How to build a system of learning with Maze MCP

Fiona

Everything is speeding up — product development, decision-making, all of it. I feel it in my day-to-day work, and I'm sure most of you do too. The question we're exploring today is: how does research keep pace with that? I'll hand things over to Noel to take it from here.

---

Noel

Thank you, and thanks to everyone for being with us today. I'm really excited to talk about this concept of making sure that research keeps up with the speed of development. I am a researcher at heart. I love geeking out about research, so I'm excited to be in this conversation today.

Before we get into the deeper conversation about MCP, I want to start with one of those challenges that I've had more times than I can count, and it goes a little something like this. Someone on a product team asks a question, they're requesting research, and someone else in the room says, "Hey, we actually did this same study six months ago." And another person in the room says, "We did? Where is that data?" I know that answer is somewhere in a drive folder, maybe organized by a date or a topic that doesn't really make sense anymore, and it's really hard to find that data.

Feel free to jump into the chat and just say yes, or "I've been there," or "I resonate with this" — just let us know if this is something you've experienced. As a researcher, I have definitely been there. This is something that, as researchers, we unfortunately have to deal with a lot. It's just part of life for a researcher.

And Patrícia and I are here to say: it doesn't have to be this way anymore.

What's really interesting about this scenario is that the research existed. Somebody asked the question, and you had already done the research. You put in the good work, you wrote the report, you took so much care, and you treated it with rigor. But those insights weren't accessible when they were needed. Maybe they were useful in the moment when you originally ran the study, but when those insights were needed for further conversation, or the issue resurfaced again, people couldn't reference that data point or use that data to drive their decisions.

What often happens is you either rerun the study, someone else on your team reruns the study, or more likely, the product team just moves forward without that information and makes decisions based on gut instinct.

So let's hold onto that moment — that feeling of frustration of "I've done this before, we're not referencing the data, we're not continuously learning from our research." And I want you to know that I truly believe this is not a you problem. It's not a resourcing problem. It's not a stakeholder problem. It's a structural problem. Research is sitting outside of the decision process. Those insights we're creating are not flowing into the areas where decisions are being made.

The good news is that a structural problem is something we can change, and that's really what we're going to be rooting our conversation in today.

On the next slide, we have a quote from our CEO, Joe, and I love the way he frames this: "The teams that win aren't the ones with the strongest convictions in the room. They're the ones that never stop learning." That's what research is all about — learning, discovering, and using that information to drive decisions.

This framing matters a lot right now because of AI. We've heard about AI a lot. It's changing the way we work as researchers, and it's changing the way our entire teams work. Teams are shipping product in days where it used to take quarters. Designers are prototyping in tools that didn't exist two years ago, and those new tools have made it really efficient to vibe code or prototype in a matter of hours instead of days. Because of this, the throughput of building has really exploded.

What that means is the bottleneck has moved. It hasn't gone away — you're no longer bottlenecked on whether or not you should be doing the work. You're now bottlenecked on whether or not that work is worth doing, whether the assumptions driving your roadmap are accurate, whether the feelings you have about the personas you're building for are true. When everybody can build anything, the quality of decisions becomes the scarce resource. And I think that's a really important opportunity for researchers, because we can help with the quality of those decisions.

Moving forward, I love this statement from Paul Graham: "If AI makes writing code a commodity, understanding users' problems becomes the most important part of building anything." Deeply understanding your users is your last source of differentiation. And the good news is that that is what research is all about. Research is the discipline of exactly that. It takes ambiguity and turns it into evidence. It takes "I think our users want this" or "I think our users do this" into something you actually know and can use to power really important decisions.

What we're seeing in our own research, and what I'm seeing in the market, is that research is becoming more and more influential over time. Research has three times more influence over business strategy than it did a year ago. We're also seeing that demand for research is up 66% in the last 12 months. But interestingly, one out of three companies are failing or really struggling to do research well. When you look at those two things together — demand for research is high, research is being used to make strategic decisions — and then you also see that companies are struggling to do research, we have to ask what's causing that.

What I think is happening is it's not a quality problem. We're doing good research. The problem is delivery. The research is being done, but those insights are not being funneled to the areas where decisions are being made.

