AI-Powered Growth: Move Fast Without Losing Control | INBOUND 2025

Foreign. Hi, everyone. We are so excited to be here. We're going to talk to you about AI powered growth and how to move fast without losing control. If you've been here this week, obviously you've heard all about how fast AI is moving. Everything is changing all the time. There's a lot to keep up with. But Most importantly, at HubSpot, we really believe that AI should not be a luxury for large corporations. And so your ability to be able to move fast and keep moving is dependent on context, learning, and having just some fundamentals under your belt. So that's what we're here to provide a roadmap for the next couple of minutes that we're going to be spending time together. First, we're going to validate why AI feels hard right now. So consider this a little bit of your support group. We're here to validate some of those things, give you a little steer on guidance. Second, we're gonna give you a framework of five questions that you need to ask to really support the conversation around vendors that you're looking at to use their AI products. And then last but not least, we're gonna circle back on a little bit of how we do this at HubSpot, pull back the curtain, share some of how Karen and I's teams work together and how we tackle some of these questions internally. Finally, we're gonna give you a couple practical tips that you can take with you, and we'll open the floor for a little bit of Q and A. But first, let's do a few intros. I'm Erica Fisher. I'm our chief legal officer, and my job is to make HubSpot a company that you can trust. And I have Karen here with me who needs no introduction. Although, Karen, if you want to go ahead and give a few points about yourself.

Hi, everyone. I'm Karen. Hopefully you caught the product he did this morning, but I run product at HubSpot. Eric. I and I work very closely. A lot of the things I talked about this morning were around data and context and even the connectors. And we think and work very deeply to make sure that it's data that you can trust and that we know where the data is going.

So Karen and I work closely together. As she said, we really build a lot of our questions around clarity and control. So more than anything with our customers, we want you to understand what we're doing with your data. We carry this principle through to vendors that we use to support our customer's data, and then we really think carefully about what Kind of control that we put in the hands of our customers so that they always feel like they're in the driver's seat with their data. We work to ship products fast together, but in a trustworthy way. And so we're gonna talk a little bit about how we do that today. But first off, Karen, let's talk a little bit about why AI feels so hard right now. What is it about this particular moment that makes it feel like a different set of questions for the tools that people are turning to use?

Yeah. As you all know, sometimes internally we'll use a phrase called AI speed, but there's a lot of whiplash in AI. Right. There are a lot of policies that are still getting defined, how training is used for models, what does it mean to use AI for things? So one of them is kind of whiplash. Why it feels so hard. It's also hard to plan. There's something, actually, I say to my team, which is, how would you build product if the ground underneath you is fundamentally shifting? And so, just from a mindset perspective, AI is really hard to plan for. And I think we need to plan our roadmaps and what we're thinking about, knowing that things will change. If you were to Dharmesh's session today, he kind of said something that was something that is hard but not there yet. And that's a little bit of the perspective we have on the world, that AI is moving so quickly and changing so quickly that even things that aren't possible today will be possible even in a few weeks or months.

Great. So Karen talked a little bit about the pace of innovation. I'm going to talk a little bit about the fact that the rules are missing. And so part of this is when you're having conversations, particularly when you're getting your legal or compliance teams involved, if things feel confusing right now, it's because they're confusing right now, even for the lawyers in the room. And so one of the things to keep in mind is that there's a lot of different laws forming about AI. The US is having one perspective in terms of being able to compete and really putting speed and innovation at the forefront. The EU as a standard with the eu, is really leading with a sort of moral compass or a moral standard for the rest of the world. And there's a lot in between there as well. The problem with this with software is that software is global. So it's not as though you can sort of say, hey, my data knows that it's in New Jersey now, and it's going to behave differently than if it's in Brussels. That's a completely different sort of set of circumstances than we see with other regulations. So what this means is that vendors really need to be building for the highest common denominator. And so some of the questions that we're gonna be going through today are some of the questions that you should be asking of vendors. And it's just a signal that they understand the nature of the world that they operate within. So this is something that we'll talk about a little bit more in depth as we unpack. But these regulations are still forming, so expect this conversation to evolve, expect the sort of confusion to persist. But the framework that we're gonna give you should help you cut through some of that noise. Finally, let's talk a little bit about, on top of that, the language that's changing. So we don't have a lot of standardized terms around some of this stuff yet. It's still forming and norming. So what does that mean for you in the product world?

I mean, there's a lot of things, I think, changing with AI and products. So sometimes as you evolve this. What does it mean for something to train, for instance, a large language model versus pre or post training? What does it mean for it to train, for instance, a prompt or tool? And I'll give you an example. Today I launched the announcement for the HubSpot connector for Gemini, we have one for Claude, and for ChatGPT, it's a big difference for your data. We don't want any of your data, for instance, trained. We're very careful that it's never trained on a model. That is a very kind of important distinction between it being trained for the tool. So today you saw, as you typed in HubSpot and ChatGPT, it knows that HubSpot as a tool, what's happening there is ChatGPT is learning which tools are the right ones for the context, so it makes the tool selection better. But that data and your data is never used for model training, which is really important to us.

