Speaker A
I want to start with a question, and I want you to answer it honestly in the comments. Drop a one-word answer as I'm talking. What is the most important decision your company is making right now that depends on something happening
Factori reveals the blind spot in AI: lack of real-world data integration and launches Factory AI to ground AI models with live physical-world data.
Ask about this video. Answers come from its transcript only — with the timestamp, so you can check them.
Generated from the transcript and can be wrong — check the timestamp.
Key Takeaways
- AI models today lack access to continuously updated real-world data, limiting their accuracy for physical-world decisions.
- The blind spot is a data problem, not an intelligence problem, requiring integration of real-time physical signals.
- Factory’s real-world graph provides a comprehensive, multi-layered data foundation updated hourly.
- Factory AI’s MCP server bridges AI agents with live real-world data, enabling grounded and reliable AI outputs.
- Grounded AI reduces costly hallucinations and improves enterprise decision-making in retail, real estate, and marketing.
What the video covers
- Current AI models like GPT and Claude excel at reasoning but lack access to real-time physical world data such as foot traffic and neighborhood changes.
- This data gap causes AI to hallucinate confident but inaccurate answers for real-world business decisions, especially in retail site selection.
- The real-world graph built by Factory integrates 12 data layers with 90 billion signals, continuously updated to reflect physical world changes.
- Factory AI introduces an AI-native access layer via the MCP server, enabling AI agents to query real-time physical data directly.
- MCP (Model Context Protocol) is an open standard that connects AI models to live internal tools and data sources for grounded reasoning.
- This integration eliminates stale data pipelines and allows AI agents to provide fact-based, actionable recommendations.
- The launch includes multiple front doors for accessing the real-world graph, including APIs, conversational web apps, and agent frameworks.
- Factory AI’s real-world grounding improves decision accuracy in areas like site selection, demographic analysis, and competitor mapping.
- The solution is validated by a 92% higher accuracy when cross-referencing multiple data signals versus single-source data.
- The video encourages viewers to engage by sharing their data vendor usage and AI frameworks, and invites early access participation.
Chapters
- 00:00Introduction and Question to Audience
- 01:14Limitations of Current AI Models for Real-World Data
- 02:11Why AI Models Hallucinate with Confidence
- 04:03Introducing the Real-World Graph and Data Layers
- 05:12Launching Factory AI and the MCP Server
- 06:20How MCP Connects AI Agents to Real-Time Data
- 07:24Demonstration of Grounded AI Responses
- 08:24Multiple Access Front Doors and Use Cases
- 09:28Summary and Call to Action for Early Access
Full Transcript — Download SRT & Markdown
Speaker A
in the real world? Where to open, where to close, where your customers actually are, where your demand is moving, where competitors are gaining ground. Drop it in chat. One word, [music] one phrase, and I'll come back to those answers
Speaker A
later in the session. Hi, I'm Amog. I head platform products at Factory. I'm now going to walk you through what I think is a single biggest unsolved problem in enterprise AI right now, and the launch we are making in early access
Speaker A
to fix it. Let me show you the problem first. Not a slide, live. [music] I'm going to ask the most generic question any retailer, real estate firm, or a marketer would ask in 2026. Where should I open my next store in Dallas? I
Speaker A
want neighborhoods with the highest foot traffic from women 25 to 44, mid to upper income, and low competitive density. Give me three ranked recommendations. [music] Look at this response. It's articulate, it's confident, it names Uptown, Bishop Arts, Legacy West. It throws in some
Speaker A
demographic theory. It reads like an analyst. But watch what it cannot tell me. It cannot tell me how many people walk past 3215 Oak Lawn Avenue last Tuesday. It cannot tell me whether foot traffic in Bishop Arts went up or down
Speaker A
after the mixed-use development opened in Q1. It cannot tell me whether those 25 to 44 women are driving in from Plano or walking from two blocks away. It cannot tell me what the competitive landscape on that specific block looks
Speaker A
like right now. It has no idea. Not because the model is bad, these are extraordinary models, but because the data does not exist on the internet. It never did. The physical world, how people move, where they actually spend,
Speaker A
what's actually happening and closing in on a given block, how a neighborhood is actually shifting, none of it gets indexed by Google. None of it gets scraped into your training data. It updates by the hour, and every AI you're
Speaker A
using today is blind to it. That blind spot is entirely the reason why we built what we're launching. I want to be precise about why this happens, because the precision matters for what we're doing about it. This is not a model
Speaker A
intelligence problem. The frontier models, Claude, GPT, [music] Gemini, they reason brilliantly, they synthesize, they hold complex context.
