How AI Handled a Biotech Support Overload

Eric Housh and Zack Terry cover how a high-growth biotech firm handled a corporate split that left their support team drowning in repetitive tickets about shipping and results. FSM helped them stand up a new Salesforce Service Cloud instance and build an AgentForce AI agent grounded entirely in a newly created knowledge base.

They dig into how retrieval augmented generation and vector search let the agent answer only from vetted business data, when it escalates a sensitive question to a human, and how a 40% deflection rate changed the team's capacity from day one.

Useful for any regulated or customer-support-heavy business weighing how to bring AI into a system without sacrificing accuracy.

For the full episode video, show notes, and related resources, visit: https://loom.ly/iFQSAjw

Transcript

SPEAKER_01 0:09

AI in Action is brought to you by Fast Slow Motion. Our team helps growing businesses put AI to work with practical, scalable solutions. To learn more about how we can help you implement AI in your business, visit Fastslowmotion.comslash AI. Welcome back to AI in Action. I'm Eric Hush, joined as always by my friend Zach Terry, director of AI over here at Fast Slow Motion. Zach, we've got a pretty cool use case today. This is a high growth biotech firm. They went through this massive corporate split. They're standing up a brand new operation. Suddenly, their inbox is drowning in the same 50 questions about shipping kits and DNA results and things like that. So that mountain of repetitive tickets that they're staring at while trying to build a department from scratch resulted in just a huge disconnect between their need to scale and actually having the people to do the work.

SPEAKER_00 1:03

It was a huge problem. Absolutely. And I think I think this is something that a lot of organizations can relate to, especially anyone that's ever done a migration, and then they're having to deal with all the work that's required just to get their data set up. And then in the meantime, they're working on the old system while they're building the new system and people are having to handle tickets on the old system, but you still need the data to get over. And so, yes, it's it's a pain point that many have felt. And in a highly regulated space, you know, this is a healthcare industry and it's it's almost healthcare adjacent because it's it's consumer biotechnology, right? You you've got to have some trust there too. You want to make sure that when people are actually getting answers to their questions, that it is accurate and it is actually reflecting the tests that you are offering and how those results should be interpreted. And so, you know, when you start thinking about having AI handling that volume, you really have to have a very, very low tolerance for error. Exactly. And it's kind of where the move fast and break things startup mentality kind of breaks down, right? You can't you can't break too many things when you're talking about a regulated industry. And so um, yeah, it's it's really just a balancing act of efficiency versus sensitivity and accuracy.

SPEAKER_01 2:16

So take us to the before state with these guys. What did that day-to-day struggle look like before we came in uh with some Salesforce AI solutions?