And as the product teams are settling on this language, we, as the lawyers and those reviewing some of the contracts, are trying to translate into that, into terms that everybody can agree upon. So you can see that the ball is sort of moving as we're trying to follow it. And so when we start to try to draft terms around what does training mean, et cetera, we really need to make sure that we're spending time with our product teams and looking under the hood and understanding what that means. This is All a landscape that is evolving. And so if you find terms where you get frustrated by them or you feel like people are talking past one another, that is a real thing that's happening. And so some of the things that we're going to talk through today help you cut through that noise to be able to ask the questions that matter most of the tools that you're using. So let's get into that a little bit. So we have a five questions framework, and we use this even internally at HubSpot, because one of the biggest things that I personally believe in is being able to give people the tools to get to the heart of what really matters. You shouldn't have to have a team of compliance professionals or five lawyers in the room every time just to get through everything. This should be sort of layperson speak and something that you can understand, especially if you're the buyer of the software. So we're going to talk through a couple of these and Karen and I are going to ping pong a little bit back and forth. But the biggest thing is these are five questions that matter anytime that you're buying software. But AI puts a specific lens on top of it. And so what we're going to do is we're going to talk through these five questions, but we're also going to give them a little bit of a, like, confidence score. How confident does AI layered on top of this make you in understanding these questions or getting the answers to the them? And in some cases, little has changed. In other cases, a lot has changed. So we're going to go through those one by one.

I get to give Erica the first question, which is, is my data used to train AI?

This is probably the number one question that we get from customers right now, and it's coming from a couple different angles. One, customers either want like a very clear yes, no answer, or two, they come to the table where someone has told them, training, you know, AI with your data is a very bad thing. You should never allow it. And the truth is somewhere in between. So, a couple of things to note. At HubSpot, we do use customer data to train our AI for a couple of different purposes. The first is for personalization. And if you sat through any of the keynotes this morning, you should kind of understand at this point that you actually really want that. If AI is supposed to be an extension of your team and a team member, you want to give it context. You want it to be highly personalized to your team. So our AI, while it looks at your data, can understand for example, how can I make this email tagline sound a lot more like your company? You heard some of the customer stories with Morehouse College and how they were able to form a real brand voice and use that and it saved them time and energy as they move through things. That's the power of AI being able to train off of your data. Now when that happens, it happens within your account, it happens on your business data and it never leaves sort of that instance. It is highly personalized to you and it's never used to support anyone else. The second way that we use data to train our own internal AI models is with your permission, when we train on data to make features better for everyone else. And this is where we look at things like what is the average deal time that we close deals within a specific industry in this area? And we can surface insights to you that say, hey, this deal typically takes about 30 days in your industry and segment to close. Here are some of the things that you should do. We do that on an aggregated basis. We never do it in a way that identifies your business intelligence. We never do it in a way that is shared with any other customer or that could be used as a competitive advantage for the lawyers in the room. This is a trade secret and we never violate trade secrets. So Those are the two ways within HubSpot that we use data to train AI. And, and in all cases it's really about adding value to the customers use of our products. Now I'm going to shift over for a minute because that's the sort of vertical SaaS application space when you're using a large language model like OpenAI, ChatGPT, Claude, whatever your model of choice is going to be, that's kind of a different ballpark. And so I really encourage people, whatever tool that they're buying, that they understand what the application is doing. A large language model is going to, is by design designed to hoover up a whole bunch of context and have any number of use cases attached to that. In that case, you do not want your sensitive business data being sucked up by a large language model. And so you want to look at the terms really closely. And here's a pro tip. In a lot of cases, if you're on an enterprise license, they already understand this and they push a lot of businesses to that license to say, hey, we've built in these protections so you don't have to worry about that. The tricky part here is that the business models are still settling. So there's sort of prosumer tiers there's consumer tiers. There's a lot of different tiers of licenses that are still emerging among LLMs. And so that's why I always advise customers, vendors, whoever we're talking, to look at a large language model, experiment with it, figure out if it works for you. But then once you do upgrade to an enterprise license because you're going to be guaranteed that protection that says I I'm the type of user that's putting sensitive business information into this model. I'm putting trade secrets in. I do not want that to be used by my competitors or used to teach other people how to compete with me. So that's the distinction there. We're going to talk a little bit more about how this works when we actually use LLMs to power our AI, which, spoiler alert, they are not allowed to train on your data. But Karen, do you want to talk about a little bit about this in the MCP connector context?

