My best build got no feedback. A BAA post hit 9,400 impressions | Ep 330

Context

A BAA, or business associate agreement, is the contract HIPAA requires before any vendor touches patient data. I posted on LinkedIn that Anthropic's own page says it will sign BAAs without the millions in committed spend an MGMA conference attendee had described. By the day I recorded this episode, October 23, 2026, the post had 9,400 impressions and over 40 comments, against the 44 to 89 impressions my posts usually draw. The demo I'd recorded the week before, a free template that keeps Claude skills in one GitHub repo and the build I'm proudest of, drew almost nothing. This episode covers why the paperwork post broke through, the three ways the comment thread mapped for keeping patient data safe with AI, the catches in each, and what I did with the thread afterward. I'm Chris Duff, an AI builder with a medical background who has directed engineering teams for 10+ years, and Serious AI Partners helps mid-market companies, healthcare first, put AI to work. Everything here is a builder's read of the vendors' own pages, not legal advice.

Key Takeaways

• A question people can answer beats a demo they can only watch. Anyone could check Anthropic's BAA page and reply from their own work, so strangers started debating each other before I had written a single reply.
• There are three doors for keeping patient data safe with AI. Strip the identifiers before the data reaches a model, run the model on your own machines and own the security yourself, or run it inside a cloud you already have a BAA with (AWS, Microsoft or Google) and configure that cloud correctly. Each one works when done properly, and each hands you a different job.
• You can't redact a voice. HIPAA's own de-identification list names voice prints, so the speech-to-text step inside an ambient scribe needs its own BAA even if the transcript is scrubbed afterward.
• One BAA covers one vendor, and the data may run through several. Anthropic's page lists MCP connectors, Claude in Chrome and Skills as outside its BAA, so each connector needs its own coverage or no patient data at all.
• No paperwork controls what an agent says out loud. A psychologist in the thread had a call agent read a patient's full date of birth back on a call. A short list of things the agent must never say, plus a regular pull of transcripts to check it held, closes that gap.
• The comments are the asset. One thread gave me nine angles, two founders offering to show their single-BAA setups (which I'll test on dummy data), this episode and the next few posts.
• Map where your patient data goes this week. Pick one AI tool your practice uses and list every piece of technology the data passes through (the microphone, the transcription tool, the model, every connector), then check that a BAA or your own security covers each one. Vendors should run the same exercise from their side.

Prefer audio? Listen to Serious AI with Chris Duff on Spotify, or wherever you get your podcasts.

(Podcast transcript below)

My best build got no feedback. A BAA post hit 9,400 impressions. Quick definition, since not everyone lives in this world, a BAA is a business associate agreement. The contract that HIPAA requires before any vendor touches patient data in the healthcare industry. So if a medical practice runs an AI tool on patient info, the company behind that tool has to sign one, BAA that is. So today, what we're going into is what that post taught me about content creation and what the common thread taught me about keeping patient data safe with AI. And, you know, why I got into, you know, or even posted this content in the first place was a few weeks ago, after I was coming home from the MGMA annual conference in San Antonio, there was a session where an attendee mentioned at the very end that Anthropic won't sign a BAA without millions of dollars committed for one.

So that really threw me for a loop since I do pretty much everything with Claude Code. And so working in the healthcare industry, I was like, oh, this could be a major hurdle here. However, Anthropic's own page, once I looked into it, which took two minutes to check, said that they will sign BAAs without requiring those huge financial commitment. So that was a clarification piece. I had just posted that on LinkedIn and a bunch of other platforms and it absolutely blew up. But yeah, why spend another episode on a piece of paperwork here? Because clearly, this is an area that's engendering a whole bunch of debate and ongoing compliance issues and trying to navigate how we work with these new AI tools, where privacy is still obviously top-notch concern and heavily regulated, especially in the healthcare space.

And recall, this is Serious AI with Chris Duff. I'm an AI builder with a medical background, which is why I'm focusing on the healthcare industry, and I've spent 10 plus years directing engineering teams, and so it was a natural pairing to become an expert with Claude Code and develop a mid-market enterprise AI implementation firm. So that's what I'm commenting on throughout this podcast and ongoing here. And just to note, everything today that I'm going to comment on is what I found reading the vendor's own pages. So treat it as a builder's read here.

It's not legal advice. Always have to put those disclaimers out there. But again, critical on why this matters again here is because AI is everywhere in every industry now. And in healthcare, it's already playing a role in patient visits. 83% of medical groups as of this year say that AI is playing some type of role with patient visits. And that's up from 71% last year. This is from MGMA's own data and they have something like 60,000 medical providers as part of their network around the US and per their stats, that's like half of all US healthcare is provided through MGMAs.