As a researcher, I used to spend a lot of time caring about the methodology and the final report, spending hours creating a detailed deliverable — and then I'd come to a readout with maybe a handful of people in the room. Those insights stayed in that report. It got archived in a drive somewhere, and it stayed with the people who were in that room. It didn't often get socialized beyond that.

In systems that can't deliver insights quickly, decisions get made without that really good research. The research might have been there, but it wasn't in the location it needed to be to inform the decision. It wasn't accumulating over time. Research was being treated as a point-in-time solution. So six months later, when you need to reference that research again, people don't know it exists — and so the study gets rerun, or skipped entirely because nobody knew it happened in the first place. It's not a research issue; it becomes a delivery issue.

That's the framing for this entire conversation. We think about research as an event, but learning as a system — and that distinction is really important. An event happens once. Someone asks you to run a study, you run it, you do the analysis, you deliver the report, and you move on. Whatever it shapes depends entirely on the people in the room, or the people who read the report at that moment, and whether or not they remembered those insights.

A system, on the other hand, is continuously running. Insights find the people instead of waiting to be found. Every study makes the next decision smarter. The baseline grows stronger with every loop. The organizations that are winning are the ones that have made this shift — the ones that have realized they need to stop treating research as a service and start seeing it as the infrastructure for how the entire organization learns.

The researcher doesn't go away in this system. They're the ones operating it, defining how the organization learns, and their research is what's fueling that learning. This distinction between service and infrastructure really matters, because in a world of budget cuts and automation, services can get cut when budgets tighten — but infrastructure stays. That's why it's critical that we make that adjustment.

Alright, that was fun. Let's talk about how we help you with this adjustment — how does the MCP fit into that ecosystem?

MCP can be seen as that connective tissue. It connects your Maze research data directly to the AI tools your teams are already working with. I've talked about how we move your great insights into the areas where decisions are being made. Decisions are being made in tools — people are using tools like Claude, ChatGPT, and Cursor to help drive those decisions. We want to make sure your research is also part of that ecosystem.

For anyone who's non-technical like me: MCP stands for Model Context Protocol, and it's an open standard that lets AI assistants like Claude, ChatGPT, or Cursor securely access your data sources. Patrícia is going to go much deeper on how this works. But grounding us in this idea of MCP as connective tissue is really important — because once you connect it, you can start asking plain-language questions about your research, and those answers come back to you not based on assumptions, but based on real data done with rigor and care, surfaced right where people are making decisions.

The last thing from me is thinking about how this changes based on your role in the organization. I like to frame it in terms of pain points. If you're a researcher, the pain point you're likely experiencing right now is: "My findings end up in a doc that nobody reads." That's a really frustrating place to be — and it's also true. We sometimes get stuck in that loop. With MCP, your studies are searchable by everyone, without you chasing stakeholders to read the report or re-presenting it to new audiences.

Patrícia

That work already exists, and it's accessible. Every answer traces back to real participants, real studies that you ran. Your rigor is preserved and not diluted. You built the loop, so your answers are trustworthy. The answers from that MCP poll are trustworthy because you're the one who was driving that research.

If you're a PM or a designer, your pain might sound a little different. You might say, "I make decisions faster than the research can reach me, or faster than I can go and dig up those artifacts that I need to make decisions." The default becomes assumptions, because finding research has friction. It becomes challenging to find research. You have to go to an archive. You have to know the naming conventions. Not anymore.

With MCP, you find out what your team already knows instantly. You don't have to go hunt for it. You don't have to go search for it. You don't have to know where to find it. You can use the MCP to bring that information into where you're making decisions — get that information before sprint planning instead of after the shipment is announced, and before you've made those product changes.

And then most importantly, if you're a leader, the pain that you're probably facing is around lack of clarity about your research investment. You're funding studies, you're buying tools, you're supporting researchers, and you're questioning whether or not the research that your team is doing is actually influencing business decisions. With MCP, you're able to make sure that every study is always available through that MCP, which means that your research investment compounds over time instead of depreciating in a folder. That's really powerful because your team is doing good work, and we want to make sure everybody is benefiting, and we continue to learn on that really great work.