Yeah, I'll add one thing kind of to that too. Erika talked about the enterprise tiers and this prosumer tier. We think of this prosumer tier as consumer tiers. So like if you're in ChatGPT, the plus would be considered a prosumer tier. The important distinction is their business tiers like team and enterprise. They're designed to make sure that data is never trained on. So by default there is no setting that says turn off training, they're not trained on. The prosumer tiers are like the plus ones. There is a data setting and we just want people to be educated about that as well. It's your choice to choose if you want to show share data to that model. That's a really important call for me. I have an 11 year old son who started to uses AI for many, many things. But that's one of the things I even teach my son. Just be conscious of it, turn off training if that's not what we want to happen. So Erica just asked me about mcp. If you went to Dharmesh's session today, which was awesome, he taught, he talked about the Model Context protocol, which is what MCP stands for at HubSpot we actually have two of them. We have two different kinds of MCP remote servers. Today is the GA, I believe if some folks are in the room of our MCP server and then we also have an MCP server that powers the connectors that we do for ChatGPT and Claude. One really important thing when I was looking at the McP servers is MC by the way is just a protocol so if you find people using MCP servers, if you see other HubSpot ones around the Internet, there are a lot. One of the reasons it was important that we built ours is because we could protect auth end user permissions. So in our connector, there are two layers, mcp, which is how the data will get used. And then we also have an oauth layer. What that does is it's kind of like a handshake. And it makes sure that your HubSpot data, when it's used with any of these connectors, has a handshake. It says, I know who you are and I know what permissions you're allowed to see. And so those two, hand in hand, make sure that we can always protect MCP servers. That is not true for any MCP server out there. It is just a protocol. So just make sure you know how that's being used. And it was one of the reasons it was so important for us to write our own.

Awesome. Okay, so on a scale of 1 to 5 of complexity, I would put the is data used to train AI at a 6 on the sort of confusion scale. Right now, moving to the question of is my data safe? And these questions are really devised for our business stakeholders, our salespeople, but I invite you to use them too. When conversations get confusing or terms get sort of muddied, cut back to these questions to kind of get to the heart of what are you concerned about? Or what is your customer concerned about? And so oftentimes we'll use them to say, like, what are you really worried about here? Are you worried about your data being safe? Are you worried about data being used to train on AI, data being safe? Karen, where would you put this on the scale?

I don't know if that one was my question. I was like, data being safe? Well, data is evolving so much. So how do you see AI changing what it means to be in data security?

Yeah, we get this question a lot. And I would say from a complexity standpoint, I would put this at like a 2 or a 3. So not much has changed here. And when you're looking at software vendors, especially those in the cloud, whether they use AI or not, there's some pretty widely accepted standards. ISO 2700127002 SOC2. You can generally find those with a little poking around on a website or the help of an LLM of your choice. This is not changing today. Now, there are some standards that are emerging that are AI specific security standards, but they are not widely accepted yet. And so I wouldn't get too sort of wrapped up here in what's your AI security standard? Sort of a false question. Stick with the basics, stick with the gold standards, and you don't need to spin yourself up too much here. Kicking the tires is good. Thinking carefully about things that are sort of newer products, you know, rather than old, tried and tested true products. But generally these third party standards are gonna come with audits, they're gonna come with consistency. So I think this is a very safe place to play in terms of what you're looking for. All right, next question. Who can access my data?

The score for this one for me is probably a four or five. So it's definitely made it more nuanced to say who can access your data. So one thing is the first bar, which is user permissions. Even in a world of humans and AI or human and agents, that agent needs to. You need to think about permissions for that agent. And if you think of it just kind of as a human license, what are they allowed to see? So I've been talking a lot about our connectors, talk about Breeze a lot as well. We think a lot about user permissions, which is, you know, if I'm a rep that has, you know, we might manage deals regionally, am I allowed to see all the deals that are inside HubSpot? Unlikely. You know, what are the deals I'm allowed to see? And so that is governed by permissions. And that has been much more straightforward. When we build things with AI, we want them to respect user permissions. The second part of it gets a little bit more nuanced, which is where does that data flow out to? So in a world where we've had APIs before, if APIs don't respect user permissions, that becomes a big issue when agents are allowed to. So I would say with AI, it gets a little bit more nuanced because you have to think about where does that data, what's the second jump from that data? Like if somebody exports it, if somebody uses an API, where does that go? And then who can use that data for what can they build, for instance, a really smart agent off of it, which we all want to do, but does it respect context, does it respect permissions? And does it then not allow people to build off of that?

I agree with all of that, including your score. And I would double down on one extra piece that you said there, which is where does my data go? And so one of the things that we hear a lot from customers and that I encourage, anytime someone's purchasing software to understand is where in the world does data go? And really understanding is that something that you care about? So some of our customers, some vendors, they do care about keeping data localized to a specific place. Whenever you have an LLM involved, that's something that you really wanna think about because oftentimes it's the underlying infrastructure, just like AWS or Azure or Google Cloud to understand where that sort of underlying infrastructure sits. So it's an extra step that you wanna think about. Usually an ask for a simple data map can answer that or answer that for the person that cares about it within your company. But I would add onto that that understanding where in the world does your data go go and therefore who can access it is an important one on top of it. All right, moving on to the fourth question. When I delete my data, is it really gone? So Karen, we get this question a lot and you know, some of our questions are fueled by a cross functional team behind the buyer and so sometimes they're just passing along questions from a security team or something else. You are on the receiving end of some of these questions. How would you score deleting data with a AI?