So pretty heady data from that side. So, you know, this BAA question sits under almost every practice using it, you know, whether anyone has asked, you know, about it or not within individual practices. And you know, again, less than a month of having started posting content through my Serious AI brand, this has become the biggest response to any professional posts that I've in my career so far on a topic that I would not have picked like if you just gave me all the various topics and reels that I was going to post even over this past month and like hey predict which one's going to blow up the most I would not have picked that one but that's why it's kind of a lesson on what we can pull out here and I'm going to dive into some of those you know pieces so here's the shape of that today and think of today as kind of a loop within three arcs here The post itself, what broke through, the thread, so what strangers knew that I didn't and helped inform my own opinions and underlying data and some new relationships that got inspired just from that thread alone.

And the follow through, like what I did with this, which feeds right back into this piece content that I'm recording here today and the following posts here. So starting with the post itself, commented at the start of this is that my piece of content that I recorded last week, which was the Claude Code build that I'm arguably proudest of so far, which is a public template that keeps all of your Claude skills and routines in one GitHub repository. So it auto updates skills and routines constantly. You never have to re-upload anything again. That was just a major issue of using Claude or ChatGPT and running into old versions of files and so forth became a huge mess.

So like on the call is like, yeah, this is, this is probably the best thing I've ever built. But so far, it's drawn basically nothing in terms of response, like zero stars on GitHub, zero forks, five page views at this point, like a week after, roughly. But like I mentioned, this compliance question took off on LinkedIn, almost 10,000 impressions and over 40 comments as of recording today. And you know, typically I'm getting like 44 to 89 impressions. And so, like, my read on why, you know, if I'm diving back into this now, like, why this post blew up so much is because, you know, Anthropic's page for BAA was like a page anybody could check.

And it was a question that any reader could answer from their own work. Whereas a demo, you know, gives people something to watch, right, and implement on their own, but it doesn't necessarily, you know, require additional engagement. But, again, like the BAA question gave them something to answer and is clearly a zeitgeist-y type of question and topic that a lot of folks are thinking about. So, you know, the demo is going to stay here, and I want to do more of those here. And even if I tried to make that one much simpler to go through, again, a lot of AI builds are still above the head of a lot of folks here.

I have to remember, yeah, this is a pretty nascent industry and operating on the frontier here. I'm sure one will eventually break through anyway with more practice and reps on my side Generally volume is what's going to win here And yeah, i'm going to judge that Response over, you know many months rather than just a week and i've like just started content posting for this brand Just over a month ago roughly So you know very minimal data to work with and yet like we're already finding some really solid outliers posts. I was remarking to one of my masterminds earlier this week is that I'm probably 5% optimized on marketing so far.

There's so much else I can do here. So just starting to figure out, okay, where is this data actually going to end up? But back to the post itself regarding the BAA, you know, the comments did something I didn't quite expect. The vast majority of folks who are commenting on this post I did were people I'd never met, not even connected to me. And they started debating with each other. Before I'd written even a single reply, like I was coming back to this thing, oh my gosh, like is just really surprising and also like just it fired me up right like I've again I've never had a professional post that performed like this before and like one comment alone ended up with eight replies under it you know some of those were mine as I you know got back into engaging with the folks who came to say some items and it was it was great to see too because there was really no you know trolls or anything maybe that's you know part and parcel For LinkedIn compared to like extra YouTube comments, for instance That was pretty encouraging to see so it was all value-add comments And all these people commenting were coming at it from different places And you know between them they laid out, you know three doors for keeping patient data safe with AI Roughly here.

So, you know door one is Strip the identifiers before the data ever reaches an AI model So your job, either as a vendor or if you're in charge of patient data that's running through any of these AI models, is checking that redaction for the data. Door two is you could run the model on your own machines, the AI model that is. And so then you won't have to go to one of BAA, and instead you're basically managing the security yourself. So you're still under the regulations of HIPAA, but you run the security yourself instead of having to get third-party BAAs.

Door number three, which I found pretty interesting and something I'm going to test a lot more too, I have it all lined up to do soon, is to run the model inside a cloud where you already have a BAA with such as AWS or Microsoft or Google so then your job becomes configuring that cloud correctly or working with another vendor who's set up a cloud model like that and working through them but again you have to be able to confirm that they're actually doing their work appropriately or have the right guard rails in place so but those are kind of the main three doors to work through here you know Each one works if you do it properly, but each one hands you a different job as well.

And there were different roles here. There was like a healthcare IT architect who suggested redacting before the model. Then there was a healthcare AI builder who said, hey, run the models locally. I mean, probably not a surprise there. There was a founder and a CTO who runs agents inside their own customers or their customers' own cloud under its BAA and like a vendor here, clearly a sales incentive. Was somebody who writes publicly just on hospitals from a reporting standpoint, and they said, yeah, redacting removes the use case of the data because it redacts too much and so you have to be careful on that balance and what's actually still useful or not post redaction.

