Episode 58: HIPAA-Compliant Intelligence – Building a Privacy-First AI Infrastructure
Hosted by Aaron Burnett with Taylor Hurff
How can healthcare and MedTech marketing leaders build scalable AI capabilities without risking client privacy or work quality?
In this episode of The Digital Clinic, host Aaron Burnett sits down with Taylor Hurff, Director of Digital Strategy at Wheelhouse. Together, they explore how to transition from informal prompt tinkering to a HIPAA-compliant, privacy-first AI ecosystem that elevates senior strategy and keeps human discernment at the center.
Listen & Subscribe:
Introducing Taylor Hurff
Aaron Burnett: Today, I am joined by Taylor Hurff, Director of Digital Strategy here at Wheelhouse Digital Marketing Group. Many agencies are treating generative AI as a shortcut to cutting costs or increasing work output. But our team at Wheelhouse has taken a very different path. Over the past year, Taylor’s led the charge in building highly tuned, privacy-first AI systems and workflows that have genuinely elevated how we deliver strategic outcomes for our clients.
In this conversation, Taylor and I unpack what it takes to move past basic prompt engineering and get to the point of building enterprise-grade AI capabilities, including how to audit organizational AI usage and create centrally deployed and governed workflows, how to use private LLM instances with a HIPAA-compliant data warehouse, and why we believe the goal of AI should be to enable deeper work, not faster or higher volume output.
This conversation is a practical look at how to build an AI ecosystem that delivers real speed, precision, and business value without compromising quality or client data. Let’s get to it.
A Message from the Sponsor
Aaron Burnett: This podcast is sponsored by Wheelhouse Digital Marketing Group. Wheelhouse provides exceptional performance marketing for healthcare and medical device manufacturers. Every Wheelhouse client saw record performance in 2025 even after implementing HIPAA-compliant data solutions. Find out more at wheelhousedmg.com.
How Should a Marketing Agency Use Generative AI?
Aaron Burnett: Taylor, so it’s been a few months since you and I met and talked about AI on the podcast, which in the world of AI is the equivalent of about 20 years. Lots has changed, right? So I think the last time we chatted, we talked a lot about futures.
We talked about where you thought AI was going. We talked about some of the experimentation that we were doing at Wheelhouse. But I know that over the past year we developed a really clear philosophy around AI, and we’ve put that philosophy into action, and you have helped to lead that. Let’s start here. What is our philosophy? What is our approach to using AI? I think it’s an important question because our approach is quite different than other agencies.
Taylor Hurff: So when we use it internally, the biggest things are: how can we deliver more value for our clients faster than we otherwise could? But it’s important that we’re not delivering AI outputs whole cloth. We’re not taking the human out of the loop. We’re not replacing smart, strategic people with AI.
We’re using it to enable us to work faster, to work better, to do more for our clients, to drive better outcomes. But we’re not replacing intelligent people, I think, is the biggest thing. You’ll never see us deliver an AI output whole cloth. You’ll never see us replace smart thinking and that discernment that we all bring to the table.
It’s really meant to be a tool that helps us move quicker, move smarter, catch things we might not have caught before. But, yeah, AI is not delivering services for us. It’s more enabling us to move quicker and do better work.
Aaron Burnett: I think that’s a really important point. I was asked by a senior exec last week how we were reducing expenses using AI, and I was flummoxed by the question, because that’s not at all what we’re trying to do. If anything, we’ve increased expenses with AI so that we can deliver more. In fact, we’ve increased expenses quite considerably.
Taylor Hurff: I think it’s important to know that we’re not trying to build AI capabilities to manage them with more junior employees. I feel like a lot of places might be standing up things where you can have someone who doesn’t really know what they’re doing run an analysis start to finish and pass it off.
We really want AI supporting our most senior, most capable strategists. So I think that’s especially important when you think of using AI in your day-to-day. If you ask ChatGPT about something you know nothing about, you’re going to be blown away by the answer. It’s going to look perfect. If you ask it something that you are truly an expert in, you’re going to see all the things it missed or all of the assumptions it made.