I think that the field is still changing quite a bit. So if you remember GDPR days, you know, was requirement that you always have to have an API that if you wanted to delete it would eradicate it in the system. I think that is still getting formed, especially in the AI worlds. Sometimes in AI you'll hear things called zero data retention. That was really popular even like a year ago. This is actually probably even eight months ago. It was really popular. Zero data retention. The speed of AI is moving so quickly. And so one of the things I announced this morning was how Breeze Assistant has memory. People love it that ChatGPT or Claude has memory. You can't have zero data retention with memory because just like humans, if I have long term or short term memory, I am retaining some of that context and data. So those are the trade offs that you're gonna start seeing. I think zero data retention, it will be really important to think about how your data, like if you want to delete it, how does it get deleted across the system? Those things are really important. If you want the benefits of things like memory and knowing you in context, those are things that will live at a certain time frame.

Agreed. And that's something that we talk a lot about actually internally because data deletion is sort of something that we have on our scorecard of questions. It's something that our teams know to check for. But we had to have a lot of internal dialogue to say, actually data retention is context. And context is magic when it comes to AI. It's just like a team member that you want to have all of the context from day one of onboarding. So I'm personally very comfortable with some level of data retention. What I look for more, especially if we're using AI based tools, is just do you have reasonable processes in place? That shows me that when we ask for our data to be deleted, and usually that's when we're terminating the services or maybe when we have a request for, from a regulator or from somebody that has rights to delete their data under, for example, GDPR or some other provision. What I'm looking for is what's your process? Is it reasonable? Does it roll out of your systems on a fairly regular basis? It's very reasonable to have backups to systems. This is where data deletion gets sort of intention with continuity of your service and your system. So I'm very comfortable with the idea that a vendor is going to retain some amount of data, especially if they're providing AI services that need that context. What you want to look for is what is your process and is it on a reasonable timeline and do you have reasonable answers to that in terms of where it rolls over? So this is one that I think is changing a lot with AI.

There's something that Erica says as we, we have a lot of discussions around data and how it's used in AI. And you know, one thing I think we say a lot is control and clarity. And so those are like two principles you can think of around, around how data is used. Do you have control over it at which timeframe and where it goes? And then do you have clarity on how it's being used? And so those are just like two ways to really like think. Use a principled way to think about your data showing up in these.

Okay, the last question when things get muddied and you're trying to cut through is what are you trying to get at here exactly? Are you concerned about whether or not we will help you meet your data obligations? And a lot of times you can use the acronym and the buzzword GDPR here to kind of stimulate some conversation. Most people know what that means. There's a handful of other regulations that sort of hit the buzzwords around data privacy, but this is really about will your vendor stand behind you when you have a request from a lawmaker, a regulator, or somebody who has rights under that to help you action that request. And so I'm going to answer this one because I'm on the receiving end of a lot of these and I would say that this is about maybe a 2 or a 3. So the basics still apply. Your vendor should be very clear on how they are going to help you answer data access, data deletion requests, that should be clear. It should be on their website, et cetera. The other piece that is evolving is for those of you who operate globally, the EU AI Act. And to really cut to sort of the essentials of what you need to know about that is that the EU AI act splits software into two broadly different categories. The first is high risk software and there's a lot of information online about this, so I'll just give you the Reader's Digest. But this is something that if it's important, you can look into further. But this is not most SaaS, business applications. So this is software that's making high risk decisions. Think about whether or not you're qualified for a loan, whether or not you get a specific job, whether or not your health data is used in a specific way. They're very high risk decisions. Things like business application software generally don't fall into that category. And so what the EU asks is that for that second category where you have business application systems that are processing data with AI, that there's a lot of transparency around it. What they want to make sure of is that the humans remain in control and in the loop. So they want to make sure that when you are making decisions or you're offering reasoning, that it's clearly indicated within the product. This is part of the reason why we use the sparkles within HubSpot for to indicate where AI and Breeze is operating. We also use it. And many of you will be familiar the large language models that show the thinking. That's not just to make you feel better about the fact that it is moving. And it's sort of like the equivalent of the hourglass that's moving, but it's actually a piece of transparency. It shows you how the software is reasoning, helps you understand its deductions. And just like you would be managing maybe a more junior team member, you're getting the answer, but also validating the reasoning behind it. That's all built into the regulation. And so that's what you want to look for. And that the vendor that you're using has some reasonable sort of, you know, discussion of that within the documentation that they're providing to you or even within their website. So this one is one that I don't think is going to change a ton. I actually think the EU has learned a lot from their first attempts to regulate some of these sort of trickier areas and are doing a much better job of balancing practicality with allowing for innovation. But transparency is really the key here. All right, so those are the five questions that we really try to focus our teams on and try to have normal layperson conversations without trying to get too complicated on things and helping our teams navigate and move fast. But how we handle this at HubSpot, we're going to talk about just briefly and maybe share a couple stories on how our team have worked together. So we talked a little bit earlier about clarity and control, and of course we talked about the fact that clarity means you really need to understand how your data is used, and control is that you have the choices built into that and around it. And Karen, maybe we can talk a little bit more about when we launched the MCP connector. And I know you came to our legal team and said this is really important for our customers. We know that we have a high demand for this type of integration and we really want to make sure that, that we are skating to where the puck is and ahead of our customers in terms of what utility they might have in other applications. Can you talk a little bit about how we went through that process and even talking to some of the teams that help support these MCP connectors on the LLM side?

