Should AI Agents Be Making Financial Decisions? It Depends.

Editor's note: This is AI Impact, Newsweek's weekly newsletter where each week, we will explore how business leaders are unlocking real value through artificial intelligence.

Tap or click here to get this newsletter delivered to your inbox.

Good afternoon.

Our Newsweek AI Impact webinar earlier this week did something I always like: It made AI sound a lot less magical and a lot more operational.

What made the conversation useful was how quickly it got into the machinery of the business itself. Yahoo’s Matt Sanchez talked about needing to modernize code and tooling before agentic development could really affect the core business. Microsoft’s Katy George explained how the company’s own Copilot AI rollout lost momentum until engineers studied how salespeople actually worked and redesigned around those workflows.

Qualcomm’s Durga Malladi brought the infrastructure side down to practical choices around connectivity, model size, cost and sensitive data. EY’s Rodrigo Madanes described another problem that can surface later, when privacy, data sovereignty and compliance requirements turn an easy-looking pilot into a much harder scaling exercise.

Among those examples, the practical challenge kept moving. Speeding up one part of a workflow can push the bottleneck somewhere else, while promising projects can run into constraints that were barely visible during experimentation. Capable tools can also lose adoption when they do not fit the way people actually work.

I pulled together the biggest takeaways from “Overcoming Barriers to AI Transformation in Legacy Industries” in my recap, and you can also watch the full webinar, presented by Cognizant, if you missed the live discussion. The video is also available at the top of the newsletter.

This week’s Signal Capture takes that same practical question into financial services. As I explore below, automating more of the process does not necessarily mean a business is making better decisions.

Read on for more.

Signal Capture

Signals from the frontlines of AI adoption

Taktile CEO: AI Agents Shouldn't Make Every Financial Decision

Banks and insurers can now automate parts of underwriting, claims and onboarding that once required people to read documents, investigate cases and apply policy. Maik Taro Wehmeyer, CEO of financial decision platform Taktile, says choosing where agents belong inside those processes is becoming as important as adopting them in the first place.

“Building an agent has become relatively easy, but putting that agent into a regulated process where a bank or insurer can trust and control its decisions is much harder,” Wehmeyer told Newsweek.

In practice, he described a financial decision as a combination of agents, data sources, fixed rules and human review, with each serving a different role in the process.

Claims processing shows how those pieces can fit together. An agent might extract information from documents and map it against a customer’s coverage before an insurer applies its existing rules. Routine claims can proceed automatically, while unusual cases are sent to a person.

“A fixed rule is the right tool when an answer needs to be deterministic,” Wehmeyer said.

His example was a policy preventing approval when a customer’s business is located in a particular country, which can remain a fixed rule. Agents become useful when information requires interpretation, such as financial data scattered across documents or an invoice written in another language.

Banks and insurers often start with a narrow group of straightforward cases and allow more to proceed automatically as the agent proves accurate on simpler work. As more cases move through automatically, the share handled without a person becomes an easy metric for leaders to watch.

“You can increase your automation rate while damaging business performance if the system starts approving bad loans or paying claims it shouldn't,” Wehmeyer said.

Tracking approval and decline rates over time can show whether the balance between the two is changing once an agent is in production. Wehmeyer also pointed to override rate, or how often a person corrects an agent’s work. Patterns in those corrections can reveal which types of cases the agent is repeatedly getting wrong and help improve how similar cases are handled.

Complex cases still move to people, which changes what review entails once an agent has already gathered information and analyzed the case.

“The reviewer should be able to see what information the agent used and understand how it reached its conclusion so that human review doesn’t become a rubber stamp,” he said.

Institutions also need evidence of how agents perform on cases where the correct outcome is already known. Wehmeyer said historical examples can provide that test before an agent reaches production.

“The failure mode the [finance] industry should be most concerned about is when an AI agent gets something wrong in a way that looks right,” he said.