SPEAKER_00 2:25

So they were essentially starting from a blank slate because they were having to migrate from an existing platform over to a new platform. Now, in this case, we're talking about Salesforce, we're talking about Service Cloud, and then we're talking about Agent Force as the AI solution. But after this sort of organizational split, they had to figure out how to get that data from this prior organization into their own organization. And so we really helped them with two things. First, we helped them stand up a brand new service cloud instance. And along with that came a knowledge base. And so not only are we moving their entire service team into Salesforce so that they can handle tickets and communicate with customers, but we're also creating a brand new knowledge base on top of Salesforce so that they have the documentation to be able to answer those questions, which naturally moves into automating that entire first touch with the customer, which in this case is setting up an AI agent that has access to that knowledge base and has the ability to not only answer questions from the knowledge base, but also understand when a question is a little too sensitive for an AI to answer, maybe understand when it doesn't have the answer to a question. And so at that point it can escalate it to a human being. And so in this case, we're really kind of talking about three things support logjam, right? They had a huge volume of these very low level one sort of inquiries like where is my kit and how long is the shipping time and how do I understand these results? All these things were hitting human beings, preventing them from handling some of the more complex needs or being able to go in and answer questions that truly required a higher personal touch. Then we've got this knowledge fragmentation, right? So when we're migrating to this new system, there is no knowledge base. And so we have to address this problem before we can have AI assist with it, because we have to have that data foundation set up. So that second part there is not only getting their support team into the system, setting up an architecture to support their ability to actually support customers, but also helping them set up this knowledge base and then getting AI connected to it. But not only connected to it, but making sure that we're connecting it in a way that all that data, all the AI responses are grounded within their specific knowledge. And so it's writing these knowledge articles in the right format, in a format that's conducive to having an AI system analyze them and produce results because we we can't have a high degree of hallucinations here, right? We we need this to be able to respond that it doesn't know. It needs to be able to identify when to escalate to a customer. And, you know, of course, we know this, we've talked about this. You can't eliminate that hallucination risk entirely. That's sort of inherent to the technology today. But there are things we can do to design these systems in a way to minimize that, grounding it just on that data, limiting it just to the data that's specific to the business, uh, and ensuring that it is formatted in a way that is conducive to retrieval. So that's essentially the the three pillars, which is getting their support team set up on a brand new system, setting up that architecture, getting those knowledge articles created, that knowledge base in a state that is ready to be leveraged by AI, and then setting up the AI tool to actually be able to answer questions specific to that knowledge base.

SPEAKER_01 5:26

I love it. Uh let's go, let's double-click on the technical substance on this, kind of unpack uh how we approach this from a technical standpoint.

SPEAKER_00 5:34

Yeah, we we really focus, and I mentioned this before, but we focus on grounding this AI in their business specific data. And in this case, we are talking about knowledge articles. So this agent force agent is not going out and reading CRM records. It is not making any updates, it is just grounded in this knowledge base, which again is a net new knowledge base that was created as they were performing this migration into Service Cloud. So, what we had to do is get these knowledge articles into the system and then set up retrieval augmented generation rag. This is connecting that AI agent to a vector database in DataCloud, which has those knowledge articles chunked inside of that vector database. And so uh the technical sort of piece there is chunking those articles, getting them turned into embeddings and enabling semantic search. But if you're not technical, all that really means is setting up the infrastructure in such a way that the AI can search using meaning instead of keywords, right? And so that's setting up these 130 or so articles initially uh in data cloud so that it's in the vector database. And that way the AI can then go and use retrieval augmented generation to go and search for those answers and provide results. So that goes also into the agent actions that are actually being used when a user asks a question. So it's not pulling from the open internet like something like you know, ChatGPT or or Claude, if you go and use just those chat tools today. That of course has access to go and search the entire web and all of its training data. That's great. But for this use case, we want to make sure that that agent is only responding using specific business information that's locked in these knowledge articles. And so that it with Agent Force, one of the benefits that you get is it's not searching the open internet. Now, can you give it the ability to do that? Sure. But what we did not in this case, right? We didn't want it to go out and search for some sort of test and then come back with something that wasn't specific to their business. And so this is grounded only in those knowledge articles that they have inside of Data 360. Of course, we have guardrails around it. We want to make sure that if somebody is asking about specific clinical information or medical advice, we have guardrails that say, hey, I cannot answer that, but I can connect you with one of our specialists. And so it's able to go and escalate that conversation to a human so that those more sensitive conversations can actually have a human touch involved there. And so that's really the seamless escalation path, which is, hey, this is what you're supposed to do, answer questions here. But if people start talking about these things, that's where we need to get a human involved. And so really you can boil it down to two capabilities answer questions, escalate to human. And then it's it's sort of just designing the right logic and pathways around, well, what questions am I able to answer? What happens if I can't find an answer? What happens if somebody asks something I know I'm not supposed to ask? And then that's how we design uh the escalation logic to make sure that we get a human involved.

SPEAKER_01 8:31

This, I think, uh correct me if I'm wrong, but this was one of the first implementations of this specific AI tech for the team. Now that the dust has kind of settled here, what's the impact this is having for the customer?