Speaker A
The problem is upstream of the model. It is a data problem. Models are trained on what's been written down, articles, books, code, conversation. That corpus is enormous, but it has a hard boundary.
Speaker A
On one side of that line, everything ever written about the world. On the other, everything that actually happens in the world. People moving through cities, foot traffic surging around new development, a neighborhood's demographic profile shifting in real time, a competitor opening three blocks
Speaker A
away, trade areas reshaping after a transit change, property values moving with mobility. None of that gets written down anywhere a model can read. And here is what makes it worse. It changes constantly. A training cut off from 6
Speaker A
months ago is already stale on anything physical. A retailer making site selection call based on what GPT or Claude knows about Dallas neighborhoods is making the call on data that does not exist. This is what we call hallucination with confidence. The model
Speaker A
is not fabricating from nothing. It is making plausible extrapolations from text and presenting them as fact. For chatbot use cases, that is annoying. For real-world business decisions, that is expensive. AI does not say, "I have no foot traffic data for Dallas." It says,
Speaker A
"Here are three neighborhoods to consider." And that confident structured answer is exactly what gets copy-pasted into the board deck. This is a structural gap on how enterprise AI is being built right now. Closing it requires something specific, real-world
Speaker A
data delivered in a way AI can natively use. Here is what most people don't realize. The data does exist. It has existed for years. For the last several years, Factory has been building what we call the real-world graph, a
Speaker A
continuously updated global intelligence layer [music] that captures the physical world across 12 integrated data layers.
Speaker A
This is not a static database. It runs on 90 billion data signals across countries and it updates continuously.
Speaker A
When foot traffic shifts, we see [music] it. When a neighborhood changes, we capture it. When a new POI opens or closes, it is reflected. Here is a proof of proof point that matters.
Speaker A
Multi-signal intelligence validated by cross-referencing 12 overlapping data layers produces 92% higher accuracy than a single source location data. That isn't a marketing claim. That is what happens when foot traffic gets cross-validated against economic indicators, retail sales, property data,
Speaker A
instead of being clustered on its own. Quick poll for everyone in the chat. How many separate data vendors does your team currently stitch together to answer physical world questions? Drop the number. I'm guessing the median answer is three. We will come back to this. For
Speaker A
years, enterprises have accessed the real-world graphs through APIs, cloud delivery, our data platform, analysts pulling trade area reports, data teams enriching CRM records, AdOps team building audience segments. [music] That works. It scales. But something fundamental has changed about how
Speaker A
enterprises consume data. The center of gravity has shifted from analysts pulling reports to AI agents making decisions. And that shift created a new problem. The data existed, but it wasn't connected to the AI stack. That is what changes today. This is what we are
Speaker A
launching, Factory AI, the AI native access layer to the real-world graph. The hero of that launch is our MCP server. If you have not worked with MCP yet, Model Context Protocol is the open standard originally proposed by Anthropic and now adopted across the
Speaker A
ecosystem that lets AI models call internal tools, data sources in real time. It is a connective tissue between agents and the system they need to reason over. I want to be clear about that claim because we don't make
Speaker A
first-mover claims lightly. Carto, Mapbox have geospatial MCPs. They answer who do where questions. We answer what's actually happening there questions. No mobility provider, no foot traffic provider, no real world intelligence platform has shipped a production MCP server before this one. Facteus [music]
Speaker A
the first. If you're already building with Claude GPT, LangChain, Crew AI, or any agent framework, drop that framework in the comments. I want to see what you're working with. We'll come back to the integration in a couple of minutes.
Speaker A
What this means in practice, an AI agent running a site selection workflow can now [music] mid reasoning call Facteus MCP and pull actual foot traffic or a specific coordinate, the actual demographic catchment for a trade area, actual competitor density for a given
Speaker A
block, and incorporate that into this answer. In real time, no pipeline, no data engineering, no stale CSV. This is what a call looks like at the protocol level. The agent asks for a trade area data for a coordinate, and the MCP
Speaker A
server returns structured real world data, visit patterns, catchment demographics, competitor proximity. The agent reasons over it. The answer it produces is no longer a hallucination, it is grounded. [music] Same question we asked at the start.