That’s the most important piece when we are delivering things to our clients that are going to have a meaningful impact. I need smart, capable people who are able to spot where it went wrong. We don’t want just junior employees passing it off not knowing better. It’s that discernment, that human in the loop, that really makes AI work for you at the end of the day, not just appear like it’s working for you.
Aaron Burnett: Yeah. Yeah, I think that’s well said.
How to Move From Prompt Engineering to AI Workflows
Aaron Burnett: You know, when ChatGPT became available, and then when Claude and other LLMs became available, we took a very active and encouraging approach with employees. Got accounts for folks, paid accounts, so that people could experiment to their heart’s content, with two constraints.
One was no client data, and the other was don’t create any client work. Experiment, figure out how to use it, but these are the guardrails to ensure that we’re safeguarding the quality of our work and the sanctity of client data. A year ago, you began to lead and, in fact, created a vision for an approach to AI that really does, and has, become a part of the way that we deliver for our clients. Describe where we were when that started, and then the approach that you’ve taken and where we are today.
Taylor Hurff: In the beginning, it was like you said: a lot of people tinkering, trying out things. But even at that time, it was a lot of, I guess, prompt engineering. It was a lot of people just trying better ways to prompt an AI to get a better output, maybe broaching some, like, more innovative thinking.
But it wasn’t using AI to its fullest capabilities, and it really wasn’t having it do client-level work or data analysis, because we didn’t really trust it then. Flash forward to today: we’re able to conduct really detailed, really high-level workflows that are analyzing data deeper than we ever could, doing more than we ever could, and really getting us closer to those final outputs, at which point our strategists step in.
It’s been amazing to see. I think the biggest catalyst, too, was enabling the teams. So it’s not a matter of having one person who’s a visionary at your company trying to build stuff they think people will use. It was being able to bring everyone along throughout the process so that they can tinker, they can come up with new ideas, analyses, ways to use it. But we’ve really come a long way by enabling the team and then setting bigger processes in place to make sure we’re not going off the rails or spending time where we shouldn’t.
How to Audit and Standardize 100+ Claude Projects Across a Team
Aaron Burnett: Let’s jump back and talk about where we started. My recollection is that we had something like 110 different Claude projects. And, by the way, we as a company decided that we were going to standardize on Claude for a number of reasons: we had higher-quality output; we also felt more ethically comfortable with Anthropic than we did with OpenAI. And so we standardized on Claude. And when you became involved and started to lead, we had more than 100 projects. How did you approach that? How did you rationalize, and what processes did you put in place that got us to where we are today?
Taylor Hurff: Yeah, that’s a good point. So when we were just letting people go off and build, it was good for getting ideas, but it wasn’t good for standardizing things or making sure that everything they built was working the best it could. What we realized when we looked at, I did that inventory of all hundred-plus projects, people were approaching them in so many different ways, not following simple best practices, not pressure testing them, or even maintaining them over time.
We had some that were built a year-plus prior and hadn’t been maintained as models launched or as new best practices came about. One of the things we realized we needed, beyond just a cleanup of what we had, was standards for how to set up a Claude project so it will work reliably, and so that as new models launch, we can pressure test it, benchmark it, and make sure it still works for us.
So now we still let people go off and build what they want to build, but when they have something there, they share it through a centralized way. We review it, we audit it, we make sure it’s up to our standards, then we inventory it and make sure anyone who needs it gets it. During that process, we found a ton of interesting projects that very few people had insight into, and unless you’re passing it around to individuals or posting it in a company channel, not everyone’s getting the benefit of the things you’re building.
So that’s one of the processes we have: we make sure it’s reliable, we make sure it’s the best it can be, but then we make sure it’s stored in a way where everyone has access to it, and we make sure it gets into the hands of the people that need it.
Building an AI Roadmap: How AI Cut Presentation Prep Time by 70%
Aaron Burnett: Coming out of that audit, you developed a roadmap, right? An AI roadmap for the company that was really, effectively, an 18-month roadmap: quarter by quarter, what would be built. Talk me through the prioritization and some of the things that have been built in the ensuing year-plus since we launched that roadmap.
