The American Transportation Research Institute puts the industrywide toll at more than $25 billion annually in lost productivity. A single roadside breakdown runs between $450 and $760 in direct repair costs before towing, rental replacements and missed revenue enter the equation. Fleet operators have long treated those losses as a fixed cost of doing business.

Artificial intelligence-driven predictive maintenance is starting to rewrite that calculus.

Turning Sensor Data Into Early Warnings

Commercial trucks are already generating enormous amounts of data. A typical heavy-duty vehicle produces more than 25,000 data points daily from onboard sensors tracking engine temperature, oil pressure, brake wear and fuel consumption. Historically, most of that data sat unused inside disconnected maintenance systems. Fleet managers learned about problems when trucks stopped moving, not before.

AI changes the model. Machine learning systems ingest real-time sensor readings alongside historical repair records, then identify the combinations of signals that tend to precede specific failures, often weeks before a breakdown occurs. The output isn’t raw data but a specific recommendation: a particular vehicle, a particular component, a service window that fits the route schedule. Repairs shift from roadside emergencies to planned shop visits.

McKinsey estimated that AI-driven predictive maintenance could reduce maintenance costs by 10% to 40% and cut downtime by up to 50%.

OEMs and Operators Move In

Volvo Trucks North America unveiled AI-powered adaptive maintenance as part of its Blue Service Contract in October 2024, replacing fixed service schedules with intervals that adjust dynamically based on how each truck is actually being used: fuel consumption, idle time and oil condition.

Magnus Gustafson, vice president of connected services at Volvo Trucks North America, said many fleets are over-maintaining their trucks, driving unnecessary cost. “Applying AI to optimize maintenance intervals based on truck specs, operating conditions and actual use ensures our customers can maximize uptime,” Gustafson said.

Volvo’s Uptime Center in Greensboro, North Carolina, monitors nearly 85,000 connected trucks across Europe, with specialists reviewing AI-generated alerts and coordinating service visits before breakdowns occur. Volvo and Mack Trucks have developed connected systems that cut the time needed to diagnose a breakdown by 70% and reduce repair time by 25%.

The shift is hitting fleet operators at a difficult moment. ATRI’s 2025 operational costs report found that non-fuel operating expenses rose 3.6% in 2024 to the highest level ever recorded, with average operating margins below 2% across most trucking sectors. Parts and labor costs are up more than 10% year over year, Fleetio fleet ecosystem strategist Stefano Daneri told PYMNTS. Fleets holding onto vehicles longer to avoid replacement costs are absorbing a growing hidden cost in the process: more downtime.

Capital and Adoption Follow

Fleet and mobility firms are directing working capital toward the technology. A PYMNTS Intelligence study found that 89% of fleet firms used at least one external working capital solution in 2024, with strategic deployment increasingly directed toward digital fleet management platforms and AI-based tools. Top performers realized an average of $15.6 million in bottom-line benefits.

The main constraint on broader adoption is data infrastructure. Many carriers still run disconnected legacy systems that prevent AI models from accessing the full maintenance history they need to make accurate predictions.

ATRI reported that the average miles traveled between breakdowns increased from 37,700 to 38,249 in 2024, crediting preventive maintenance practices as a key factor. Whether AI systems can push that number further will depend on how fast fleets close the data gap.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.

The move underscores how competition in AI centers not only on chips and infrastructure, but also on attracting a relatively small pool of researchers capable of pushing model performance forward. Anthropic, best known for its Claude family of models, has emerged as one of OpenAI’s strongest rivals in enterprise AI and coding applications.

Karpathy announced the move on X, saying he was excited to return to research and development work during what he described as a formative period for large language models. According to Reuters, he will work under Anthropic’s pretraining lead Nick Joseph, another former OpenAI employee.

Karpathy carries unusual influence across both research and developer communities. Beyond his work at OpenAI and Tesla’s Autopilot division, he became widely followed for educational content explaining neural networks and transformers to engineers and nontechnical audiences. He also founded Eureka Labs in 2024, an AI-focused education company that remains active alongside his new role at Anthropic.

The hire comes as Anthropic continues to expand aggressively across enterprise AI. The company has positioned Claude as a model family optimized for coding, enterprise workflows and AI safety, areas viewed as critical battlegrounds for commercial adoption. Anthropic was founded in 2021 by former OpenAI researchers including Dario Amodei and Daniela Amodei.

Karpathy’s move also adds to the growing list of high-profile departures from OpenAI over the past two years. Former OpenAI leaders including John Schulman, Ilya Sutskever and Mira Murati have all left the company amid broader debates over AI commercialization, governance and research priorities.