And that's the why — why we built it, what we believe in. I'm going to go ahead and pass it over to Patrícia to show what this actually means in practice.

---

Patrícia

Hi, everyone. I'm excited because I get to show you how you can make everything that Noel just explained a reality for yourselves.

Let's start by understanding MCP — and Noel already got a bit into that. MCP, as Noel said, is a connective tissue. You can see it as a standard, or the protocol for how you can connect your AI agent of choice to your data or tools in your ecosystem that you use for work every day. Here we're talking specifically about how you can use Maze as a tool, as the center of your research, directly in your AI agent, and that's what Maze MCP does — it connects the two things together. As Noel said, you connect it once, and from then onwards, your agent will be able to have access to all the Maze research that you have already conducted, whether you're a leader, a designer, or a researcher. If you connect to Maze MCP, you have access to the data that you can see when you go to the Maze platform.

In terms of the data that you can see, you're going to be able to see all the studies, the research, and the learnings that came out of them. This includes not just the insights and the learnings you got, but also the evidence behind those insights. As you ask questions to your agent about what you have in Maze, you're always going to be able to trace back what the evidence in the research was that brought you that specific insight.

Now, let's move to the next slide, and I'm going to start a live demo. But before we go into the live demo, I just want to show you how you can get started in four simple steps.

If you go to the next slide — yes. The first thing is, Maze MCP is currently available as a beta, and you can quickly join the beta. You just need to go to a link we're going to share in the chat, and you're also going to receive it in the follow-up email for this session. You can click the link, go to this landing page about what you can do with MCP, and register immediately. We are going to review the applications. At the moment, the beta is only available for enterprise customers, so if you're an enterprise customer, you'll get access and will receive an email informing you that you now have access to Maze MCP. Your CSM will also reach out to help with any questions, help you onboard, and unlock use cases for Maze MCP.

Noel already touched on this a bit. You just need to connect your AI tool, and then you can start asking questions — questions like Noel shared before about the research that you already have, how to run a report, how to create a report more quickly instead of taking a long time to analyze all the results.

So now let's get to the fun part. I'm going to start sharing my screen and walk through a demo. I'm going to use three agents at the same time so you can see the different experiences. We're going to be using some studies that we have just for demonstration purposes, so don't be too particular about the details of the studies we're sharing.

So where can you find Maze MCP before we start asking questions? When you get access, you just go to your settings — specifically personal settings — and then within the integrations area, you're going to be able to see everything about MCP. You're seeing my real account here, so you can see that I'm connected to all of the agents. You just need to click Connect.

One thing to know is that for Claude and ChatGPT, if you click Connect, you'll be directed into the apps of Claude and ChatGPT, and you can connect in one click because we got approved to be showcased in their marketplaces. For Copilot, Cursor, and other apps, you're going to be linked to our help center article, where you'll get step-by-step details on how you can connect. You just have to copy a couple of details into your own tools, and you'll be set up.

So now let's go to the agents. Just to see where you can connect, I'm going to showcase the example of Claude. If you go to Customize, you can go to Connectors, and if you browse Connectors for Maze, you're going to be able to see our Maze MCP there.

So now let's start actually asking questions. I'm going to pull in some prompts that I already saved that I thought could be helpful to walk through today. I'm going to start with probably a very common case for you, which is you just want to know what the key pain points are of a study that you ran. I'm going to be specific and ask Claude for the pain points ranked by criticality, with a supporting participant quote for each one so that I can really see what the participants of the study actually said.

I'm going to ask ChatGPT as well, and Cursor too. While they're processing the information from the study, I'll give you a glimpse of what the study looks like.

This is a dummy study, and it's about a Bridger app — an app where you can book your travel plans. Specifically, this is an AI-moderated study where we wanted to deep dive into what the specific friction points in the app are, what the trust factors are that influence use of the app, what the potential information gaps are that create uncertainty for travelers when they're planning their vacation through the app, and to evaluate potential usability challenges.

Let's see what our different agents have come up with. We can see here in Cursor that it has analyzed the different sessions and put together the key pain points, signaled the level of criticality for each, and — exactly as I asked — put together a quote, in some cases more than one, along with a link to the participant who said that specific quote so I can click and go to Maze to see the specific participant session.