Taylor Hurff: When we started that roadmap, we already had a handful of ideas of things people said they could use. But really, when we started out, we went to the teams and asked, one, like, “What do you spend the most time doing that AI could potentially help save time on? What are things that you don’t do that would drive significant value, that are just too time-consuming or too difficult, that we don’t spend the time we should on?”
But essentially: where are we wasting time, and where could we deliver more value? And from there, we identified dozens of really important workflows that we could build, Claude projects and skills and automated workflows, and then prioritized them based on time savings and value added for our clients.
One of the biggest ones, we didn’t expect this when we went in, was even just building presentations. So we have a number of analyses we run, different deliverables we do, but then it takes very senior people to come together and put that into a deliverable format that’s going to resonate with the client, that has a strong narrative, that is very compelling.
It gets the buy-in you need it to have. That’s not the best use of time for your most senior strategists. What we were able to do pretty quickly was stand up a process that would intake all of the findings, analyses, and recommendations, and create a very compelling narrative, a very compelling outline, and then give you a first draft of the slides.
But essentially, it’s saving seventy percent of the time you would be spending, getting you to an almost-finished presentation that your strategic people can then come in and take over the top. That was one of the biggest time savers we had, and it didn’t really tie to a new capability. It was almost strictly time savings. But as you can see, the outputs we’re delivering now are so much better and so much faster, and it’s getting us buy-in quicker, which is in turn creating better results for our clients, all just by saving us more time on the things that don’t require your smartest, most capable people to be spending time there, when they could be digging into data, coming up with recommendations or strategic pivots.
How Automated AI Analysis Improves Paid Media Performance
Aaron Burnett: What have you and your team developed that has delivered the greatest value?
Taylor Hurff: So we do a lot of paid media, and analyzing our paid media data at scale is time-consuming: being able to drill down by landing page, by channel, by segment, seeing what’s working and what’s not working, and using that to inform bidding adjustments or targeting adjustments. Being able to stand up a workflow that is working in the background, it doesn’t require hands on keyboard to run it. It’s an automated analysis that looks at every dimension you could think of and surfaces what’s working, what’s not working, where to pivot and where not. And putting that into the hands of our strategists overnight lets us move way quicker and have much better insight into those strategic pivots.
Essentially, we’re able to run these workflows in the background, so I wake up to a report of: this is what’s going on, here’s what’s changed, here’s where you need to dig in further. So then I can hit the ground running and come up with really impactful recommendations for our clients. I think that’s probably the most impactful thing: highly detailed analyses that are informing our paid strategies at a clip that you couldn’t really do manually, at least not without a massive team of data analysts.
How to Use Claude in Chrome for Landing Page CRO Analysis
Aaron Burnett: I know that of the things you’ve built, one is a landing page analyst, looking at the performance of landing pages in the context of organic or paid, and looking specifically for conversion rate optimization opportunities. Can you describe that workflow and what it’s delivering?
Taylor Hurff: Yeah. So it starts with what I described: finding out what’s working and not working by channel. But there are a number of different frameworks you can use to do a landing page analysis. You can do a LIFT analysis, heuristic, there are a number of different ones. We were able to codify that into a Claude project that goes out and assesses, you know, where are the distracting elements, where are the value propositions, are they front and center, are they not?
And then it gives us recommendations for how to better message to our visitors. It’s not something we deliver, you know, as is to a client, but being able to say, “We ran these structured analyses; here are the things that you’re doing and not doing. Here’s a graded rubric of, you know, where you’re doing well, where you’re not, and what might move the needle,” gets our strategists a lot of really great thought starters instantly. It’s able to save us hours and hours per landing page, and we can run it at scale across every landing page where we’re driving traffic.
That’s it at a high level. I could get into the specifics of how we actually have Claude look at the pages, how we put in guardrails so it’s not just… We don’t just say, “Do a LIFT analysis on this page.” We have very detailed instructions for what counts as distraction, value prop, urgency, things like that.
But being able to have a high level of control over how Claude analyzes, to do it our way, and then giving it to a very senior person who can pick out what is and isn’t important, or where it might have missed, it’s really expedited things on our end and caught a lot of things we probably would not have caught otherwise.