SPEAKER_00 8:41

I think immediately is uh about a 40% deflection rate. And 40%, I think, doesn't sound huge, but that's almost half of all inquiries getting handled by AI. So if before all we had was the ability for somebody to chat with a human, call in, email in, 40% is pretty strong for version one coming right out of the gate to be able to say, okay, you know what? 40% of these initial inquiries are never making it to a human because this AI is able to answer that question and it's not requiring an escalation. So that I think is just a huge win. And then of course, yeah, do we want it to be more than 40%? Who wouldn't, right? That's where you get into optimization and understanding what knowledge gaps you have and trying to figure out reporting on what questions are being asked but are not being answered, but maybe they should be answered. That can help inform you which articles are maybe missing that you need to write so that you can give the AI access to that information. And then, of course, the other side of that, is there anything that the agent failed to escalate when it should have, right? And that, of course, is going to bring that deflection rate down, but that's an important sort of safety metric, which is let's make sure that anytime somebody asks something of a certain content, we want to make sure that that's that's getting escalated to a human. So that that was kind of the big piece there. And then um, additionally, because we built this on a Salesforce knowledge base, let's say they do some metrics around, okay, we're getting this question a lot, but we don't have any way to answer it. All they really need to do is write a knowledge article, and it is immediately able to sync back over to the agent sort of within a day, you know, depending on kind of their configuration. But when a new knowledge article is created, it's automatically synced into Data Cloud, which is automatically a part of the retrieval mechanism that the agent can use. And so they're not having to do a lot of work to maintain that knowledge base. They're just able to add new articles, publish those articles, modify existing articles, or even if they remove an article, right? That's going to be reflected back to the AI agent. So it's a it's an easy way to uh maintain that. And I think it's a nice sort of uh symbiosis between your metrics and the action that you can take because you can you can take those metrics, determine what's missing, go and immediately draft something that can answer that question, and then almost immediately you have an answer to that question. And so you can say, okay, we weren't able to answer this before. It had to go to a human. We write one article, it's already set up for the pipeline to ingest that article and get it into the system. So then all of a sudden, now, once you've identified it, then the AI is able to answer it because you were proactively able to measure, understand the gap, and then design something to fix that gap?

SPEAKER_01 11:20

Yeah, I think when reflecting on this, and and I think a lot of leaders would agree, this project wasn't really about saving money on support reps. It's about a better customer experience. Uh what what what role does having that right data how how is that important for the AI tooling that we build?