The recruitment also reflects how AI companies compete through researcher ecosystems and developer mindshare. Karpathy’s technical credibility, combined with his popularity among engineers building with generative AI tools, gives Anthropic both research depth and cultural influence at a time when coding assistants and agentic AI systems are becoming major commercial markets.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.

Because the barrier to artificial intelligence isn’t the technology. It’s the operational infrastructure underneath it.

New research from PYMNTS Intelligence shows that financial services, healthcare and media and advertising companies face very different obstacles to deploying AI at scale. And while ambition is shared across all three sectors, what’s blocking progress is unique to each one.

What the Data Shows

AI budget growth is real across the board. Eighty-five percent of financial services and insurance firms plan to increase AI spending over the next 12 months. Media and advertising companies are close behind at 80%. Healthcare trails at 60%, though for reasons the data quickly explains.

The reasons behind that spending reveal just as much. Financial services firms tie their investment to productivity gains and competitive positioning, both at 65%. Healthcare, by contrast, leans heavily on pilot funding with no formal ROI requirements, cited by 60% of firms. That’s not a sign of lower commitment. It’s a sign of an industry still figuring out where artificial intelligence fits in clinical and operational workflows. Media and advertising points to executive-driven strategic alignment, at 50%, but only 25% of firms in that sector can point to hard financial metrics to justify the spend.

The barriers follow a similar pattern. Financial services firms are held back mainly by data quality problems, cited by 30% as the single biggest obstacle. Clean, standardized data is the foundation for nearly every AI use case in that sector, and too much of it still isn’t there. Healthcare faces two equally serious problems at the same time: system integration and data quality are each cited by 30% of firms. That double constraint is what makes AI so difficult to scale in clinical environments, where patient data lives in dozens of disconnected systems. Media and advertising has no single dominant barrier. Instead, organizational problems spread across skills gaps, governance and leadership alignment are each cited at 15% to 20%.

3 Key Findings

• Financial services has the most mature AI infrastructure of the three sectors, but data quality is the ceiling. The systems are largely in place. The inputs aren’t reliable enough yet to support broader use cases.
• Healthcare’s challenge is structural, not philosophical. Clinical data exists in abundance. The problem is that it is fragmented across systems that don’t talk to each other. AI cannot do much when it can’t reliably access consistent data.
• Media and advertising needs organizational alignment before it can scale anything. Governance, talent and leadership issues must be resolved together. Fixing one without the others doesn’t move the needle.

What This Means for Banking, FinTech and Digital Economy Leaders

For banks and FinTech firms, the data is both encouraging and instructive. Financial services leads in deployment depth and budget commitment. But productivity gains will plateau if data quality problems aren’t addressed directly. That means investing not just in AI tools, but in the data pipelines, governance standards and infrastructure that feed them.

For FinTechs building products for healthcare clients, the integration constraint is a market opening. Solutions that connect fragmented clinical and operational systems are in direct demand.

The Bottom Line

Across all three sectors, more than 80% of leaders say artificial intelligence will augment human decision-making over the next five years, not replace it. The goal is widely shared. The path to it is not. What separates AI leaders from laggards isn’t budget or ambition. It’s whether the operational foundation exists to support the tools they’re buying.

Findings in the inaugural “ABCs of AI Credit: A Playbook for Issuers,” a PYMNTS Intelligence collaboration with Thredd, confirm that’s the case with the credit landscape, where legacy architectures designed for the single moment of origination are finding their “if-then” workflows becoming structurally misaligned with how money actually moves.

Today’s environment is one where digital channels dominate transaction volume, commerce is always-on, and user behavior is fluid across devices, geographies and contexts. The emergence of embedded finance, real-time data processing and adaptive underwriting models has turned credit into an event-driven decision rather than a pre-approved condition.

Fraud has also evolved, with synthetic identities, AI-generated deception and real-time social engineering attacks challenging traditional defenses.

As a result, the competitive battleground is moving away from the binary discipline of declining risk toward the more complex challenge of approving it intelligently, continuously and contextually.

At the center of this shift is the rise of artificial intelligence (AI) agents. These autonomous, real-time decision engines sit inside payment flows and can determine credit outcomes at the moment of transaction. This evolution represents more than incremental innovation. It signals a redefinition of how financial institutions manage risk, capture revenue and engage customers.

From Static Rules to Continuous Intelligence

The legacy credit stack operated by defining rules, applying them consistently and managing risk by limiting exposure.