Aaron Burnett: So I think it would be worth digging into how this is built and how it runs. One of the things that I experience when I listen to people talk about the value of AI is that it ends up being very arm-wavy: we built this thing, it delivers fantastic results. It’s just “a workflow,” and you have no idea what went into that workflow, or if it’s even a workflow, or if someone just wrote a really long prompt and uploaded some good knowledge. So maybe describe what it took to build this particular workflow and how we ensure quality output.
Taylor Hurff: Yeah. So the first step of building this was taking those analyses and, I say codifying them, but basically, step by step, giving instructions on what to look at and what counts as different elements of the analysis, and being able to give clear direction on: when we say we need a sense of urgency on this landing page, what does that mean?
What are the signals it’s looking for? So Claude, when we start one of these, will align with us on, I guess, the rubric that it’s going to judge it against, which is different from client to client. Someone who is in a lead gen medical device space is going to be much different than an urgent care landing page.
And so we first tailor the analysis to each individual client, and then it goes out. We use Claude in Chrome to feed that tailored rubric to Claude on the landing pages that we’re going to be optimizing or digging into, and it’s going to view the actual page across different devices and give us that detailed readout.
It’s going to look at every factor that we said is important. It’s viewing the actual pages and doing the analysis together and giving us the output. But essentially, it’s not a one-size-fits-all workflow. It tailors it to each client, with a strategist there guiding it to make sure the analysis it does truly makes sense for that client. And then Claude in Chrome is able to fully see the pages and analyze them without us guiding it throughout. So it’s getting that tailored. Essentially, it’s curating a very customized prompt for the situation and then running it in Chrome to view the page, deliver the output back to our Claude conversation, and analyze it against the criteria we had in the first place.
I do think the best workflows we have leverage different capabilities, whether it’s data connectors or something like Claude in Chrome, or connecting to our Compass warehouse. There’s a lot you can do out of the box. If you’re stuck with just really fancy prompt engineering, you’re missing so much of the possibility.
What Does a Privacy-First, HIPAA-Compliant AI Ecosystem Look Like?
Aaron Burnett: So we’ve talked about a few workflows. Give us a sense of the AI landscape at Wheelhouse today: the number of workflows we have, skills, governance.
Taylor Hurff: Yeah. So I think I did a tally earlier. We’re at 41 projects and automated workflows that are tailor-built to things that we either would’ve been doing or would’ve liked to have done, but found a possibility to expand our capabilities. So, 41 workflows and projects, and eight organization-wide Claude skills, with more per team.
So there are skills that, when you’re doing SEO, you might need, but digital advertising, marketing, all those other teams don’t need. But we have eight organization-wide skills that everyone needs and has enabled. And then, currently, six MCP servers that are connecting to all the different platforms we use most often and our data warehouse, plus a series of additional connectors.
Our Claude instance, which we mentioned before, on the last podcast, is a private instance of Claude. We’re not feeding data back to Anthropic, but it’s so much more powerful than I think most people would think when they hear that we’re using Claude. It’s connected to our Slack conversations.
It’s connected to ClickUp, which we use to manage all of our projects and store information about the clients. It’s connected to SharePoint, where all of our documents live. It’s connected to Otter, which has our call transcripts and meeting notes, and a number of other places. So essentially, I can go into Claude now and say, “What is my team working on this week?”
As simple as that. It knows my team, it knows where to look, it knows the ClickUp boards to look at. It knows the Slack conversations. I can really ask it what I want to and get something back. But then we also have tailored workflows to make sure it does it exactly how we want. So that might be a good example.
If I wanted to get a really detailed look at the next six months of my team’s work, I could just put it into a chat and trust that Claude’s going to find it in the right place. I also have a tailored project that tells it exactly where to look, exactly where to double-check to make sure that Claude is doing it how I want it to do it.
But, yeah, a series of those workflows, a handful are those always-on scheduled analyses. So, as an example, a content gap analysis. Before AI came out, every agency I was at would do a content gap, but they might do it every six months or once a year. We’re able to run that weekly and surface any new opportunities that come up, because it’s scheduled; it’s in the background.
It doesn’t take hands on keyboard to get that going. We run that in tandem with a series of other analyses that we would like to run as often as possible to surface any new opportunities for our clients. So that’s where we’re headed: not one-off projects where I have to be at the helm. It’s how can we package these up so they all run in tandem and then surface the biggest opportunities to us.
