Scribble: An AI Agent for Web Mapping – Beyond the Chatbot Wrapper
Jonathan Wagner, CEO of Scribble Maps, is back on the podcast, and this time we’re talking about Scribble—an AI agent he’s built into his platform.
AI agents are appearing in web mapping platforms, but not all AI implementations are created equal. The difference between a chatbot and a true agent lies in how deeply the AI integrates with the underlying platform and what actions it can take on behalf of users.
Chatbots vs Agents: More Than Semantics
Most AI features in software products function as conversational interfaces—chatbots that answer questions and provide information through text. Agents go further by taking actions, manipulating data, and interacting directly with platform tools.
Scribble, an AI agent built into Scribble Maps, demonstrates this distinction with access to 140 tools. It can view maps, select tools, build plugins, fetch data, and send emails. These capabilities extend beyond question-answering into direct manipulation of the mapping environment.
A perception problem exists in the market. Studies show that adding AI to products can make them appear “cheap” to users, particularly when implementations amount to thin wrappers around existing large language models. Users have grown skeptical of “AI-powered” labels after encountering forced integrations that don’t provide meaningful utility or can’t be disabled.
Building an AI Agent: Technical Decisions
Scribble runs on Google’s Gemini via Vertex AI rather than OpenAI’s models. Cost efficiency drove part of this decision, but native voice capabilities played an equally important role. The agent includes full voice interaction—users can speak directly to it rather than typing commands. This voice-first design creates a different user experience than text-based interfaces offer.
The implementation currently covers approximately 80% of Scribble Maps’ tools, with ongoing expansion. The agent interacts directly with the interface, manipulates data, and executes complex workflows that would traditionally require multiple manual steps.
Privacy in Agent-Based Systems
Privacy considerations become more complex when AI systems can take actions rather than just provide information. Vertex AI doesn’t train on user data, which addresses baseline privacy concerns, but the system still requires some data transmission to function effectively.
Data minimization principles guide the implementation. The system transmits data to the AI model in only two specific scenarios:
- When building plugins or filtering data, only attribute column names are sent—not the actual data values
- When users explicitly enable screen viewing, images are transmitted for visual analysis
The screen viewing feature requires explicit user permission and operates similarly to video conferencing, with clear indication when active. This balances functionality with data protection.
The Model Context Protocol: Modular AI Capabilities
The Model Context Protocol (MCP) represents a shift in how AI systems access external functionality. Rather than building monolithic agents with all capabilities embedded from the start, MCP enables a modular architecture where tools can be connected as needed.
MCP functions like “Lego blocks” for AI agents—discrete tools that can be added, removed, or swapped without rebuilding core systems. Developers create tools following the protocol specification, and AI agents can discover and use these tools through standardized interfaces.
The QGIS MCP server exemplifies this approach. It potentially allows AI agents to interact with QGIS functionality without requiring direct integration between every AI platform and QGIS. Users can bring their own tools into AI workflows, customizing agent capabilities based on their specific needs.
This modular approach has significant implications for geospatial tools. Instead of each mapping platform building isolated AI features, interoperable tools could enable richer functionality with less redundant development. An agent could access QGIS processing tools, PostGIS databases, and web mapping APIs through a consistent protocol.
Risks of Integrating AI into Established Products
Adding AI to production software introduces several distinct risks that extend beyond technical implementation challenges:
Perception Risk: Users may assume any AI feature is just a chatbot wrapper, missing deeper integration and functionality. Demonstrating real capability differences becomes a marketing and education challenge.
Product Devaluation: Research shows that AI-branded features can make products seem cheaper or less valuable, particularly when implementations feel forced or unhelpful. The proliferation of superficial “AI-powered” labels has created skepticism.
Expectation Management: Users familiar with general-purpose AI assistants may not understand domain-specific capabilities or limitations. They expect ChatGPT-like flexibility while encountering constraints designed for specific workflows.
These risks explain why some companies hesitate to prominently feature AI capabilities despite significant development investment. The challenge isn’t just building functional AI—it’s overcoming market perception shaped by poor implementations elsewhere.
Early Usage Patterns and User Segmentation
Early adoption data for Scribble shows a clear pattern: newer platform users adopt the agent more readily than experienced GIS professionals. This raises questions about the primary value proposition of AI agents in geospatial platforms.
Are AI agents primarily onboarding tools that help new users navigate unfamiliar interfaces? Or will they eventually support advanced workflows that experienced practitioners rely on daily? The answer likely depends on how agent capabilities evolve and whether they move beyond task automation into genuine augmentation of expert work.