Yeah, actually, so kind of a little bit of behind the scenes, too. One of the reasons we moved so fast to become the first connector CRM connectors for these is we kind of have a deep belief that if you think about every disruption wave that has happened, let's just pick, you know, Windows and PC. In the Windows era, there were an application ecosystem that happened on top of operating systems. That same thing happened with browsers. So Enterprise SaaS was born out of browsers and Chrome, Mozilla, a whole bunch of different ones where you decided as an app if you should work on each of these browsers. Same thing happened for mobile. I'm a deep believer that in AI, it's, it's going to happen again. And so if you think like that, it's. You want to bring context to the next operating system. In a way, it's very kind of abstract that. And so as we build these connectors, it was. We should allow people to grow everywhere and do it well and. But we also should make sure that we want HubSpot to be easy to use. We want it to also be easy to trust. So we need to do everything ahead of time before we even release these to ask all of the hard questions for are we gonna protect this data? Are you gonna ensure that people can see it? And that led back to, okay, we have to write our own OAuth layer so that there's this handshake that protects user permissions, seats, how data is used and how data is trained on. So we spend a lot of time with all the providers, OpenAI and Claude and, and Google, talking about what we will allow and how our connectors and how our data can be used.

Yeah. And I would add to that, using the framework of clarity and control, it wasn't enough for us to say, you're connecting to this third party tool for which you also have a license. Check their terms. When Karen and I sort of tried that on for size, we were like, if I were a customer, I would not feel great about. And so I would really want HubSpot to have the answers of exactly what happens with my data from this point forward and what are my choices. So this is just sort of how our teams then, you know, use guideposts to operate, what kind of questions we ask of our vendors, how we build partnerships, how we, you know, fast we allow our teams to move, but really always through the lens of if I were a customer, would I be cool with this, would I be clear on it and do I feel like I'm being given of control in this situation? And it's the same standard that we hold for our vendors. So that's one of the tests that we use. And where a vendor isn't super clear or we don't feel like the controls are super straightforward, that's an area where we tend to dig in a little bit more and move a little bit more slowly.

Yeah. And to give you a little color behind the scenes as well. It meant we fought really hard to make sure that it would never be trained in the models. And if it ever compromised that one of our decisions and choices might be that we don't allow the HubSpot connector to work in some of the tiers. So, you know, as we go forward, we just want to make it easy that any AI that we offer is easy to use and easy to trust and that we are like really thinking if we even want it to be available in places. Luckily for us, as we worked with a lot of these vendors and teams, we were able to do that, which is this data is not used to train models. It will help tools get better. The same way that features in a great product that you might love helps get that value, but it won't be used for another purpose that way.

All right, we promised some practical tips. So before we open the floor to questions, Karen, what are your top three tips from this conversation?

It's probably the same that I tell my son. So check your defaults, know what they are. It was really important to me that you know you need to know where your data is being used and you often do have control. So check your defaults, make sure it's showing up the same way. It happens in a lot of different applications. Another is kind of like know where it's showing up, right? And it's always a value trade off. You know, we've seen that for a lot of different products and a lot of different things. So make sure you're getting your value trade offs. Like I love Netflix. I watch a lot of recommendations, I watch recommended things, but it is using preferences that I use and what I watch. So, you know, that's a good trade off for me, which is it gives me great things on the top 10 movies that are happening right now, but it uses some of my watch preference. So just know what those defaults are. And then, you know, there are things that I think the AI field is just evolving so much. I mentioned this morning, I'm a machine learning grad and it's so powerful to have memory. So things like zero data retention that might have been so hot, like such a hot topic eight months ago. Like, it's amazing where AI is taking us and just be kind of conscious of like what that means with some of those policies. But it is a trade off where you want control and clarity.

So I would say yes to all of those and add a few more. One would be training isn't a dirty word. And I think that that is something that like we see a lot right now just like a knee jerk reaction to training. What I would say as a tip here is training can actually help make the product that you're using better. However, if you're using a large language model, I would really either check your default or upgrade to that enterprise license because half, more than half the time.

Or team. Or team.