Aaron Burnett: There are a few important things to emphasize here. So one is that, as I mentioned, a year ago we had a hundred-plus individual Claude projects. We weren’t using client data. We weren’t producing client work. It was just a lot of experimentation so we could all figure out what was possible, what was good, and what wasn’t.
Today, as we’ve integrated AI properly into our business, we, A, have, as you mentioned, a privacy-first environment. So it’s a private instance of Claude that doesn’t share data back. The connections are private as well, and our data warehouse is a HIPAA-compliant data warehouse. And so data privacy is sacrosanct in what we do.
The second thing that’s important to understand is that the workflows that we have today all have to be submitted for approval to you and your team, and they have to be vetted. And then they are, in many instances, tuned and enhanced to ensure that they conform with data privacy, brand guidelines, style guidelines, data accuracy, skills, that sort of thing. So we ensure we’re not putting workflows into the environment that will hallucinate or create content or designs that are inconsistent with the quality or the ethos that we want to put out there.
How to Govern, Benchmark, and Pressure Test AI Workflows
Aaron Burnett: Talk with me about that governance aspect. We’ve kind of talked through the approval process, but how do we ensure consistency of output, consistency of quality, and accuracy?
Taylor Hurff: It’s really three main steps, apart from centralizing this review process as a whole and making sure things all come through me and my team. The first is reviewing for best practices. I say best practices, but even things like making sure instructions are in markdown, and that we’re not doing two thousand lines of instructions where we know Claude will forget pieces of it or ignore different pieces.
But making sure it’s set up in a way that we know common failures won’t occur. But then the biggest two things are benchmarking and pressure testing. So we come up with standardized prompts that really push the limits and make sure it does everything it’s promising to do, and then we grade it against that. So this also feeds into how we choose what model for what project. We check Haiku versus Sonnet versus Opus, and we say, “These are the things it needs to be able to accomplish.” It does that. We review the outputs against this graded system. This is benchmarking, and we see: is it effective? Is it missing anything?
Do we have to revisit instructions or clarify anything? Is there a skill that’s missing that needs to be included? And then pressure testing from there. So we have our standardized prompts, but we really try to push the limits because we know once it’s in the wild, people are going to be pushing the limits.
But it’s making sure you have that process in place, so it’s not just getting submitted to me, I log it, I let people use it, without knowing through and through that it’s going to do what we need it to do, and do it within the constraints of our data privacy, making sure it’s not going to go off and use a tool that is not approved, or making sure it wasn’t set up in that way in the first place. Really, centralizing that review process and thoroughly vetting everything before it goes out is the key to having a sustainable, secure inventory of projects and skills and workflows.
Aaron Burnett: I think the other thing that’s been important is that because we have an organizational account, when we have skills shared at an organizational level and we refine or improve those skills, that becomes common to the organization. We don’t have to push anything out individually.
Taylor Hurff: Yeah, exactly. That’s been one of the biggest things, too: making sure everyone gets the skills, the projects, everything they need. And we already talked about that, but the biggest miss historically for us had been that someone builds something fantastic and no one else knows about it, and then someone else builds a very similar thing.
We’re doing double work. We really solved that by centralizing it, and then also just making sure things are consistent, because there’s nothing worse than trusting AI in a certain way and it fails you, and then you spend more time than you would have otherwise going back and trying to refine it. But pressure testing and benchmarking are huge—especially, Claude just rolled out the 5.5 updates to its models. We had to go and benchmark again, make sure they’re doing what they intend to do. Luckily, we didn’t have many issues. These models are great, but in the previous version of our AI infrastructure, that wouldn’t have been a check we had, and we wouldn’t know if things were as reliable as they were when they were built.
How Does an AI-Powered Technical SEO Audit Work?
Aaron Burnett: Tell me about the technical SEO audit workflow. What do you get from that workflow that is better than you might get from a manual audit process? And then, secondarily, what do you get from being able to run such a rigorous analysis more frequently?