Let's see if the other agents have come up with answers. Claude also produced a very similar structure based on the specificity of what I asked, and you can see the pain points here. And ChatGPT already has things here for us — a table with the different pain points, again ranked by criticality, the supporting evidence, and obviously the link so that you can always trace it back.

Now that we've deep-dived into a specific AI-moderated study, you can continue asking questions here in Claude. For the sake of this session, I'm going to ask our agents to deep dive into an unmoderated study about navigation. I'm asking Claude to summarize the key takeaways of a navigation study, and I've shared the link so that it goes straight to the study I'm looking to get answers from. I'm going to ask the other agents as well.

For context on the answers we're going to get, let me show you what this study is about. We're testing a new navigation, and we're comparing a hamburger menu navigation and a sidebar navigation. Where an AI-moderated or moderated study would give us transcriptions and quotes, an unmoderated study gives you more quantitative data in addition to open-ended answers and quotes. Let's see what our agents come up with.

Here in Claude, it has found that the tab bar navigation performed better than the hamburger menu navigation in this study. You can see how the different prototypes compare across each of the metrics, and here are the key takeaways — with links, as you can see, to the source of each insight and the specific evidence.

The other agents are going to run the same thing. ChatGPT produced a very similar assessment and recommendation. Cursor is asking me to authorize something, so I'll let it run while we jump to the next study.

Because this is a fairly small unmoderated study, I wanted to raise the bar and go a little further with what an unmoderated study could be. I'm going to pull in a different unmoderated study and ask Claude, then open a separate conversation in ChatGPT for clarity.

Cursor had already analyzed the study in the meantime. We have here an unmoderated study, and I didn't give it much context — just the URL. If we go to this specific study, we can see it has a more complex setup. This is a usability testing study where we're trying to understand the best way to showcase the instant booking feature in our Bridger app. You can see that we have multiple questions, a variant comparison between a version with a badge only and a version with a badge combined with a filter, and additional questions about travel — for example, what categories you'd normally search for and what features would be most helpful in this type of app. So there's a lot of data to go through.

If we go back, we can see in Claude that it has come up with the key takeaways. The headline was that adding a filter to the instant booking badge resolved the task completion issue. We can see that the completion rate increased from the badge-only variant to the badge-plus-filter variant, along with additional evidence and insights from the study.

But this is a lot of text, and sharing it in a Slack post, a Teams post, or a document would likely be where it stops. Since this is in the Maze platform, anyone can get access to it as long as they're connected to Maze MCP and have the appropriate permissions — so it won't be locked here. But imagine you actually want to share this with executive stakeholders in a presentation.

I'm going to ask Claude to go further and aggregate these results into a visual, easy-to-digest report targeted at an executive audience, with the appropriate level of detail. I've also provided hints on the format I'd like — headlines, supporting metrics, and recommended next steps. You can see that Claude is already pulling in all the data. I'm going to ask the other agents as well, if you want a glimpse of how they visualize things differently.

Let's see how Claude is running. You can see it's following the structure I asked for — the headline finding, the—

Patrícia

...the quantitative data that we have to back up this finding. You can see here what the success rate was in each option, what are the features that users said are a must-have, and clear recommended next steps. And obviously from here you can then share further.

Let me see if the other ones — normally it's a bit of a heavier task for the agents. I think maybe we are okay with Claude. As you can see — of course, it is going to come up with the canvas so I can show. You can see here how those reports that could take long hours, now you can easily do them based on the research. You don't have to export or copy-paste anything. You just need to use your agent and ask for whatever you need. I'm just going to pull up this canvas from Cursor just so that you can get a glimpse, because it's a slightly different structure, but following what I asked — having a headline, having the evidence, the quantitative data to back it up, and finishing here. This is a longer one, but it has a list of recommended next steps.

I'm going to stop sharing now and go back into the presentation so that I can tell you what's coming up next.