Or team. Yeah. So I'm using enterprise license broadly, but know the license that is fit for business use. And the whole point of that is these areas are still evolving. The way that these companies go to market is still evolving. And really making sure that you're in the license tier that's fit for the purpose of your Profile, it's gonna be an enterprise license tier. It's gonna allow you to sort of bypass a lot of these questions because a lot of those defaults will be built in. The second thing I would say is always look for ISO 27001 and SoC2. So if you want the gold standard from a security perspective, if you have either of those, from a vendor perspective, you're in a pretty safe zone that that vendor knows what they're doing and that they are abiding by the best security practices that are audited by a third party. And then the last one I would say is know who powers your AI. And so you know, if you go to the Trust website for HubSpot, we have our models that power all of our AI listed. Almost every SaaS application right now is using some sort of third party LLM to power their own enterprise AI models. And that's because this technology is so powerful and that's kind of what it's built for and part of the business model scale over time. However, that should also come with promises that that engine that's powering is not training your data within the product. And so that's something just to look for and understand what models are they using. There's often several because different models are good for different things. But they should be able to clearly articulate for you which models are powering those and whether or not they allow those models to, to train on your data, which as we discussed, the answer should be no. So those are my key takeaways that hopefully give you a little bit of a practical roadmap. If you're the one buying software or maybe you're interacting with a legal or compliance team, that helps you get a little bit closer to the line of being able to adopt AI, which is the commitment that we have here at HubSpot.

You know, one thing I'll add, I'm gonna add a last tip. I think we've talked a lot about data and training and I think what we don't want is, you know, I'll actually say it internally a lot, which is how do we move at AI speed? To move at AI speed means you also need to be curious and you also kind of need what need to know the guardrails for testing out new tools. So you know, for testing out new tools I'm like test quickly. And the guardrail for, for the team for instance, is like, don't put too much, don't put work sensitive information in there. But if I test, I test a ton of AI Tools. And I'm fast to test and try things in POCs. I'm fast to test things like what should I plan for my 11 year old's birthday party? Those are places that it doesn't matter to me much that like I'm not working with sensitive data. So I think that's my last tip too, which is we're all kind of evolving at AI speed. I know it's important to test out different tools and that would be my guidance, which is, yeah, I want my team to test quickly. It means that we at HubSpot, we kind of have an experimental budget that we allow people to quickly use some tools. And if it starts working, if it starts scaling, that's when we start looking at. Okay, if we need to put sensitive data work information in. Now we talk to legal or procurement about 8 enterprise license or a team license. So I'll leave you with that because I know many of us are still exploring many tools and to keep moving at AI speed, you also need some kind of guardrails to think about how to experiment.

All right, so hopefully that gives you some tools leaving here today to move a little faster in adopting AI. And at HubSpot, we really are committed to our customers being some of the best companies growing in the world. And we know that for all of you today, that means being able to adopt tools that keep you at the most competitive edge. So that was the goal of our conversation and hopefully you leave having a little bit more clarity. We're going to take a few minutes for Q and A, if there is any, and I'll open that up now.

Hi, I'm Garrett Hogan. I work with Jackson and Coker in Atlanta, Georgia. You mentioned what you guys do with your MCP and your security. And I know there's a lot of discussions right now about that is a big kind of like concern as to how do you really secure those MCP servers to make sure that they don't access data incorrectly or don't take actions that somebody shouldn't be able to do? Do you think that the way you guys approached it is something that will become the standard? Or did you guys base it off of someone else who you think is going to become the standard? And when do you think that's going to be something that's shared so that other people can start utilizing it?

Yeah, I definitely think it's something that's shared. So we based it off of the OAuth protocol, which I think because it already is such a good standard, we knew that handshake would be trusted. And you're seeing that actually a lot with some other MCP servers and connectors. And then in ours with our remote server, we wrote that layer as well. So I think I would say be aware of the millions of MCP servers out there, because not all of them implement OAuth. You really do want to make sure that there is a good handshake and that's something we've used as a standard and we share it back as well. So we built our ChatGPT MCP connector off of the way that they extend ChatGPT. We were the first. We helped shape that and we used OAuth for it. So I think as you look at the protocols that are coming mcp ada, the AUTH protocols are not all there yet. So look for ones that are leaning off of the standards like OAuth. And I expect things to also evolve and change so that those get plugged.

Thank you.

Hi there, my name is. I'm a lead innovator at Wallet. My question would be as both the chief legal officer and the product head. Head of product, sorry, how. What's your experience? Or what tips would you give for development teams wanting to move fast at AI speed while also leveraging, let's say, key tools such as the breeze ecosystem you guys just recently are expanding upon, to move and develop a sustainable AI strategy for the company.