Taylor Hurff: So the biggest thing with this: it starts with a Screaming Frog crawl, which a lot of technical analyses do, but it also connects other data sources, like Google Search Console. As any SEO knows, I mean, there are the common flags, 404s, broken links, stuff like that, that Screaming Frog can flag to you.
And you could have a simple technical audit that just connects to your crawl, pulls out the findings, and puts it into a document. We take it up a level. We are able to run a series of different analyses using that same crawl, using Search Console data, and even connecting to Ahrefs and other sources, which lets us do a series of analyses in one.
So, one of them: we use vector embeddings to do an internal linking audit. That in itself would take hours to do manually. Even using Screaming Frog to get the output, we’re able to, through Claude, connect to that crawl, conduct a vector embeddings analysis, conduct a content gap analysis, internal linking recommendations, cannibalization.
All of these different analyses that before we would have scoped out separately, we would have scheduled out throughout the course of a year, we’re able to run that all in one go, and run it weekly, to be able to find new broken links, new 404s. The client publishes a new page, doesn’t connect it to the other content that it should, or it cannibalizes something, we’re able to catch that immediately.
Whereas before, we would have just caught it the next time we did a cannibalization audit. And we’re able to get that all run in the cloud, so it’s happening automatically. It doesn’t take one of our strategists opening his computer, spending significant bandwidth just to have Screaming Frog running, then prompting all of these analyses. It’s happening in the background, in the cloud, and getting delivered to his inbox every Monday.
How to Use AI for Competitive Monitoring in Search and Paid Social
Aaron Burnett: One of the other things you’re working on is competitive monitoring. Tell me a little bit about that.
Taylor Hurff: Yeah. So that’s an interesting one. Having a close eye on the competitive space is incredibly important. As anyone in SEO knows, if your clicks are dropping, it might be because there was an algorithm update and you’re ranking lower. It could be because competitors are publishing new content, they’ve pivoted strategies, they’re starting to eat into your market share.
But also with paid search, as well, and paid social. Having a close eye on competitive strategies, competitive messaging, and the target audiences they’re going after is hugely important, especially in medical device spaces, which are incredibly competitive. What we’re able to do is closely monitor the search landscape.
We’re pulling that data into our warehouse. We’re identifying what ad copy they’re using, what kind of content they’re publishing, where they’re outranking us, or where they’re starting to show up more in ads where we otherwise were getting most of that market share. And then we’re also looking at the Facebook Ads Library and other social ad platforms to identify messaging, new ads going live, new strategic pivots.
Any of that stuff I just described would take a lot of time to collect manually. Before AI, you could set up API connections for some of this, but having this kind of detailed analysis, this kind of finger on the pulse of the competitive space, has been huge for us. It’s able to help us identify their pivots, their shifts, and we can even see when they’re testing a new strategy and then when they double down and start showing up more frequently, putting a certain type of messaging into more ads. We’re able to get really good insights that you can’t get anywhere else. That kind of data is not publicly available. We’re able to collect way more data than we ever had been and analyze it at a clip that lets us quickly pivot or counter when our competitors are making shifts.
How AI Improves Cross-Team Collaboration in a Marketing Agency
Aaron Burnett: That’s great. Thank you. One of the things that sets us apart as an agency is the way that we engage as One Wheelhouse. Regardless of where people have entered the agency, whether it’s for paid or organic or analytics, we will bring to bear all of our services. In what way has what we’re doing with AI either facilitated better cross-practice collaboration or changed and elevated the quality of the conversation that we have in these multidisciplinary conversations?
Taylor Hurff: So our Marketing Sciences team collects tremendous data, and they’ve been able to put that in the hands of everyone at the company through our Compass MCP. That’s been just one way they’ve done that, but now, instead of us having to ask them a question, submit a ticket, or just have a conversation and they get back to us days or weeks later, we’re able to have full access to everything they have and be able to run it through any of these projects that we’ve stood up, any of these workflows.
I’ve been able to get my team doing detailed paid analyses very quickly by standing up a workflow that conducts it. I have to train them on what it is, what to catch, where to dig in deeper. But instead of me having to do hours of training and running them through it over and over again until they’re true experts in it, it’s made that learning curve much quicker.