The first thing is marketplace listings. What this means is having Maze available in your agent so you can connect in one click. As I showed you in Claude, you can already go to customize and you'll see Maze there. For ChatGPT, it's already the same thing. For Cursor, Codex, and Microsoft Copilot, for now you have to follow a guided tutorial that we have — just a few steps, it's pretty easy. Soon enough, we hope they will get us approved and you'll be able to connect in just one click.

After that — and this is something we're currently already working on — is making cross-study analysis better. When you ask questions that are going to analyze multiple studies with an infinite number of participant sessions, we want to make sure that you get accurate answers to your queries. We're optimizing to make sure that you get the proper level of detail when doing a cross-study analysis.

Finally, the later next step is allowing you to create research directly from your agent and generate it automatically in Maze. This will mean that you can ask your assistant to create the study with our expertise, and then it will be automatically created in Maze.

This is it for the product deep dive today. I believe we're going to move forward to some Q&A, and we can answer the questions that you've been adding throughout this conversation.

---

Noel

Yeah, thank you so much. One quick note to be clear on access today. The data is open to current Maze customers on enterprise plans. I've seen a couple of questions in the chat about access, so I just wanted to address that specifically. If you are a customer on an enterprise plan, just know that when you go to request access, you're not just filling out a form into the void. We personally manage each application, follow up, and our team reaches out for a walkthrough to make sure that you've been able to set up your first connection successfully.

If you're not on an enterprise plan — and maybe you're not even a Maze customer — go ahead and apply anyway. That's the fastest way to start a conversation about what access might look like for you or your team, and to kick off that conversation.

I just wanted to cover that quickly before we jump into questions. It looks like we've had quite a few come through the chat, so thank you so much for taking the time to write those in. Let's take a quick look at where to start. I'll stop sharing so I can present some of these questions.

One question that I'm seeing a lot is about connecting with specific tools such as Copilot. Patrícia, feel free to jump in, but I will say — the way that MCP works is that any client that can connect to an MCP server can connect to Maze, which includes Copilot, Studio, and 365. It just has to be set up with a manual link at this time. I don't believe we have marketplace access today.

---

Patrícia

Yes. For now, we don't yet have it available on Marketplace, but for Copilot specifically, you can follow the step-by-step guide we have on our help center. You'll probably need help from your IT team, but it's quite a quick setup to make. We already have some users in beta using it in Copilot, so you should be good to go.

---

Noel

And the same goes for questions around any tools that we didn't mention — Gemini is the other one I'm seeing in the chat.

---

Patrícia

When you go into our help center article, you can see the key basic data needed to set up an MCP. In this case, it's — not to get too technical — at least two elements: the connection URL and the authentication method. We have those written in our help center article, so for any tool you're using, you can create a customized connection and just add in these URLs.

For example, to speak a little more to Gemini: to set up Gemini, you'll likely need your IT team, similar to Copilot, because it requires admin access for the whole Google Suite. But with the information in the help center article, you can already do that setup. Unfortunately, for the moment, Gemini doesn't allow a single user to add a customized MCP — it has to be done at an admin level, so your IT team should be able to handle that for you.

---

Noel

Another question we have here is whether it's possible to ask broad questions that could be answered by one or more studies, and to ask them without directly referencing those studies. I can take a stab at answering that. Patrícia, feel free to jump in, but I think this touches on the cross-study analysis piece that we're working on improving — the ability to ask something more general or broad, and then MCP can go ahead and try to find the relevant studies to answer that question.

One best practice that we've discussed is dropping in the link to a study, which helps you get to the right answer faster. But the goal is eventually to be able to do this successfully without that.

---

Patrícia

For example, if you say, "I want to understand the key takeaways from the studies we've run around our onboarding," you'll probably have several studies on onboarding, not all of them relevant to the specific part you're asking about. It's easier if you put in the URL, but if you don't, your agent will likely say, "Here are the studies I found — do you want me to cover all of them, or would you like to select some?" And from there you can continue the conversation. You'll have more back-and-forth in the conversation, but it's possible. Yes.

---

Noel

"What is your view of the role of a UX researcher in light of tools like this? Can a PM now do the work of a researcher?"