The business case for Scribble emerged from an operational need: automating customer onboarding and education previously handled by staff. This practical driver—rather than following industry trends or competitive pressure—shaped development priorities and feature sets.
Two Value Propositions: Democratization vs Expert Augmentation
AI agents in geospatial work offer two distinct value propositions that serve different user needs:
Democratization: Enabling users without GIS expertise to accomplish mapping tasks through natural language interaction. This expands the user base for geospatial tools beyond traditional GIS professionals by lowering technical barriers to entry.
Expert Augmentation: Helping experienced practitioners work more efficiently, explore alternatives, and validate approaches. This focuses on productivity gains and enhanced capabilities for users who already understand GIS concepts.
Expert augmentation may be underexplored compared to democratization use cases. Experienced users can leverage AI to test ideas, identify edge cases, and work through complex problems more quickly than manual exploration allows. The agent becomes a tool for rapid iteration rather than a replacement for expertise.
The tension between these value propositions affects product design. Features that help beginners (extensive explanation, guided workflows) may frustrate experts who want quick execution. Balancing these needs while serving both audiences presents design challenges.
The Overuse of Maps in Decision-Making
Maps get used when they’re not actually needed—when questions could be answered with simple yes/no responses or numbers rather than visualizations. The visual appeal of maps can obscure whether they’re the most efficient tool for a specific decision.
AI agents could address this by providing direct answers to geospatial questions without requiring users to interpret maps. A query like “Which store location is closest to demographic target X?” might return coordinates and distance rather than displaying a map. The geospatial analysis happens in the background, but the interface focuses on the answer rather than the visualization.
This suggests a future where geospatial capabilities become embedded in applications without exposing users to mapping interfaces. Location-based logic operates invisibly, triggered by questions that don’t explicitly mention mapping or geography.
Invisible Geospatial Integration
The next evolution of geospatial AI may involve capabilities embedded invisibly in broader software systems. Users wouldn’t “log on to a mapping platform” but would instead ask questions that trigger geospatial analysis in the background.
This model relies on protocols like MCP to connect AI assistants to geospatial functionality on demand. A user working in a business intelligence tool might ask about regional performance differences. The AI agent calls geospatial analysis tools through MCP, performs spatial calculations, and returns results—all without the user seeing a map or knowing they triggered GIS operations.
This architectural approach separates geospatial capabilities from geospatial interfaces. The tools remain specialized and powerful, but access becomes abstracted through natural language and AI agents. Geography becomes a dimension of analysis rather than a separate domain requiring specialized software.
Technical and Strategic Implementation Challenges
Building functional AI agents for geospatial platforms involves several technical hurdles:
Hallucination Management: AI models generate plausible but incorrect responses. When agents take actions rather than just providing information, verification mechanisms and error handling become critical. An agent that confidently creates incorrect spatial joins causes more problems than a chatbot that gives wrong answers.
Tool Selection: With 140 tools available, agents must correctly interpret user intent and select appropriate functionality. This requires sophisticated natural language understanding and context awareness beyond general-purpose AI models. Domain-specific training or fine-tuning may be necessary.
Cost Management: Running an AI agent accessible to thousands of daily users requires careful resource planning. Model selection affects costs significantly—Gemini offers advantages over some alternatives, but even efficient models accumulate expenses at scale.
User Education: Communicating the difference between simple chatbots and functional agents remains challenging. Users arrive with preconceptions about AI capabilities shaped by their experience with ChatGPT or Claude. Setting appropriate expectations requires clear documentation and onboarding.
These challenges extend beyond technical implementation into product strategy and user experience design.
Industry Patterns and Market Evolution
AI agents in geospatial platforms represent test cases for broader industry trends. Several patterns are emerging as more companies explore similar functionality:
Voice interfaces may become standard features rather than optional add-ons. Natural speech provides a more intuitive interface for spatial tasks than typing commands or clicking through menus.
Modular architectures using protocols like MCP enable faster iteration and customization. Rather than waiting for platform vendors to add features, users can extend agent capabilities by connecting their own tools.
Privacy-conscious implementations using services like Vertex AI address enterprise security concerns. Organizations with strict data policies can deploy AI agents without data leaving controlled environments or being used for model training.
The distinction between “AI-powered” marketing and meaningful agent functionality will grow more important. Users will develop better intuition for differentiating substantial implementations from superficial integrations.