So I’m able to build out detailed workflows that analyze our paid search terms at scale to identify areas of wasted spend or opportunities for new landing pages. Instead of that being a detailed training for each individual person and taking significant time, I’m able to pass that off to them quickly, and they’re able to pick it up much faster.
But that’s enabled better collaboration between digital strategy and the paid media team, because now, in a matter of minutes, I’m able to get insights into where we are not as efficient as we could be, and is that a landing page issue, or is that something in the platforms? And it’s enabling really great conversations between the two. But I think data access is one of the biggest ones, and then building out workflows that enable us to better collaborate has been the other.
Why Build an MCP Server for Your Marketing Data Warehouse?
Aaron Burnett: You mentioned something really key there that’s part of our sophistication. We built our own MCP for our data warehouse so that all of our analysts can talk to the data warehouse and run their own analysis, and run very sophisticated analysis, without requiring, except in the most extreme instances, a BI analyst.
Taylor Hurff: So when you stand up a data warehouse, oftentimes in the beginning it’s to inform your reporting or whatever capabilities you’re building off it. Like for us, we do MMM or propensity modeling, things like that, that we need this data for, but a lot of it informs reporting. When we started connecting all this data and building out these workflows, we realized there’s a lot of data analysis that we used to do in-platform that we could do much better if that data was in the warehouse. So we’re actually getting team members from across the organization flagging gaps in data: “Hey, we’ve got to collect this. We have to backfill this.” It’s going to unlock so much potential for us that it’s really been the catalyst for enhancing our data collection beyond what you would typically expect for performance reporting and our typical optimizations. We’re able to collect much more granular data that we previously didn’t need in a warehouse, and now it’s able to influence so much of our strategic decision-making.
What’s Next for Automated AI Marketing Workflows?
Aaron Burnett: Where are you headed next? What’s most exciting for you that is coming up on the roadmap?
Taylor Hurff: So we mentioned the technical SEO audit, where we have a series of workflows. What I’m most excited for is building those out for the different verticals we focus on. So I mentioned the paid search terms analysis and the paid landing page analysis. I want to be able to build these out into one automated workflow that looks at paid from every angle, all simultaneously, and gives us a prioritized output.
I don’t want to have forty-one different workflows that we can use at any time. I want to have three or four main ones that run dozens of workflows simultaneously and then prioritize outputs against each other. That’s really where this is going. It’s great to have the opportunity to run one specific one, say, a content gap, when we need just that, but it’s even better to run a content gap every week, and when a big enough opportunity arises, it surfaces to the top.
That’s where this is going. It’s connecting all these together, doing them in tandem, so that instead of just running one thing here and there, it’s running absolutely everything at once and flagging opportunities much quicker than ever before.
What Will AI-Powered Digital Marketing Look Like in a Year?
Aaron Burnett: Things change so quickly in the world of AI and in the world of digital marketing, so I’d be remiss if I didn’t ask you the same sort of forward-looking question I asked last time. What do you think digital marketing at Wheelhouse will look like a year from now, given the anticipated continued evolution and advancement of AI?
Taylor Hurff: I think in general, it gets even more strategic. We’re using AI to free up our time from the pieces that aren’t strategic, that don’t necessarily require a senior person involved. Even running a Screaming Frog crawl and pulling out the key findings, you could train a newer person to configure a crawl in a specific way.
What we’re able to do is save that time, get that time back. Even formatting a document in Wheelhouse branding used to take a lot of time, and now we’re focused strictly on: what is the message we’re conveying, what are the recommendations, where do we go from here? I see us replacing as much, if not all, of the manual, mundane, tedious tasks, so that when we step in, start to finish of our day, it’s that higher-level strategic thinking, guiding businesses where they need to go.
I really feel like that’s where we’re going. I don’t see a world where my team is bored because they’re making slide deck after slide deck, or presentation after… you know, spending time configuring crawls. I really see us spending the entirety of our time being very strategic and being stewards of the business, more so than executing tasks that they don’t need to be spending their time on.
Aaron Burnett: I agree. I think that’s a great vision.
Taylor Hurff: We’re building it.
Aaron Burnett: Yeah.
