.
Why Health Is Where AI Assistants Matter Most
More and more, an AI assistant is where that question goes. Records from every doctor and health system you’ve seen, prescriptions, lab results and the steady stream coming off a watch or a ring add up to far more than anyone can sort through on their own. Frontier models can work across years of visit notes, labs and medication history, connecting a prescription started at one office to a symptom recorded at another.
What we’re hoping for is an n of one. A web search tells you what lightheadedness usually means for most people. An assistant with your records should tell you what it means for you.
What Finance Taught Us About Missing Data
It’s the same thing we’ve come to expect on the financial side. For years, many of us have linked checking, credit cards and retirement accounts to budgeting and investing apps so we can see everything in one place. Now we’re handing that same connected picture to an AI assistant and asking it to help manage our financial health portfolio. When I connected my accounts, everything came in except my mortgage. The lender wasn’t one of the big names, and the tool I was using had no way to connect to it, so every answer about my finances was missing my biggest bill.
The difference is that I knew the mortgage was missing. With health, you usually don’t. If the assistant can’t see last week’s prescription, nothing on the screen says so, and the answer comes back just as fluent and just as confident either way. Models make their own mistakes, but this one starts upstream. The model can reach a careful, plausible answer from the part of your history it can see. You went looking for an n of one and got an n of some.
Three Ways Your Health Record Goes Wrong Before the Model Ever Sees It
1. The Records Were Never Found
Some of that gap starts with finding your records at all. Many ways of connecting ask you to pick your health system from a list, which assumes you know it.
You know your doctor’s name. You may not know the group she practices under, or that the specialist she referred you to is part of a system you’d never heard of. That system is healthcare’s version of my mortgage lender, and when it’s missing, there’s no error message. The assistant answers with what it has.
2. The Records Went Stale
Even the records you did connect have to stay current. When you check on your finances, you expect today’s numbers, because the point of asking is to know where things stand right now. You expect the same from your health, usually at the moments you least planned to be asking, like a new symptom late at night.
But a connection can stop updating without you knowing, and the access you granted can expire without much warning. The specialist visit from last week isn’t there, and neither is the prescription you picked up after it. A record that stopped updating in March is still a March record, even when you’re asking in November.
3. The Records Disagree With Each Other
Then there’s whether the record agrees with itself. Your history is pieced together from every place you’ve been seen, and those places don’t always tell the same story. Your doctor stops a medication, but another office’s list still shows you taking it.
The model can only work with what’s in its context window. That outdated list can even carry the more recent timestamp, because it was sent later, and the model can reason carefully all the way to the wrong answer.
Proactive AI Agents Change the Stakes
Until recently, all of this ended with a person reading the answer. You asked, it responded, and sometimes you caught what was off. That’s changing. Proactive personal agents are a new class of assistant that arrived this year, and they don’t wait for a question. They work in the background toward goals you’ve set, checking what they’re connected to and acting within the permissions you’ve given them.
For health, that’s overdue. A goal-oriented agent that keeps working on your behalf can help with the follow-through between visits.
What Happens When an Agent Acts on Half the Record
It also raises the stakes on every gap above. An agent works from the sources it’s connected to and whatever those sources last sent. It can’t know about a record no one found, a connection that went quiet or a list that’s out of date unless something tells it. So it acts on what it has.
It could:
- Remind you to refill the medication your doctor stopped
- Book a follow-up you already had, because that visit happened somewhere it couldn’t find
- Prepare a summary for your next appointment that leaves out last week’s specialist visit
Approval steps usually cover the consequential actions, like sending a message on your behalf. None of those steps tells you what the record is missing.
What Changes When the Record Is Found, Current and Reconciled
Now give the same agent a better record.
Found
A provider directory maps your doctors’ names to the health systems where they practice and to each system’s available FHIR endpoints, including many smaller and rural ones, so you know what to connect.
Current
Connected sources are checked for updates every day, and when a connection lapses, the agent or app it’s integrated with is notified, so you can be asked to reconnect.
Reconciled
Records from each connected source are matched, de-duplicated and compared, and medications are sorted into active and past before the model sees them. Where sources still disagree, both versions come through with their source attached instead of one being picked for you. Not every conflict can be settled from the data alone. Some still need you or your doctor.
When the record shows the medication was stopped, the agent has no reason to send the refill reminder, and once last week’s visit posts, it shows up in the summary for your next appointment. No record will ever be perfect, but when the gaps and conflicts are known, the agent can say what it’s missing or unsure about, and check with you, instead of acting as if the record were complete.
Trust Is What Gets People to Connect
We’ll keep managing more of our health through AI, and we should. Trust is what gets us to connect. If the answers turn out to be built on half the record, the utility isn’t there, and we’ll hesitate before relying on them again.
Agents can finally do the follow-through health has always needed, as long as they know what’s in the record and what isn’t. Before you trust one to act, ask what it searched, what it couldn’t reach, when it last refreshed and what it couldn’t reconcile.
I’ll be at HLTH 2026 in Las Vegas discussing this with peers and showing what it looks like in practice. If you’re attending, I hope you’ll stop by the b.well booth #4751 to see how health.network connects AI assistants and agents to a consumer-controlled health record. If you’re building assistants or agents that touch someone’s health and can’t make it to Las Vegas, reach out and I’d be glad to walk you through how we address this.