The geospatial industry faces fundamental questions about how AI agents coexist with traditional GIS workflows. Will they replace map-based interfaces for common tasks? Will they remain primarily onboarding tools? Or will they evolve into sophisticated assistants that experienced practitioners rely on daily?
Data Access: The Limiting Factor
Much of the best geospatial data remains private rather than publicly accessible. This creates challenges for democratization efforts, as AI agents can only work with data that users can legally access.
Data quality and availability may ultimately constrain what AI agents can accomplish more than the technology itself. An agent with perfect natural language understanding and tool selection still can’t perform analysis on data it can’t access. Proprietary datasets, licensing restrictions, and data costs create barriers that technical improvements won’t overcome.
This reality suggests that democratization has limits determined by data economics rather than AI capabilities. The most powerful analyses may remain restricted to organizations with access to premium datasets, regardless of how sophisticated AI agents become.
Open Questions and Future Development
AI agents for geospatial work remain in early stages. Scribble’s 140-tool integration represents one architectural approach, but many questions remain unanswered:
What’s the optimal balance between general-purpose AI models and domain-specific fine-tuning? How should agents handle uncertainty and ambiguity in spatial queries? When should agents request clarification versus making assumptions? How can implementations verify spatial analysis correctness before taking actions?
The Model Context Protocol and similar frameworks suggest a future where geospatial capabilities become more modular and accessible across platforms. Rather than each company building isolated AI features, interoperable tools could enable richer functionality with less redundant development.
For geospatial professionals, understanding how these agents work—their capabilities, limitations, and architecture—becomes increasingly important. AI agents will become more prevalent in mapping platforms and GIS workflows. The question isn’t whether they’ll arrive, but how to work with them effectively when they do.
Related Resources:
In Conversation
What Is Scribble Maps?
Daniel: You’re the CEO and founder of Scribble Maps, and you’ve been on the podcast before. This time I want to talk about the agent you’ve built into the platform. First, some context — what is Scribble Maps, who’s it for, and how long have you been working on it?
Jonathan: Thanks for having me back. Scribble Maps has been around for 15 years. We’re a horizontal, multi-use Web GIS platform with roughly 10 to 15,000 users a day.
Daniel: What are those users doing?
Jonathan: A lot of different use cases — site selection, drafting, sales territory mapping, construction, all the usual suspects when it comes to GIS. The platform is targeted at mid-range data, roughly the 20 to 200 megabyte range, rather than really big data. By and large our biggest users are small and medium-sized businesses — sometimes just an individual doing some mapping, though we do have enterprises with 50 or 60 seats.
Why Build an Agent — and Why Now
Daniel: You call the agent “Scribble.” Why build it, and why now? Was there a technology shift, or were people asking for it?
Jonathan: It’s a little interesting. A couple of months ago I found out my wife was expecting, and she does a lot of the onboarding and customer education. So I sat down and asked myself: can I have AI do this, or do I need to bring on another employee? I ran some tests and thought, I think I can do this — I think I can make something quite intensive. It’s still early innings, and it could be a huge risk for a variety of reasons, but I’m pretty confident it’s going to have a huge impact.
Chatbot vs. Agent: The Perception Risk
Daniel: What are the risks?
Jonathan: One of the biggest is that a study found including AI in your product is perceived as making the product cheap. Which is interesting, because a lot of companies are slapping “AI-powered” on everything despite the studies saying you really shouldn’t. Big companies jamming AI into their systems where it isn’t useful — or where you can’t disable it — has poisoned the well, so to speak.
Daniel: Is there a danger people think this is just a thin wrapper around OpenAI’s model?
Jonathan: It’s a huge risk, and one of our biggest challenges. A lot of the AI out there is just chatbots, whereas Scribble actually interacts with the interface, selects tools, pulls in data — it can even email you information about your map. Most chatbots don’t do that. People assuming it’s just a chatbot is a real struggle, because if they make the mistake of thinking it’s a wrapper on top of ChatGPT, they’ll just go use ChatGPT instead. One difference — I use Gemini, not GPT — is that it talks to you. It has a full voice; you can speak directly to it rather than typing. That, and it’s cheaper than OpenAI, is why I went with Gemini.
What Scribble Can Actually Do
Daniel: Tell me what it can do. Does it have access to geoprocessing?
Jonathan: It has access to almost all the tools. Say “I want to use the draw tool” and it selects it. Say “filter on an attribute greater than 1,000” and it does that. Say “I want all the building footprints for a polygon” and it fetches them; ask for the population of a polygon and it fetches that too. It has 130 — actually 140 — tools, and because of AI programming I’m adding two to five a day based on user feedback. What’s interesting is it sometimes surprises me. Early on, before I’d built a tool to switch between imperial and metric units, a user asked for it — and Scribble went to our documentation, figured out the setting lived in the UI settings panel, and used the tool it did have to open that panel. It went to the docs, worked something out, and satisfied the request with the tools it had.