Once an agent is live, organizations can continue comparing its decisions with the data and context behind each outcome to catch recurring errors before they become patterns.

The same testing becomes more consequential if agents are eventually allowed to propose changes to the logic behind regulated processes. Wehmeyer said those changes should be evaluated against past decisions before reaching customers, with people retaining approval authority and the ability to reverse them if outcomes deteriorate.

Responsibility for the final decision remains with the bank or insurer, even when an outside model or technology provider plays a role. Business owners should have direct visibility into and control over the agents involved, including how they affect approvals, declines and other outcomes. The record of how an individual decision was reached also matters months later, when the institution may need to revisit the case.

“The clearest warning sign is when nobody can easily explain why the system made a particular decision,” Wehmeyer said. “If a team has to reconstruct what happened after the fact, governance hasn’t kept pace with automation.”

Upcoming Webinars

What Enterprise AI Looks Like When It Works

On November 4, Dr. Ranjit Tinaikar, host of Newsweek’s “AI Impact Forum,” will sit down with Firdaus Bhathena, executive vice president and chief technology and transformation officer at S&P Global, to discuss how large companies are turning advances in AI into business results, and why trusted, differentiated data is becoming increasingly important.

Drawing on Bhathena’s experience at S&P Global, FIS Global and CVS Health, the conversation will examine how organizations make data governed and reusable, redesign workflows around new capabilities and balance speed with trust, security and resilience.

They’ll also explore how leaders can keep technology investments focused on customer problems and identify where genuine competitive advantage can emerge.

Join the live discussion Wednesday, November 4, at 11 a.m. ET. Register for free.

The AI Election: How Artificial Intelligence Shaped the 2026 Midterms

What will the 2026 midterms tell us about the emerging politics of AI?

One week after Election Day, Gabriel Snyder, Newsweek’s executive editor, enterprise, will moderate a discussion with policy experts and Washington insiders examining the midterm results through the lens of artificial intelligence.

The panel will feature Nathaniel Persily, professor of law at Stanford Law School; Brent Buchanan, founder and CEO of Cygnal; Nick Penniman, founder and CEO of Issue One; and Andrew Mamo, communications advisor at The Bench.

Using Newsweek’s AI Policy Scorecard as a starting point, the panel will explore where AI policy became a campaign issue, how positions crossed traditional party lines and how debates over regulation, data centers, jobs, privacy and national security played out in individual races.

The conversation will also look at how campaigns and outside groups approached these issues, including the role of industry spending, and what the results could mean for the policy fights ahead.

Register and join the live discussion on November 10.

Enterprise AI in the Physical World: Navigating the Supply Chain’s Messy Frontier

Warehouses, factories and transportation networks put enterprise AI to a different kind of test: digital decisions have immediate consequences in the physical world.

Duncan Angove, chief executive officer of Blue Yonder, joins Dr. Ranjit Tinaikar, host of Newsweek’s “AI Impact Forum,” on November 19 for a conversation about bringing increasingly autonomous systems into complex supply chain operations.

The discussion will cover the role of specialized models, the expertise companies need to preserve as more decisions become automated and whether the economics of enterprise technology could move toward measurable business outcomes such as productivity, operating costs and the movement of goods.

Register to join the live discussion Thursday, November 19, at 12:30 p.m. ET.

Prompt Injection

What’s one recent insight you’ve learned about AI?

“One recent realization is that AI has dramatically reduced the time it takes to produce an answer, but not necessarily the time it takes to prove that answer is right.

In marketing, AI can analyze performance data, generate scenarios and recommend budget changes in seconds. But those recommendations are only as reliable as the quality of the data feeding the model. If that data is incomplete, biased or built on flawed assumptions, AI does not correct the problem. AI trained on bad data just gives you bad answers faster and more confidently.

