The Geospatial Edge: Issue 14, Q3 2025

The Geospatial Edge is Esri Canada’s periodic newsletter for managers and professionals tasked with growing their organizations’ geospatial capabilities. In this issue, Matt Lewin discusses an approach to modelling an AI agent from a business-first perspective using a simple one-page solution canvas.

The buzz around AI agents is everywhere—on stage at conferences, in product roadmaps and in headlines promising transformation. But behind the buzz, reality is setting in. According to Gartner, 40% of agentic AI projects will be shelved by 2027. That’s a sobering statistic and a clear signal that excitement alone doesn’t guarantee success.

What's to blame? Costs, risks and unrealized benefits are all highlighted. But a key statement in the report stands out as an underlying problem that I've personally observed, and I believe is the crux of the issue: many of the use cases identified as agentic never really required an agentic implementation in the first place.

It would seem we have a problem with how agents are conceived and conceptualized in the first place, long before they enter development.

In my observation, approaches to AI agent design often begin with data and algorithms. While technically rigorous, these methods can overlook the broader business context, leading to solutions that are misaligned with organizational priorities or are too complex and costly to justify the investment. A more prudent approach would be to flip the script and start with the business capabilities your agent is modelling and the outcomes you're trying to achieve. On that basis, you design the agent's cognitive architecture around those needs.

In this article, I look at an approach to modelling an AI agent from a business-first perspective using a simple one-page solution canvas.

About business capabilities

Business capabilities represent a high-level abstraction of an organization's core business functions. Specifically, they represent what a business does, as opposed to how those abilities are executed. In this way, business capabilities focus on outcomes rather than processes.

Typically, business capabilities are organized into nested tiers, with the top-level capabilities representing the broadest and most strategic outcomes of an organization, and lower-level capabilities representing more granular sub-capabilities associated with more tactical outcomes. Customer relationship management is an example of a top-level capability, with customer onboarding and customer retention being lower-level sub-capabilities.

For a more comprehensive understanding of business capabilities, read this excellent article.

What makes AI agents well-suited to manifest business capabilities is their outcome-driven nature. Traditional software is functionally aligned: it performs tasks based on predefined inputs and workflows. In contrast, AI agents can be capability-aligned: they operate autonomously to achieve a goal, drawing from data, models and reasoning frameworks. Give an agent a goal and the necessary data, and it can plan, reason and adjust its own steps while it works. That makes agents an almost perfect mirror of a capability whose essence is also outcome-focused ("ensure 80% customer activation rate" as opposed to "run this batch job at 3 am").

The anatomy of the canvas

The AI Agent Solution Canvas is structured around several key sections, each serving a distinct purpose in the agent's design.

To illustrate how the canvas works, I'll use an example from an article I wrote a few years ago, which discussed using GIS technology to help manage election-day voting locations for a city government. This example is in many ways a geographic problem, and I'm interested in how advances in AI and geospatial reasoning technology could help tackle this issue.

Elections management is a critical function of city governments. Few things are more vital to citizen confidence than a well-run election. And few things can damage public relations worse than a bungled election. On election day, effective voting place management is a primary concern. Among their many responsibilities, city officials and volunteers must ensure voting locations are set up and open on time, ballots are available throughout the day, electors with disabilities are provided with accessibility accommodations and voting places close on time.

How do you decide if an AI agent can help with this process? Let's use the canvas to organize the key components and make the determination.

Start by clearly defining the name and purpose of your AI agent. That means giving the agent a descriptive title and a definition that reflects the relevant business capability you're modelling. In the voting location case, let's call it Voter Flow Manager and define it to focus on managing polling locations on election day to ensure smooth operations and minimal wait times. There's much that goes on on election day, so we want to be clear about which aspect of election day management our agent is scoped to support. In this case, we're focused on the operational aspects of managing queues and wait times on the day of the elections. Other election day activities (e.g., results reporting) and longer-term planning activities (e.g., polling location selection) are out of scope for this agent.

Business Context

With the overall purpose of the agent clearly defined, your next step is to define the business context. Every AI agent needs a clear articulation of the business capability it supports. This is the foundation of the canvas as it defines the operational environment within which the agent functions.

In our example, the agent is designed for a city's election office, which is responsible for planning and operating polling sites to ensure voters can cast their ballots efficiently, equitably and in compliance with legal requirements.

This section of the canvas includes three critical components:

Business Capability. What function does the organization perform? In this case, it's the orchestration of polling stations.

Issue. What problem is the organization trying to address? Long lineups at polling stations can discourage participation, especially among seniors and people with disabilities, and erode public trust.

Desired Outcomes. What does success look like? The key outcomes include keeping average wait times below 20 minutes, ensuring no location exceeds 30 minutes and achieving target throughput per hour.

The key step in this process is clearly defining the desired outcomes, as these ultimately become the goals of the agent. AI agents, by definition, are goal-oriented, meaning they don't simply respond to queries; like AI assistants, they map out and execute complex tasks under the guidance of an overarching set of goals and outcomes. As stated above, in our voting location example, our primary goals relate to wait times and throughput. A clear set of goals enables the AI agent to prioritize actions and measure their effectiveness.

Note: the goals of an AI agent must be bounded by a set of constraints that guard against negative or unintended consequences. These are identified as "guardrails" in the "Design Parameters" section.