How MCP Servers Connect Claude to Your Marketing Data and Tools
Aaron Burnett: All right, so we just talked about the MCP that we built to connect to Compass so that our analysts and our strategists can talk to the data warehouse. Let’s talk about the use of MCPs more broadly. How do we employ MCPs at Wheelhouse in a way that gives us a significant advantage over a more conventional use of AI?
Taylor Hurff: The biggest thing we do with these MCPs: any data we would need to conduct an analysis to better inform what we’re doing, we’re able to connect that today into our Claude infrastructure. So, our Profound AI monitoring instance, we’re able to connect that data directly into our Claude, so that when we’re running a content gap, we’re not just looking at, say, Ahrefs data or SERP data from our API.
We’re also able to layer in our AI monitoring data: where our gaps are, where our competitors are beating us. I think the biggest thing is being able to connect as many data sources together as possible, so that we’re not just relying on Claude to ask us for the data. We’re not just using Claude as a chatbot; we’re truly empowering it with as much data as possible.
I wish, so if you ever feel the need to join one of our onboarding calls with our new hires, they’re always blown away with what Claude can do. Like, not just connected to these data sources, but even connected to our Slack, our Outlook, SharePoint, our Otter transcript notes. It goes from me asking Claude, like, “Oh, what are some ideas for a landing page for this kind of business?” To, “Look at the call transcripts from the last ten weeks where this client attended, and give me themes of things they’re asking for, any trends you’re noticing.” It knows to go to the Otter transcripts. It knows who the client is. It knows what our strategies are, what’s important, and what is and isn’t in scope. It’s able to do a very detailed analysis based on things that otherwise live outside of this infrastructure. So we’re able to bring even more into our ecosystem, so our Claude knows who we are, our vision, our clients, our scopes of work, how we operate. I’m missing a lot, but it knows our clients. It knows what we talked about on the last call.
Again, all of this is in our private instance, so none of this is being shared out. But it has everything from our client contacts and what we talk about day in and day out, to our performance reporting and every platform from which we get data. Really, I wish I could just pull it up and show some of the great use cases I show in our trainings, but that’s the biggest differentiator for us internally: it knows everything you would expect it to know.
Aaron Burnett: Yeah. I think in my experience, right, there’s tremendous power on the end of the continuum that has to do with very sophisticated analysis and nuance and strategy. But in a more mundane sense, it also saves a lot of time for me and for others, in that there are lots of instances where I need to know the answer to a question, or I need to find an asset that I haven’t looked at for seven months and I can’t remember where it’s filed, because I don’t work with that client regularly.
These are things that I historically would’ve interrupted other people to find out, and now I just ask Wheelhouse in Claude, and I know: oh, that file’s here, and the context around it is this, and here is its history, and here’s some other information. So I find I’m much less of a bother to other people in Wheelhouse, because I can use Claude to answer those more mundane questions.
Taylor Hurff: That’s very true. And the last piece of the process I’ll touch on is with every new hire, we not only do a detailed training of all the capabilities, but we make sure all of their settings are set up in a way where they’ll get the most out of it. So it’s not just: we add them to Claude, go figure it out.
We make sure they’re fully enabled with everything that they need, and then we walk through every connector, every setting, every capability, the projects, and make sure they’re fully equipped. And it’s so great to see their reactions. Everyone at Wheelhouse is so excited to have these capabilities, whereas before they got here, AI was just what you’d expect, like ChatGPT is in your regular life.
You can ask it questions, you can do some nifty stuff, but it’s so much more here.
Aaron Burnett: Yeah, I agree. Thank you, Taylor.
Taylor Hurff: Yeah, of course. This was fun.
Key Takeaways: Building a Privacy-First AI Strategy
Aaron Burnett: As Taylor shared today, truly integrating AI into a marketing organization requires building secure infrastructure, establishing rigorous benchmarking and quality controls, and developing workflows that enable your senior team to focus on deep strategic work. You can connect with Taylor on LinkedIn to follow his work in digital strategy and AI architecture.
And if you enjoyed today’s discussion, be sure to subscribe to the podcast, leave us a review, and share this episode with anyone on your team looking to build a privacy-first AI strategy. I’m Aaron Burnett, and I’ll see you next time on The Digital Clinic.