So first off, I would say, you know, and I think this was covered in. I think it was your talk today. I'm trying to remember which. Which talk I listened to that was about. Or maybe it was Dharmesh's talk around. Just, I think your AI strategy has to be focused right now on adoption and use, just given the competitive advantage that's there. So, for example, internally within HubSpot, when we started to get a lot of requests for AI first tools from our internal teams, we took a step back to say our approach to how we normally onboard tools and vet them is. Cannot be the same today as it was yesterday. Things are moving too fast. Our teams need to learn faster. And so that was like one thing that we agreed upon just from a risk posture perspective. Then we looked at, okay, how do we allow our teams to move faster? And we put in place some things kind of what Karen was pointing to earlier around, if you use this type of data, it's lower risk and you can experiment and then figure out if you have a value proposition here and whether or not you actually want to, you know, move forward with a stronger protection. Or maybe we want to reach out to the vendor and say, is there a broader set of data that we could use in between? But I think that's important to get the learning loops going because there's so much right now on the market that there's no way in a year's time you're going to be using all of the same tools that you're testing today. So testing and learning is the key. And then having a process by which you say there's an accountability, right, from whoever's using that to say, yes, this returns value to me, I want to actually invest in it. And. And then putting the full weight of your teams behind that to say, all right, let's reach out to the vendor, say, this is something that we want to move from just experimentation to a true commitment. Can you meet us where we are? And here's the data that we plan on putting into your product. And those five questions are ones that you can kind of use to sort of poke at them a little bit more. Like, how do you keep the data safe? Do you use it to train AI? This is the type of data I'm using. And that can get you a long part of the way in the conversation before you even. And get to a contracting phase or terms.

Yeah, some of my tips. You said you're in engineering. Yeah, yeah, yeah. Some of my tips would be.

For.

Instance, for Inside HubSpot, all of our engineering uses AI tools, but when we picked the first one, we kind of had a small group go do kind of proof of concepts where we used a couple of different ones and then they came back with a recommendation. I would recommend, if you don't have to sign a full contract, don't. If you have to sign one, sign a one year one. So what we have done is we've given, we have a certain budget that we're like, okay, if you want to use an AI tool, it's under this amount. Go and just apply and use it. So that's how we kind of unlocked the exploration then as we saw people really use it more like in engineering, we knew we wanted to use one of them, right? It's, you're going to use GitHub, Copilot, you're going to use cloud, you know, you're going to use cursor, you're going to use cloud code, you, you're gonna use one of those. And so we wanted to enable people to use it broadly. And then when there was enough like there was a recommendation, then we said, okay, we need to move to a contract. We recommend short contracts because it turned out the one that we were using 18 months ago is actually not the ones that we're using today. It's just advancing so quickly. The Codex agents. Same thing for PM products and even design. For instance, there's been a really big shift in prototyping and UX. So many of us used V0 dev for a while in product, and now many of us are using Lovable. So, you know, if you think how quickly things are expanding, Lovable wasn't even a product like 12 months ago that people were talking about, and now so many people are using it. So those are my, like, kind of practical tips, which is how do you unlock the short POCs as you get those POCs? Like, have a small group make recommendations and then sign short contracts.

Thank you.

Hi, I'm in the payment space, FinTech. So we have a lot of PII, and I've been getting a lot of pushback on launching the connector, either ChatGPT or Claude. And my CEO founder, who's a developer, found an article where our competitor Stripe had one of their employees use Anthropic to basically do some corporate espionage and pull out some of the data and supply it to a competitor. With the use of this, obviously, we can lock it down to certain individuals, whether it's development team, Rev Ops, finance. Do you foresee an area where there'd be like an audit log to be able to control and monitor how these MCPs are used with this system?

Yeah, we do. So, like, as we've learned a lot about the connectors, one of them is giving more controls to what we call the Super Admin, so that you can kind of see what's being used today. Because it's an OAuth handshake, it's, I think, refreshing like multiple times a day, like multiple times, like every hour. So that means that even if somebody like is no longer in your company, it would be refreshed. Refresh. But we've learned a lot about the super Admin controls and I do expect us to think a lot more about what needs to be built. And that's, you know, we've talked a lot about how data is also changing. The AI landscape is evolving. The connectors themselves and those, those LLMs are also learning a lot about that. It's kind of the difference between the Pro and Plus tiers in the ChatGPT world and the team and enterprise. There are just certain kind of admin controls that you would expect. Both team and enterprise, as example, have a compliance API. So those are things that you can See, And I would expect us to look more at that as well.

Awesome, thank you.

Yeah.

This one's a little adjacent. I don't know if it's entirely on topic, but we've talked a little bit today earlier about marketplaces and then you're talking about these connectors, just your vetting process and how we can see what you've done and why you choose to trust them. Obviously you talked about the standards that you have, ISO 2701 or these other standards, but how do we, how can we investigate these ourselves to make sure that we want to trust these and see the process that you've done, whether it's the connectors or whether it's apps that you have in your marketplace.

Yes, I can take a little bit of that. So I think it depends on which application that you're connecting. We try to publish really easy to understand knowledge base articles that get at some of the FAQs, and we certainly have teams that sort of are on hand to help with RFPs, et cetera. A lot of our marketplace vendors are also our solutions partners. And so they're there to sort of help understand and navigate your specific vertical, your specific segment. And then where we have our development platform for our marketplace vendors, we also require them to abide by security standards, you know, practices audits, et cetera, that we would hold our other vendors to. And so we try to keep as much of that knowledge public facing as possible in FAQs that you can then use for your compliance teams to be able to find.