The thread also found some of the catches in addition to that. So catch one is that you can't redact a voice. So HIPAA's own de-identification list names voice prints. So the first company that hears the visit, the speech-to-text step inside an ambient scribe needs its own BAA. So you can scrub the transcript afterward and the recording is still patient data. So have to make sure that you're covering all pieces of data that might have some patient info in it. Catch number two related to that is that a BAA can cover one vendor, but the data may run through several different ones.

So for example, Anthropic will sign for a HIPAA ready Claude Enterprise account set up and parts of its own API, But its own page lists connectors like MCP servers or Claude in Chrome and Skills is outside the BAA. So I mean, whose Claude setup doesn't have a bunch of MCP connectors and so forth. So, you know, you're going to need coverage for each of those connectors as well, or just really make sure you're not having any patient health data going through any of those connectors that are exposed. So like there was a physician founder commenting in thread like okay well how do you verify every connector before AI goes near a clinical workflow.

So you know open question there that some folks are again trying to work through in real time. Catch number three is that the paperwork says nothing about what agent says out loud. So for example a psychologist building a patient engagement AI had a call agent read a patient's full date of birth back on a call so no BAA would have caught that necessarily. Again all these are like of related here insofar as it's like, okay, yeah, where are each of these points where like anything related to patient health info might get exposed where all these catches that need to be caught so you know and i had mentioned you know back for that catch number three from the thread that you could just include a short list of things that the a agent should never say and a regular pull of transcripts to you know check that it actually held and didn't say anything and like this is a real problem here too is There was, in November 2025, again, AI is way worse than it is now and not probably as widely distributed, but there was a patient that sued a San Diego medical group roughly a year ago, alleging an ambient AI tool recorded visits without consent and sent the audio to a third-party vendor.

So, you know, there's real liability that practices are facing here. And you know, it was just that there was a comment on the thread as well, too, where founder of an AI company just responded, hey, says who in relation to BAA costing millions of dollars from Anthropic here, which was the whole point of the post is, yeah, you have to ask. And I've always referred to the Latin phrase nullius in verba, credit to Naval for popularizing that, take no one's word for it. You know, what did I do with all of this, you know, this really interesting post that was absolutely blowing up tons of comments on it is the actual follow through here.

So, you know, one, me being, you know, an AI builder here and just trying to create engaged, or, you know, engaging content routinely, and also wanting to build relationships from it. So I used three habits here. One was just to answer every comment with something new, like if the comment deserved, it was just like, hey, a thumbs up, whatever, fine. And I'd probably like over comment on some of these earlier blow up posts versus later on when things might become more routine, but can You never want to take anything for granted. And you want to see authors that participate.

It's always going to be more exciting for folks to engage. And so anytime I had additional value to add, I would suggest a fix or a view or a next step the person didn't already have. So if some of them were asking for questions or misconceptions and so forth, then yeah, let me comment on that. Two is to log every angle along with the comment that raised it. So this is part of how I built this post on, OK, here's all the catches related to setting up BAAs with HIPAA regulations and how folks are going about tackling this live issue here. And so nine angles alone came out of one thread.

So this is just interesting both from a business opportunity perspective, additional content creation, and just additional value add to the industry here. And three was to turn the offers into real work. So there were two AI founders who offered to show me their single BAA setup where you're using a cloud that's governed by one BAA. So that's another piece of content that I actually do and potentially recommend to other health care practices if it works well. So like now I would have never uncovered what these companies were but you know because of this post they came to light and so now I can run through a few practice workflows on them on dummy data of course and you know put the setup side by side and so that's additional value that can be back to the industry here.

So the lesson here is that like the comments are the asset. So one post gave me nine angles, two potential partner conversations, this episode that I'm recording right now and a next few posts here. So it's just a constant content engine that can start running here. Here's one thing to do this week if you're in the healthcare industry pick one AI tool your practice uses and figure out each and every piece of technology that patient data goes into and out of, whether that's the microphone, transcription tool, an AI model itself, every single connector.

You have to have everything mapped out to really understand what could really be exposed and what needs to be governed by BAAs or your own security measures. And if you are using BAAs, is like really looking at those underlying agreements with any vendors that you might be working for, or working with rather, and making sure that there's real coverage for each of those areas that your PHI is being exposed to. And of course, on the other side, if you are a vendor or you're building or selling some of these tools, run the same exercise from the other side, because you are facing very real exposure as well, too, where more and more regulators are going to be paying attention.

And just from the content creation side of things, read your own comments closely, because that's where your next ideas are. So, yeah, this is Serious AI with Chris Duff. If you're interested in seeing the weekly newsletter, it goes out every Monday morning at seriousai.partners/newsletter, 100% free. And if you are a mid-market enterprise with a top-line revenue of $5 million or more, and want to potentially work with us, feel free to reach out on our website seriousai.partners. Healthcare is the primary vertical here, but we're looking at any industry that fits within that band.

So looking forward to next time, subscribe and share everybody. Take care now. Bye.