This is a great question. What we really love about the MCP is that it makes research accessible at any stage, regardless of your role. Whereas research could possibly have been skipped or deprioritized in the past, it gives you the ability to find out whether research already exists to answer a specific question — and if not, take the steps to make it happen. It's definitely relevant whether you are a PM or a UX researcher. It gives either role the ability to make a research-backed decision, and even to work together to make that happen. We really like that the MCP allows you to surface the right information while still keeping a human in the loop, so that those decisions are made at the human level.

---

Fiona

Yeah, I can jump in here too. One of the things that's important to remember — going back to what I talked about earlier — is that when we think about moving from a single event of research to building a system of learning, the researcher is really meant to be the architect in that ecosystem. We're defining the rules of how research gets done, we're setting the rigor standards, and we're doing a lot of that sophisticated research work that is ultimately being fueled into what the MCP is pulling out.

That also means we have others — people who do research — following our lead and wanting to answer questions themselves. I don't think it's a this-or-that situation. I think it allows us to ensure that as an organization we're learning, and that we're learning on factual information we feel really confident in. I actually think the role of researcher is even more important right now, because what we're essentially doing is giving access to a lot of people, while making sure that access is credible. And that's our goal.

---

Patrícia

"How do you avoid people from other teams getting incorrect answers if they don't understand the context of a specific study?"

I think there are two ways this can be addressed. First, in the same way that one person has access to that knowledge, everyone else in the organization will have access too. It's very easy to have fruitful conversations about research that has already been done, and that just proves the case for getting research even wider in the organization, rather than having it closed in a document that no one will see.

The second part is that one thing we made sure was part of the Maze MCP is that you can always trace back. As I was showing you, I was going back and forth between the study itself and what I was getting in the agents, pulling the links directly from the agent output. You can always trace back where that source of information came from. You can understand what questions were asked in the study, so you can always get back to what led you to that learning and mitigate those kinds of situations.

I don't know if anyone else wants to add something — Noel or Fiona, feel free to contribute.

---

Fiona

Yeah. I think you nailed the big pieces — that everything is traced back to a real participant and not generated. With AI tools, there's often the fear that output has been hallucinated. But with Maze MCP, and a lot of the testing that has confirmed this, everything comes from your actual Maze workspace, traceable to a real participant, to actual data, to actual insights. It really solves for the fear of the LLM running with an assumption, by giving you a data-backed response.

---

Patrícia

"Do PMs have to have a seat in Maze in order to use the MCP?"

Yes. For users to have access to the Maze MCP, they have to have a Maze account on the specific team they want access to, and they have to have permissions to access that given team. Everything that applies to a user in Maze will be replicated in the Maze MCP. So yes, that will be required.

---

Noel

"When I'm referencing research, I know newer research is more relevant, but there is value in older, more high-level research. With Maze MCP, is the context given so I know a user's opinion is five years old versus a month old, or if feedback was given in passing rather than as an answer to a specific question?"

---

Patrícia

On this one — you can ask specifically, when you make your query, something like "Show me the pain points of our onboarding and how those pain points have evolved over time." You're able to distinguish between older and newer research. You can also ask, in the same way you'd ask for key takeaways, to include the time the research was conducted — or even when the participant provided that specific piece of information.

The great thing about our agents is that you can continually iterate on the information you're seeing. If you have more questions about the context, you can ask. For example, if you got an insight that your onboarding is bad because the form is very long, takes a long time, and requires checking multiple fields — leading participants to give up halfway through — you can follow up by asking, "Can you provide when that feedback was given and the circumstances in which it was given? Was it explicitly asked for, or did the participant mention it while answering something else?" You can always deep dive to get additional context. And obviously you can always trace back, but you can also ask your agent if you want to do everything within your tool.

---

Fiona

We're just combing through the list of questions. It looks like we have a lot of duplicates, but if we didn't get to yours...

Please follow up with us. We will be sending out a follow-up email where you can find information on how to reach us and ask questions.

We will wrap up a little bit early for now. That's all from us today. Thank you so much for joining, for all the great questions, and for making this a really fun conversation. We're so excited about MCP and where we're going, so we appreciate you all taking the time out of your day to hear a bit more about it.

Keep an eye on your inbox for that recording and those resources. We hope to see you all at another Maze event soon. Thank you.