Cognitive Capabilities

The next section defines how your agent will behave. An agent's cognitive capabilities define how it perceives its environment, reasons about it and acts to achieve its defined goals. This is known as the agent cognitive cycle and determines how an agent operates autonomously in dynamic environments.

See (Perceive). The agent must ingest and interpret data from various sources. In the election office example, this includes live queue monitoring via sensors, turnout projections, weather and traffic data, and updates on staff and equipment availability. These inputs allow the agent to build a real-time picture of the operational environment.

Think (Reason). The agent must be able to analyze the data and make informed decisions. It predicts where and when surges in voter turnout will occur, simulates interventions such as reallocating staff or equipment, assesses the impact on compliance and equity, and chooses the most optimal and fair response.

Act (Execute). Finally, the agent must be able to take action. It recommends resource reallocations, redirects voters to alternate sites, updates public wait-time dashboards and notifies officials while logging its actions for accountability.

By framing the agent's capabilities in human-like cognitive terms, the canvas makes technical functionality more accessible to those who might not be as familiar with AI or software terminology.

Design Parameters

The final section of the canvas addresses the technical and operational scaffolding around the agent. These design parameters include the key data inputs, user interface, human oversight and guardrails, and are essential to ensuring your agent is not only functional but also trustworthy and usable.

Key Data Inputs. The data your agent leverages is perhaps the single most important design consideration. The effectiveness of your agent largely rests on the quality of the data your learning models are trained on, in addition to the external sources of data your agent accesses during the cognition cycle.

Training data forms the foundation of the agent's predictive and reasoning capabilities. It is used during the development phase to build models that can anticipate patterns, simulate outcomes and make informed decisions. For the voter flow manager agent, relevant training data includes:

• Historical voter turnout patterns
• Past queue lengths and wait times

• Weather and traffic correlations
• Staffing and equipment deployment histories

These datasets are typically cleaned, labelled and used to train machine learning models that underpin the agent's reasoning capabilities.

Data consumed by the agent during operation, often streaming and in real-time, is vital for situational awareness and responsiveness. These power the perceive and act capabilities of the agent, allowing it to perceive the present moment and execute decisions, such as updating dashboards or notifying officials. Datasets could include:

• Queue length counts (updated on a mobile app by staff monitors)

• Current weather and traffic feeds
• Staff and availability updates

• Polling site maps

User Interface. The agent must interact with different stakeholders through appropriate interfaces. These include an operations dashboard for headquarters, a mobile app for site supervisors and volunteers, and a public portal for voters. Designing these interfaces thoughtfully is critical for adoption and transparency.

Human Oversight. AI agents should not operate in isolation. In the election office example, election officials approve resource allocations, poll supervisors act on recommendations and field staff update the app and provide feedback. This ensures that human judgment remains central to decision-making.

Guardrails. Ultimately, the agent must operate within clearly defined boundaries. It must comply with election bylaws and key regulations, maintain transparent audit trails, protect voter privacy and avoid outcomes that disadvantage specific voter groups. These guardrails are essential for building trust and ensuring fairness.

Practitioner’s insights

As I worked my way through the voting location example using the canvas, I found it surprisingly clarifying, but also a bit sobering. The idea of building an intelligent agent to manage voter flow sounds great, but as I filled in each section of the canvas, especially the cognitive capabilities and data inputs, I started to question whether this agent was truly feasible. Could an election's office reliably access real-time queue data from every polling site? Or could they rely on voting location attendants to continuously update queue lengths? Basically, would the cost or complexity of building such a system deliver the necessary ROI?

That’s precisely where the canvas proved its value. It didn’t just help me imagine what the agent could do, it forced me to confront what it would need to succeed. That's the essence of this tool. It's not meant as a technical specification, but as a conceptual blueprint. It's a working document designed to help you align on what the agent is for, how it will operate and what success looks like.

In the end, I decided that a narrower, more focused use case would be a more effective starting point. Something like optimizing staffing levels at a single site or predicting peak hours based on historical turnout and weather conditions. These are simpler use cases, but still in the spirit of agentic AI, meaning they are goal-oriented, require persistent memory of interactions, and orchestrate tasks, tools and other agents (potentially).

As AI agents become more embedded in businesses, the need for thoughtful, purpose-driven design grows. A canvas such as this one (or one you build yourself) can help you and your team hone in on vital design criteria before launching into expensive implementation work. Over the years, I've leveraged many design canvases such as the Value Proposition Canvas, the Business Model Canvas and the Geospatial Strategy on a Page (a co-creation). All have been useful in their own way, but the key is they all drive a process that helps you to think things through before you act. Always a good thing!

Let’s talk

Are you exploring the potential of AI and AI agents for your business? Are you trying to integrate AI into your GIS program? Are you wondering how to get started? Send me an email or connect with me on LinkedIn. I’d like to hear from you!

All the best,

Matt

The Geospatial Edge is a periodic newsletter about geospatial strategy and location intelligence by Esri Canada’s director of management consulting, Matt Lewin. This blog post is a copy of the issue that was sent to subscribers in September 2025. If you want to receive The Geospatial Edge right to your inbox along with related messages from Esri Canada, visit our Communication Preference Centre and select “GIS Strategy” as an area of interest.