SPEAKER_00 11:38

Aaron Powell We talk about this a lot, but it's it's huge. It's incredibly important. The context and the data and the ability to articulate your business process is directly correlated with the responses that your AI is gonna provide. And so if you have inaccurate data or bad data, I think I've mentioned this before, but but AI is a mirror and a multiplier. And so if I have poor, bad data, I'm not only gonna reflect that back, the AI will reflect it back, but it'll also get multiplied across all the surfaces where that AI has been deployed. So then all of a sudden, that poor data or that bad answer or that that thing that isn't exactly the way that I want it to, maybe beforehand that was a simple, hey, you told the customer this, but you should be telling them this. Well, now that AI could do that to 20 customers in a single day that ask the same question. And so that's it's multiplying it, right? And so that's why it's really important to audit your knowledge base, to audit the data that's being connected, to only connect the sources that are absolutely necessary for an AI to be able to respond back to a customer. We don't want to introduce other noise, right? Like we don't want a bunch of extra data that might not technically be wrong, but it's not specific to the use case because all that introduces noise into how the AI is going to generate those responses. And so I think the main thing is it's it's not a one size fits all approach. It's specific to every business. And what I have found to be very common is uh most businesses don't come into an exciting new AI implementation with the understanding that there's likely a lot of foundational work to be done first. And that foundational work is again making sure that you have the right data at the right time. That's sort of the whole context idea, context engineering, and making sure that you're you're designing a system to where the AI can go and grab the right data at the right time. But then it's also that process. If if you don't have a really good understanding of how your business works, well, you maybe you do, but it's not documented anywhere. That that's something we find a lot of times is that there's one person who owns a process and they've done it for a decade and nobody else knows how to do it and it hasn't been written down. That's one of those sort of homework items that almost everyone runs into when you start to think about how to implement AI. Because you have to be able to articulate that process, right? So you can't expect good results if you have a messy knowledge base. You can't expect good results if you can't articulate the business process that you're you're trying to solve for. And so I think um, you know, this was a good example of the whole end-to-end process, really, because we're we're migrating into a brand new system. We have the opportunity to make sure that all the data that we're putting into the system is good, is accurate, is ready to be used. We're rewriting that knowledge base from the top down. And so this is one of those rare situations where you're starting with the end in mind, as opposed to having 10 years on a CRM and now you have to kind of go back and figure it out. This was a unique opportunity for someone to say, okay, we know we want to implement this AI agent at the end of this. So let's make sure that as we build, we're doing it in such a way as to make sure we're only putting in the data that we want, that our knowledge articles are accurate and correct and written in a format that's going to be conducive for Agent Force to be able to surface those answers. Um, so yeah, I think I think it was a really good use case just to say, yeah, data is super important. It's a high sensitivity industry. Let's make sure we're only grounding it in the information that we want to surface back to those customers. Um, and I think it it ended up being a great result.

SPEAKER_01 15:11

Zach, I love hearing these stories about how we're thoughtfully implementing AI and solving real problems for our clients. Uh, for folks out there that may be dealing with similar problems, and I think you just uh hinted at this, but really trying to go at it alone and figuring it out is probably not the wisest move. Like bringing in someone like us that's got the reps, that understands how these systems operate and some of the pitfalls if you do it incorrectly, um, definitely a better move, huh?

SPEAKER_00 15:42

I'd say if this sounds familiar, then don't guess your way through it. You know, Mark Benioff would say don't DIY your AI. And I really I do like that. Um, but yeah, reach out to us. We're we're very eager to help out, take a look at your workflow, see what your knowledge base looks like, help make some recommendations, make sure that you're designing the right escalation paths and um hopefully help you get kind of from zero to an autonomous solution that's able to make real impact in your business and actually save time and drive those results. I think it's really common for a lot of teams to jump straight into the technology before they've mapped out the experience, before they've cleaned up the data or really defined what the automation sheet even do. I think if you kind of look at the scalable business framework, it really is all about understanding the process before you can automate it. And that's especially true when we talk about adding AI into the mix, because usually AI is is automated and it's doing a bunch of probabilistic work on your behalf. And so it's it's all the more important to sort of follow the process of defining a process first and then implementing automation and AI. So that's the kind of work that we help teams with every day. And so if that's something that resonates, we'd love to chat with you. Slow is smooth and smooth as fast. Hey, well, something like that. I can't remember what I said a couple episodes ago, but it was not that.

SPEAKER_01 17:02

That's good stuff. And just a reminder for everybody out there listening 40% win rate right out of the gate can fundamentally change the trajectory of your company. Uh again, Zach invited you guys to engage with us. If you want to dig in, uh, see how your organization can solve these kinds of bottlenecks with operational AI, hit us up fastslowmotion.com slash AI. We would love to help you. Zach, any final words?

SPEAKER_00 17:26

I'd say if you're building AI into your business, yeah, reach out to us. We'd love to help. But also just remember it's not set it and forget it. It is something that requires ongoing care and feeding and reporting and iterations and making those updates. And so I would say if you're willing to put in the work on your data and your processes, that is when you can get an AI that can put in work on behalf of your customers. So invest in that foundation so that you can actually see the results of that investment down the road.

SPEAKER_01 17:53

Well said, my friend. Thanks everyone for listening. We'll see you next time.