Awesome, thank you.

Hi, my name is Miha. I'm a co founder and CTO at the Swarm. You launched a lot of products today. I think somebody said like 200 new features where you like, did you stop any AI related features from launching because there were, you know, legal concerns, privacy concerns, or were you like close to like, you know, stopping a feature or a product from launching because of these concerns?

One of the reasons you kind of see Erica and I on stage is we work so closely together now with product and legal. And one of the most important, like when we, we have a. A lot of what I announced this morning was around context, unified data, how data gets transferred. A really important goal as part of that strategic initiative was actually how your data is treated. So I give you that because I think before we've even entered some of these discussions, like we knew we wanted connectors, we knew we wanted like the Breeze Assistant connectors. Those are ones that bring in Gmail, those are ones that bring in like email, your own Context. And like, because we knew that that was so important, we also launched a note taker that I didn't talk that much about. I talked about meeting brief. All of those are actually pulling in unstructured data. Unstructured data itself pulls in a whole bunch of stuff because much of that data was sitting in your CRM. But now you have PDF files, you have emails, you have called transcripts, you have PII in that. And as we even started development of those features, we were talking hand in hand about how to think about like data and legal ones. So I would say I don't know of any features that we stopped because of the data and legal reasons, but we were talking so early about them. And then because we were talking so early, some of the early conversations were nuanced. So like the ChatGPT connector we just talked about, it was very nuanced. And if we were, if we couldn't prevent training of the models, you may not see the connectors in all of the tiers. And that was kind of how we thought about it.

I'll just add on to that. I agree with that and I would add on. So all of our teams go through training about how to develop software and develop products in a secure way, in a privacy compliant way, et cetera. It may not be the most fun or popular part of their training, but they all go through it. And so the yield of that is we rarely come across product or even product features. We're like, oh my gosh, I can't believe that you're building this. What are you talking about? Like that, that rarely is the conversation. What often is the conversation is, okay, this is what we're gonna do and this is the value it adds to customers, but this is how it's handling their data. Did we go far enough in explaining that to the customer? Did we go far enough in like, would they get that from where we are or at the point that they're integrating this thing, have we made it clear enough to them at the point that, that that information is relevant? Those are usually the conversations of just putting ourselves in the customer's shoes and saying, okay, based on our principles of clarity and control, if we were the customer, would we be surprised here? Would we be disappointed? Would we understand what happened? And so that's usually where we are seeing the sort of finishing conversations happening between product and legal teams.

I think one other thing we didn't probably bring up that I think we deeply believe is your data is yours. So you see this actually a lot of it in policies happening now where people might say, okay, you can't have access to your data depending on what you're using. And so far that has been a principle where we think it's your data, what should you be able to do with it? Where should you be able to take it? And that's kind of an important concept too as we think about this. So hopefully that helped answer some of it.

Thank you. Can I have a follow up question?

Sure.

Erica, I don't have a legal team. Many startups don't have a legal team. Is it okay to consult ChatGPT about the security and privacy aspects of my idea?

Yeah, I think, I mean I'm a huge fan and we talk a lot as a legal team about like where the legal industry is moving, you know, and we talk a lot about the fact that most of our customers are small and medium sized businesses that do not have a legal team or a compliance team. And oftentimes they'll even show up on our like dust demo calls with their legal avatar attached and we're like, awesome. I think ChatGPT is the perfect use of, you know, getting like making you supercharged. Right. So it's very, as a, as a small business, medium sized business, you know, practicing, it's a pretty small risk that you're gonna be sued or you know, that there's going to be a dispute and there's a pretty high likelihood that you're going to get most of the way there with ChatGPT versus, you know, $1,000 an hour lawyer that you probably can't afford. So the incremental improvement on top, I think like this is a perfect example of AI democratizing access to, you know, competitive advantages and operational overhead that small and medium sized businesses should have access to. So yes, I'm all for it.

I will say I'm not a lawyer, I'm a product person. But I definitely tell ChatGPT you know, I just bought a home, I have warranty, I will upload every single doc I have on every warranty and every policy in my state and, and ask it to be a brilliant lawyer and find different things, give advice. It's great.

I'll add one piece to give you a little bit of comfort. You know, like many teams, we are, we are as a legal team within HubSpot, trying to identify the areas where we deliver the most value and where we, we as humans do not deliver the most value. We are building our own versions of GPTs to help our business. Answers now like, are they going to be better than more accurate than ChatGPT for you? Maybe slightly. You know, but. But this is something that I think we all have to accept that there are areas where we can add a lot of value as humans, whatever our specialty area is, and there are other areas where AI is actually well suited to keep people moving fast and at scale. All right. Okay.

I think we just hit time, but we'll be around if you have more questions.

Awes.