Daniel: Does it use MCP — connecting to those tools through a Model Context Protocol server?
Jonathan: It doesn’t use MCP at all — it’s probably one of the biggest client-side implementations. It goes to Gemini, Gemini has the config and sends back a tool call, and that call is executed inside the app. I am building the ability to add your own MCP servers, though, so you could create a specialized tool and bring it into Scribble. We’re hoping Scribble can be the front end for a lot of these geospatial tools.
Reliability, Hallucination, and the Multilingual Trap
Daniel: How do you deal with quality assurance — knowing it’ll do the thing it’s supposed to?
Jonathan: Reliability is a huge issue. I’ve improved it by 30 or 40%. One specific lesson: you have to force the agent to always speak in English, because your tools are defined in English. Something as simple as “respond in the language the user talks to you” can interfere with tool usage — the model decides it can’t call English-named tools. So we do multilingual support only on the front end. As for correctness — “draw a red polygon” should change the colour to red and select the polygon tool. It does that maybe 90% of the time; the rest of the time it might set the colour and then just tell you how to select the tool. You can never fully stop that, so we focus on getting it right 80 to 90% of the time, at least in the ballpark otherwise, and we invest real time when something is way off.
Daniel: Have you had major problems with hallucination?
Jonathan: Yes. Early on, with fewer tools, it would go off on tangents guessing how the UX worked — completely wrong. Counterintuitively, the more tools I add, the lower the hallucinations, because it’s more likely to pick the right one. Defining limitations is just as important as defining capabilities — AI can’t really draw for you, so if someone says “trace this on my screen,” Scribble says “I can’t draw for you, but I can give you the tools to do it.” And samples beat logic: providing examples of how it should respond is far more valuable than trying to program it from a logical standpoint.
A Multi-Agent Architecture
Daniel: Is this one agent, or agents talking to agents?
Jonathan: Scribble is the conductor, and there are five or six other agents — one for map styling, one for building plugins, one for looking up documentation. You split it up because of context and cost: building a custom plugin pulls in a ton of examples and documentation, and you don’t want all of that sitting in the main context. But the deeper agents need a running memory, otherwise you get a telephone effect — and agents can even get stuck in loops, sending information back and forth. There’s a real limit to how big and complex an agentic network can be.
Why LLMs Can’t Touch Spatial Data Directly
Daniel: Some people say spatial is special. Have you noticed anything special about implementing this in a geospatial system?
Jonathan: You can’t have these LLMs operate on geospatial data directly — you can’t rely on it; it will hallucinate and break. You’re never going to want the AI to perform an operation on the actual spatial data. Ultimately your AI agent is a router for natural language input — it decides which tried-and-tested spatial functions to fire, and in which order. For “all the points within 50 metres of this line,” it creates a buffer around the line and filters on the buffer. We’re building a database of samples that outlines those multi-step processes. It’s the same with conversions — when Gemini converts a GeoJSON to KML, it writes Python code and runs it in an isolated environment; the LLM itself doesn’t do it.
Advice for GIS Professionals
Daniel: Having built a system like this, what’s your advice to other geospatial professionals — panic, or think about a few things?
Jonathan: Did I come on a GIS podcast to say you’re all doomed? For all its intelligence, if you don’t tell it and guide it, it won’t necessarily do the right thing — it can go down really wrong rabbit holes. I wouldn’t be super concerned right now. These tools help people get up and running very quickly; I see them as educational and as assistants, not replacements — kind of like the nurse to a surgeon. One thing I didn’t expect: Scribble turned my feedback faucet back on. When it can’t do something, it asks the user if they want to submit feedback, and it drops straight into a Slack channel. I went from getting very little feedback to five or ten useful suggestions a day.
Daniel: What are you most excited about when you think about AI and geospatial?
Jonathan: Scribble Maps’ mission was to make mapping accessible, and this drops the learning curve significantly — if you show up, you can just ask, and it can walk you through. I’m excited about the capabilities that open up to the average person who wasn’t in GIS before. But a lot of the opportunity in the future will be geospatial integrated invisibly into broader products — helping people make choices without them logging on to a mapping platform at all. And often maps are used when they’re not needed; that’s one place AI can help a lot, making those decisions without a map.
