Esri's approach to responsible AI

Organizations are making real progress with AI. Workflows are faster. Repetitive tasks that once absorbed days of analyst time now get done in minutes. Most technology leaders would agree the results, in the right applications, are encouraging. But what is becoming clearer is that speed and efficiency alone do not translate into defensible decisions. These decisions require understanding: what the data reflects, why an outcome was reached, and whether the analysis accounts for the full picture. For many of the most consequential choices an enterprise faces, that full picture requires geography.

Esri has spent more than 50 years building a comprehensive geospatial platform that governments, utilities, public health agencies, and large enterprises globally depend on due to the real consequences their decisions have on people, communities, and business outcomes.

When generative AI came into the picture, Esri faced a choice. Should it move fast and figure it out later, or take the time to get it right in a way that reflects what customers need and expect? Esri chose the latter. That meant treating transparency and trust as requirements rather than afterthoughts. It also meant holding firm to human oversight, keeping people in control of every outcome, and ensuring that those accountable for outcomes review, guide, and correct any automated output.

Three areas define what that looks like in practice.

#1 Context: Geographic understanding is what grounds AI in reality

AI does a lot of things well. It processes data at scale, finds patterns, and generates outputs at speeds no human can match. But output quality depends entirely on the data the model was trained on, and the context it receives. For high-stakes decisions, where getting it wrong carries a real cost, data quality and analytical methods matter. That is where geographic understanding makes the difference between an informed decision and a fast one.

Why ontologies don’t tell the whole story

At the 2026 Gartner Data & Analytics Summit, Distinguished VP Analyst Rita Sallam described the problem, “Agentic AI outcomes depend on context including semantic representations of data. Without context — a clear understanding of the specific relationships and rules within an organization’s data — AI agents cannot operate accurately and are far more likely to hallucinate, introduce bias and produce unreliable results.” Organizations are already working on this by building ontologies and semantic layers that map out what concepts mean and how business processes connect. But most of these ontologies describe what things are, not where they exist or how location influences their relationship to everything around them.

Two distribution centers may be identical in an ontology and worlds apart in operational reality, separated by infrastructure gaps, regional demand differences, and climate exposure that only geography can explain. Ontologies map what organizations know. Adding geographic context reveals where and when that knowledge plays out. It shows how location shapes relationships between things that look identical on paper but behave very differently in practice. That distinction is what most AI systems, and most knowledge-mapping initiatives, have not yet accounted for.

Where the gap shows up in practice

The gap shows up quickly in industries where location drives value. Take a retailer evaluating market expansion. Demographics are a great starting point but cannot be the only lens through which this decision is made. The retailer needs to know how people move through an area, what businesses are nearby, how the neighborhood is shifting. A spreadsheet won’t capture any of that. A real estate firm managing a large portfolio must weigh environmental risk, zoning, infrastructure proximity, and market dynamics together, not in disconnected reports siloed across departments. A financial institution assessing portfolio risk needs to know not just who its clients are, but where they are. Physical conditions, climate exposure, flood risk, and aging infrastructure all affect the underlying numbers in ways balance sheets do not capture.

ArcGIS is built for these questions. ArcGIS brings together decades of authoritative geographic datasets, spatial analytics capabilities, machine learning and deep learning tools, and hundreds of pretrained models. Together, they give enterprises the grounding and capability to go beyond what general-purpose AI can reach on its own. These tools and models are available and are already being used today.

#2 Trust: Responsible AI, designed around the people working with it

As AI becomes more capable, the question of who is actually in control becomes more pressing. Esri hears this consistently from customers: genuine excitement about what AI can do, paired with real concern about depending on systems that are not fully understood. That concern sharpens when decisions carry direct accountability.

Those concerns are not abstract. A financial services firm applying AI to credit or insurance decisions is, in effect, determining who gets access to opportunity. A city using AI to prioritize road maintenance or utility repairs is making choices that affect residents’ daily safety. A public health agency using AI for disease monitoring informs interventions that touch vulnerable communities. A government agency applying AI to emergency response is operating in conditions where timing and accuracy are not recoverable if things go wrong.

Six commitments built into how ArcGIS is made

Knowing how a recommendation was reached, what informed it, and where the limits of the underlying models lie is a business requirement. It is the baseline condition for AI that can be trusted at scale.

Esri’s approach to responsible AI is built on six commitments: security, privacy, transparency, fairness, reliability, and accountability. These commitments govern how Esri designs, evaluates, and releases AI capabilities in ArcGIS.

What that looks like in practice is concrete. Transparency Cards give organizations clear visibility into how Esri built its AI models, what data trained them, and where known limitations exist. This empowers practitioners to apply AI with informed judgment. AI assistants inside ArcGIS live directly within the applications users work with every day. They provide coding support, workflow guidance, and in-context answers using natural language without taking action on the user’s behalf or making autonomous calls where human judgment belongs.

A release cadence that reflects the stakes

The ArcGIS release cadence reflects this same philosophy. AI capabilities in ArcGIS spend a considerable amount of time in beta before reaching general availability, and this is intentional. Each stage is an opportunity to collect feedback from users, validate that a capability solves actual problems, and confirm it meets the security and compliance standards Esri’s customers depend on. Much of the agentic AI work currently in development for ArcGIS is still in early beta for this exact reason. Getting it right matters more than getting it out fast.

#3 Interoperability: Location intelligence that works with what you already have

No enterprise is building its AI strategy from scratch. Most are managing a layered mix of platforms, data systems, and investments that took years to put in place. The practical challenge of adding location intelligence is making it work alongside what is already there, not competing with it or replacing it.

Esri’s design priority is straightforward: enterprises should not have to choose between the AI platforms they have already committed to and geospatial capability. The goal is to strengthen those existing investments by adding geographic context they currently lack.

Through support for Model Context Protocol (MCP), ArcGIS connects geographic context directly into the agentic AI ecosystems enterprises are already running. A government agency can connect ArcGIS to the platforms its teams use daily. This grounds AI-driven decisions in authoritative geographic data, rather than letting those recommendations float free from the places and communities they affect. For organizations that want to go further, ArcGIS also supports building spatially aware agentic applications using open standards. This allows organizations to keep existing investments intact while leaving room to adapt as the technology keeps moving.

At enterprise scale, ArcGIS gives AI what it often lacks: a grounded, geographic picture of operational reality.

Geography grounds decisions in reality

Where to open the next location. How to price risk across a large portfolio. Where to direct infrastructure investment. Where a public health intervention will have the greatest effect. How to plan for conditions that keep changing. Geography sits at the center of every one of those questions.

Enterprises that bring geographic understanding into their AI strategies early will make better decisions than those that treat it as an add-on later. More than five decades of experience alongside some of the most consequential decision-making organizations in the world has made that clear.

Esri’s approach to AI is built around that reality. It is grounded in authoritative data, transparent by design, and committed to keeping human judgment at the center. It is also built to work with the platforms and teams enterprises already depend on.

To read more about how Esri is advancing AI across ArcGIS, read The Next Era of AI in ArcGIS.