Speaker A
Same model, but now the agent has Facteus MCP wired in. Left side, the hallucination confident generic based on nothing real. Right side, the grounded answer, [music] specific candidate addresses, actual foot traffic indices, real demographic catchment, real competitor landscape with distance
Speaker A
and visit overlap. That is not a better model, that is the same model with real world grounding. Every one of you watching could be running this in your own stack inside the a
Speaker A
server connected. You are now authenticated with Facteus MCP configured in the client. That is the entire setup. I'm going to ask, "Recommend the top three locations in Dallas for a new women's apparel store.
Speaker A
The target customer, women 25 to 44, household income 75,000 plus. Show foot traffic, trade area demographics, and competitor proximity for [music] each, and cite your data." And watch what happens. The agent is not just generating text, you can see it making
Speaker A
tool calls. Each one is hitting the real world graph, pulling structured data back. Now, look at the output. Three specific candidate locations. For each one, weekly visit volume, demographic makeup of the actual catchment, distance, visit overlap with the nearest
Speaker A
three competitors, the agent sites which data layer each in fact came from. So, that is the MCP. But, here is the thing.
Speaker A
Not every team that needs real world intelligence is going to wire up Claude with a custom MCP integration. Some teams want it packaged. Some teams want a working application they can deploy and modify. Some teams want it conversational. That is why we built
Speaker A
Factory AI as one launch within the five doors. Same real world graph underneath, different surfaces for different buyers arranged on an abstraction ladder from raw developer access at one end to no install conversational at the other.
Speaker A
[music] As I walk through these five, drop in a 1 2 3 or a 5 in the chat for the one that matches your team. I will prioritize Q&A on whichever front door wins the vote. [music] Front door one,
Speaker A
the MCP server. This is what we just saw. Raw, flexible, real time access to the real world graph from any AI agent, Claude, GPT, custom enterprise models, any framework that speaks MCP. You can query any of the 12 data layers, combine
Speaker A
them, build complex spatial reasoning into your agents. Full developer control. This is for the teams already building agents, tired of stitching three binders together to give those agents physical world granting. Front door two, the Claude plugin. This is the
Speaker A
lowest friction door, one click. Installs the MCP server and our pre-built skills bundle directly into Claude desktop. Zero to querying real world data in roughly five [music] minutes. No API key juggling, no config files, no DevOps tickets. If your team
Speaker A
already uses Claude and you want real world intelligence inside it tomorrow morning, this is your door. Front door three, pre-built skills. These are packaged agent workflows that sit on top of the MCP. Purpose-built for the high-value analytical tasks our
Speaker A
customers run over and over. Run a trade area analysis for a coordinate, build an audience segment from behavioral profiles, score a site against demographic and competitive criteria.
Speaker A
Forecast demand based on event and mobility signals. Instead of your team writing prompt engineering and tool orchestration from scratch, you drop in a skill and the workflow is done. This is for AI products teams who want the magic moment without the build. Front
Speaker A
Door 4, open source apps. These are full working applications we have built on top of the MCP for the use cases our customers run most often. Site selection, audience segmentation, demand prediction, trade area intelligence, and [music] there's more on the way. Each
Speaker A
one is deployable, customizable, and open sourced on GitHub. You can run them as is, fork them on your own workflow, or use them as a reference implementations to see what's possible in the real world graph end to end. This
Speaker A
is for the team that wants a working solution on day one and the freedom to modify the code. It is also frankly the fastest way for a CTO to evaluate whether the platform delivers. Clone the repo, point it at your market, and see
Speaker A
what comes back. Front Door 5, Ask the Real [music] World. This is a surface for everyone who is not going to install anything. A conversational web app, no install, no integration, and no API [music] key. You ask a question in plain
Speaker A
English, "How has foot traffic to coffee shops in Williamsburg changed over the last 12 months?" And it answers with real data, real charts, and real sources. This is for the marketer, the planner, the executive who needs answers grounded in the real physical world, but
Speaker A
is never going to write a tool call. We built this specifically to extend who gets to use real world intelligence beyond the data team. Five front doors, one real world graph underneath.