That has shaped how I think about using AI. The technology is extraordinarily valuable for identifying patterns, exploring possibilities and accelerating decisions, but it needs to be grounded in trusted data and evidence. For marketing leaders, that means using incrementality experiments and other rigorous measurement methods to establish what is actually driving business results.

As AI increases the speed and volume of decisions, the value of trusted proof only grows. The goal should not be to automate judgment. It should be to give people better evidence on which to exercise it.”

Have your own lesson to share? Email me at: a.mills@newsweek.com

Run Log

AI use case of the week

Fullstory AI Reveals Where Digital Experiences Break Down

Game Lounge, an iGaming affiliate that operates comparison and review sites and directs users to partner operators, often has two minutes or less to understand what a visitor does on one of its sites before that person clicks through to a partner. Because the company does not control what happens after users leave, traditional analytics gave its product team an incomplete picture of where people were getting stuck.

Lane Greer, head of customer engineering and strategic solutions at Fullstory, a behavioral data company that helps businesses understand how people use their websites and digital products, said Game Lounge used the platform to analyze visitors behavior during that limited time on its sites.

Fullstory captures interactions across Game Lounge’s webpages, while its StoryAI tools analyze that behavioral data for signs of friction. The system can surface problems such as clicks on elements that do nothing or interactions that trigger errors, helping the UX team identify where the experience needs attention.

Game Lounge reported a 20 percent faster validation time for new product updates, along with a 70 percent reduction in dead clicks and a 50 percent reduction in error clicks.

“To understand a digital experience holistically, tons of data must be collected, processed and analyzed,” Greer said.

The lesson? When a business sees only part of the customer journey, extracting more meaning from the interactions it does control becomes more valuable.

Have an interesting AI use case to share with me? Email me at: a.mills@newsweek.com

Context Window

■ Anthropic recently launched Claude Duet, an agent for professionals and small businesses that can work on tasks in the background, connect to more than 6,000 apps and services, and require user approval before actions such as sending messages, making payments or signing documents. [Anthropic]

■ A VentureBeat Intelligence survey found that 81 percent of companies running AI in production reported using half their GPU capacity or less, even as most of that group planned to adopt new AI infrastructure within a year. [VentureBeat]

■ Microsoft is expanding Project Perception to discover AI agents across enterprise environments, assign them separate identities and permissions, continuously score their risk and cut off access when their behavior becomes anomalous. [Dark Reading]

■ AI agents helped security researchers identify how critical vulnerabilities in PHP could be exploited and test fixes, prompting organizations worldwide to patch or temporarily take websites offline. [Associated Press]

■ Finance leaders warn that automating entry-level accounting work could weaken the pipeline for future controllers, finance directors and CFOs, forcing companies to rethink how junior employees build technical and leadership skills. [Fortune]

Transfer Protocol

Tracking executive moves across the AI landscape

Joe Castellanos, who led Bugcrowd’s product portfolio, including its expansion into AI penetration testing, has joined RemoteThreat as vice president of product, directing product strategy and development for its AI-enabled offensive cyber operations platform.

Vasu Jakkal, former corporate vice president of security at Microsoft, has joined Armadin as chief strategy and business officer, leading strategy, partnerships, customer success and corporate development as the company scales its autonomous security platform.

Jim Freeman, after serving as senior advisor at Permira, has joined Ticketmaster as chief technology officer, leading AI transformation and agentic tools, while Rachel Smith, coming from Block, has been named its first senior vice president of data science and machine learning, building a data science center of excellence.

Per Jr. Greisen, former chief scientist of AI biology at BioMap, has joined Abiologics as chief artificial intelligence officer, leading its AI and machine learning strategy for generative protein design and development of programmable D-protein therapeutics.

Priyanka Kharat, former vice president and general manager of Skylar AI at ScienceLogic, has joined Ribbon Communications as senior vice president and Acumen AI autonomous networking business leader, driving growth for its AIOps and automation platform.

Tap or click here to get this newsletter delivered to your inbox.