Geospatial Analysis and Generative AI: Banking and Telecom Views
In a 2024 AI Connect webinar, leaders from Webster Bank and Cox Communications described how generative AI could change geospatial analysis. They also explained why banking and telecom teams verify outputs, protect customer data and adopt it in baby steps.
By The AI Enterprise Magazine Team4 min read
Key takeaways
- Location is one more feature. Both panelists use geospatial data for forecasting, marketing tests, risk and fraud.
- Conversation lowers the barrier. Generative AI may let non-technical users ask for maps and analysis in plain language.
- Keep it out of decisions for now. Webster Bank does not feed generative AI output into models such as credit decisioning.
- Protect customer data first. Cox keeps work in house, using on-premises GPUs or tightly limited knowledge bases.
- Not everything needs generative AI. Proven analytics can be a better fit when risk outweighs benefit.
In a 2024 AI Connect webinar, "Navigating New Territories: Exploring Geospatial Analysis with the Power of Generative AI," a bank data science chief and a telecom data scientist described where generative AI helps with location data and where they are holding back. The session was part of the Generative AI for Data Executives Webinar Series and was sponsored by SandboxAQ.
Natalia Kuklova, Enterprise Chief Data Science Officer at Webster Bank, and Evan Walker, Senior Data Scientist at Cox Communications, spoke with moderator Brian Mink, co-founder and president of AI Connect.
How do banks and telecoms use geospatial analysis?
Kuklova said geospatial analysis connects data to physical locations. At Webster Bank, the team uses it for fraud detection, since the bank needs to know where a transaction took place, as well as for forecasting traffic and occupancy at branches, tracking properties in flood zones, evaluating commercial mortgage risk and targeted marketing. It also informs decisions about where to open or close branches.
Walker described location as one more data point in forecasting and prediction. At Cox, Walker said, it appears in zip-code-based marketing tests, in choosing test and control markets, in tracking competitive pressure by area and in adjusting offers and pricing.
How is generative AI changing geospatial work?
Both panelists pointed to a lower barrier to entry. Walker expected conversational requests, such as asking a tool to show something in an area and layering the analysis, to bring geospatial work to non-technical users.
Kuklova said generative AI already helps fill gaps in location data, for example when a city is known but the zip code is missing. Kuklova hopes it will improve visualization and routing too. As an everyday driver, Kuklova said, a map app should let people specify preferences such as avoiding left turns. UPS, Kuklova noted, had found right turns to be faster and safer for deliveries.
Where is generative AI too risky?
"We do worry in banking," Kuklova said when Mink asked about hallucinations. Her team uses generative AI for advanced analytics with humans verifying results, such as spotting an ATM transaction plotted in a country where it could not have happened. It does not feed generative AI output into its models. A hallucinated data point in credit decisioning, Kuklova said, "would actually hurt us," and Kuklova described the approach as taking baby steps.
Walker said that in telecom, "protecting customer data is number one." That means keeping work in house, whether on premises GPUs or knowledge-graph RAG that strictly limits the knowledge base, so that questions and customer information do not flow out to a public model.
How do you decide what to test and what to ship?
Kuklova's test is the use case. Anything that drives a business decision, such as approving a loan, is "a big no no" without human verification. For exploratory research, Kuklova said, a 5% or even 10% error rate is acceptable, and the team documents the shortcomings of each model. Kuklova added that outputs differ between large language models, so the bank's team compares them against fully labeled data and might accept a 3% to 5% mismatch.
Walker said any testing must first clear customer safety and data protection, and that outputs should be verified as true, interpretable and useful to the business.
Mink added a caution: "not everything has to be a gen solution." Mature, well-vetted analytics often remain the better fit, Mink said, when risk outweighs benefit.
What about emergencies and stress testing?
Asked how generative AI could help in a crisis such as a natural disaster, Kuklova said Webster already stress tests its loan portfolio using scenarios from a third-party provider. Generative AI could create such scenarios, but it is hard to explain to regulators why one scenario is likely, so the bank has not started. Walker saw potential for speed, but warned that misinformation and hallucination could be dangerous in a life-or-death situation.
What production use cases are emerging?
Walker's strongest examples came from conversational data. Cox analyzes call data to map customer intent, churn and likelihood to purchase, and some uses supply text-based guidance to field technicians visiting homes. Kuklova said the bank is still using traditional machine learning for tasks like drive time to the nearest branch, and has not tested heat-map analysis enough to adopt it.
On integration, Kuklova said major cloud providers make it easy to run LLMs next to geospatial data, and that role-based access and PII masking help control who sees what. Walker listed security, finding good use cases and hallucinations as the main hurdles. For smaller companies without large infrastructure, Kuklova said downloadable models and outside consultants are options, but "don't start uploading your legal documents to ChatGPT."
What comes next?
Kuklova expected generative AI to become the final natural-language layer on top of other applications, but said "Hallucination is something that stops us from actually adopting the technology in full." Walker expected tools to become more conversational and humanlike, and was watching for a tipping point in efforts to avoid hallucinations.
Continue the conversation
Data and AI leaders from regulated industries will meet at COLLIDE 2027 on Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center in Atlanta. AI Connect also has 24 free live webinars on its 2027 schedule; see the webinar page. Registration is coming soon.
Frequently asked questions
How is generative AI used in geospatial analysis?
Panelists cited filling gaps in location data, conversational map and analysis requests, scenario generation and analyzing call data, with human verification.
Why do banks move slowly with generative AI?
Natalia Kuklova said hallucinations and inconsistent outputs between models make it risky for decisions such as credit, so Webster Bank uses it for exploratory analytics.
How do telecom companies protect data when using generative AI?
Evan Walker said Cox keeps work in house, using on-premises GPUs or knowledge-graph RAG with a strictly limited knowledge base.
Can small businesses use generative AI safely?
Kuklova said downloadable models on their own virtual machines or a consultant are options, and warned against uploading sensitive documents to public tools.
This article is based on remarks by Natalia Kuklova, Enterprise Chief Data Science Officer, Webster Bank; Evan Walker, Senior Data Scientist, Cox Communications; moderator Brian Mink, Co-Founder and President, AI Connect at AI Connect webinar, Generative AI for Data Executives Webinar Series (2024) on October 16, 2024. More coverage: Webinars.
Filed under: AI Use Cases & Applications, AI Governance & Trust, Security & Fraud