Speaker A
Whichever door you walk in, you get to the same physical world truth. 90 billion daily signals across countries with 92% higher accuracy than what your team is probably using today. [music] That is the launch. Let me show you the
Speaker A
non-technical front door end to end because it is a surface most of you on this call probably haven't seen yet.
Speaker A
This is Ask the Real World running on the same real world graph as the MCP. No install, browser-based. You're going to plan a retail expansion from scratch.
Speaker A
Quick one in the chat. Type yes if your team has ever produced a brief like this manually.
Speaker A
[music] I want to see how many of you have spent 2 weeks on something this app is about to do in 90 seconds. I'm going to type, I want to open three women's apparel stores in Chicago. The target customer
Speaker A
women 25 to 44 and household income above $80,000. Find me the three best candidate neighborhoods, show me foot traffic patterns, demographic catchment, and the closest competitors for each.
Speaker A
That is it. [music] Plain English, no filters, no parameters. Watch what happens. You can see it resolving the query, pulling visit data across candidate zones, running demographic matching on trade areas, checking competitor proximity. The progress is visible. You're not waiting on a black
Speaker A
box. Three neighborhoods ranked, each plotted on the map, each scored on three axes, foot traffic match, demographic match, and competitor position. Now, let me click into the top recommendation.
Speaker A
[music] This is the candidate card I can see. Weekly visit patterns, peak days, peak hours, [music] demographic breakdown of the actual catchment, age, income, and household size. Where visitors are coming from, [music] the origin heat map. Competitor location
Speaker A
within 10 minutes with their visit share. The trade area polygon, not a 1-mile circle, the actual behavioral catchment. This is a brief that used to take an analyst 2 weeks of stitching reports. It is live, it is current, and
Speaker A
a non-technical user just produced it in 90 seconds. [music] Here is the bit that matters for the rest of your team. Every output is exportable, every data set has a stable identifier, and every workflow on the app has a corresponding MCP call.
Speaker A
So, when your marketer in Astra Real World finds the three candidate locations, your data team can rerun the exact same analysis through the MCP, pipe it into your existing BI stack, and your AI agent can hand it off to the
Speaker A
next step in the decision chain. One decision, one workflow, one real world graph, five front doors that all see the same truth. Let me bring this home. AI knows the internet, it does not know the real world. And for any businesses
Speaker A
making decisions about where people are, how they move, what they want, what is actually happening on the block in a corridor or in a market. That blind spot is the most expensive thing in your AI stack right now. We are not solving it
Speaker A
with a better model. We are solving it by grounding AI in the real world intelligence data that is live, global, multi-layered, and built for how AI actually consumes data in 2026. That is the real world graph, and Factory AI is
Speaker A
how you connect to it. [music] We are in early access right now, not general availability, but early access. That is deliberate. We are working hand-in-hand with our first cohort to make sure the developer experience, the docs, the skills, and [music] the apps are sharp
Speaker A
before we open the floodgates later this year. So, I have two asks. Pick the one that fits you. If you're already in our beta cohort, today is your activation day.
Speaker A
[music] The link is in your invite email. We want your first MCP query in the next 48 hours, and we want your honest feedback in the shared Slack channel. We will read every line of it. That is how this
Speaker A
gets sharp before launch. If you're not in the beta yet, and you want in, drop the word real in the comments right now, and our team will follow up with an early access link before the stream ends. [music] Or, go to factory.ai/ai
Speaker A
and request access there. Spots are limited, and we are admitting in cohorts. And if you're a developer who wants to look at the code right now, the open-source apps I just walked through, site selection, audience segmentation, demand prediction, or trade area, are
Speaker A
all live on our GitHub today at github.com/factory-ai. Clone the repo, run it against your market, and see what the real world graph returns. No waiting for a beta seed. And one more thing, if you're building anything, site selection,
Speaker A
audience intelligence, demand forecasting, trade area analysis, market intelligence, auto form planning, retail analytics, anything that touches the real world and runs on AI, I want to hear from you directly. My email is in the description. Email me your queries,
Speaker A
or drop in your questions here, and we will take it up.
Topics:AI blind spotreal-world graphFactory AIModel Context ProtocolMCP serverenterprise AIfoot traffic datasite selectionreal-time dataAI grounding