But across today’s landscape, hewing too closely to these traditional static rules can create two critical problems. First, they can generate false declines by blocking legitimate transactions that fall outside predefined patterns. Second, they can fail to detect increasingly nuanced fraud signals that do not conform to historical thresholds.

AI agents can help address both issues by introducing continuous intelligence into the transaction layer. Rather than applying fixed criteria, these systems are able to evaluate behavioral context, transaction intent and real-time data signals simultaneously. The result is a shift from retrospective analysis to live decisioning.

Each transaction becomes an opportunity to reassess risk, informed by up-to-date behavioral, transactional and contextual data. Advances in machine learning and data infrastructure now allow lenders to evaluate not just who a customer is, but what they are doing, where they are doing it and how their financial position is evolving in real time.

Read the report: ABCs of AI Credit: A Playbook for Issuers

Instead of acting as credit gatekeepers, AI agents, in this emerging environment, are becoming orchestrators by balancing risk, customer experience and revenue in real time.

This transition mirrors broader trends in digital commerce, where personalization and immediacy have become baseline expectations. Consumers no longer accept one-size-fits-all financial products; they expect credit to adapt to their needs as fluidly as the platforms through which they transact. In this environment, the value of credit lies not in its availability per se, but in its relevance at the precise moment it is needed.

For financial institutions, the implications can be both immediate and measurable.

When credit decisions are made continuously, the marginal cost of an approval or decline is lower, but the opportunity cost of a missed approval is higher. Declining a transaction that could have been safely approved is no longer a neutral outcome; it represents lost revenue, diminished customer engagement and potential attrition to competitors who can underwrite more effectively in real time.

As a result, the competitive question is shifting from “Who can say no most effectively?” to “Who can say yes most intelligently?” This may require a fundamentally different set of capabilities.

Ultimately, the transition to moment-of-spend credit is redefining what it means to be creditworthy. In the traditional model, creditworthiness was a relatively static attribute, derived from historical behavior and expressed through a score. In the new model, it is a dynamic state, continuously updated based on real-time information.

At PYMNTS Intelligence, we work with businesses to uncover insights that fuel intelligent, data-driven discussions on changing customer expectations, a more connected economy and the strategic shifts necessary to achieve outcomes. With rigorous research methodologies and unwavering commitment to objective quality, we offer trusted data to grow your business. As our partner, you’ll have access to our diverse team of PhDs, researchers, data analysts, number crunchers, subject matter veterans and editorial experts.

The app handles informational queries in natural language. Users can check account conditions, compare card types and explore savings products by asking questions the way they would ask a person. Conversations include direct links to relevant product pages. No separate login. No navigation to a bank-owned interface. The banking experience starts inside the chat.

Meeting Customers Where They Are

The logic behind the move is straightforward. ChatGPT has become one of the primary channels through which people seek information online, BBVA noted. Building a presence inside it is less about technology and more about distribution. A bank that shows up where a customer is already asking questions does not need to convince them to open a separate app first.

Murat Kalkan, global head of digital banks at BBVA, told the bank’s website the integration marks the start of a new chapter where AI becomes part of everyday banking. The current version is designed for product discovery. BBVA expects it to evolve into a digital companion that helps customers navigate financial decisions conversationally over time.

The initiative is part of BBVA’s broader strategic partnership with OpenAI. The bank has already deployed ChatGPT Enterprise to more than 11,000 employees, with plans to expand to 120,000 staff globally, PYMNTS reported. During the initial rollout, 80% of participants accessed the assistant daily and reported saving an average of three hours per week on routine tasks.

Broader Race for Financial Interface

BBVA is not alone in watching where the customer relationship is moving. ChatGPT now lets Pro users connect financial accounts across more than 12,000 institutions through Plaid and receive answers grounded in their actual spending and cash flow, PYMNTS reported. More than 200 million people already ask ChatGPT questions about personal finance every month.

That number is what banks are paying attention to. A conversational interface that sits between a customer and their financial decisions is not neutral infrastructure. It is a relationship layer. The institution that owns that layer, whether a bank, an AI company, or a platform, owns the first point of contact for every financial question a customer asks.

The consumer data reinforces the urgency. Thirty-seven percent of power AI users reported using native AI platforms as their primary tool for managing finances and banking, PYMNTS found. Among mainstream users, that share doubled in a single month. Dedicated AI platforms are gaining ground specifically in finance and banking as a category, while embedded tools inside merchant and bank apps are losing relative standing, PYMNTS Intelligence reported.

BBVA’s ChatGPT app does not move money or execute transactions. It is positioned as a discovery and information tool. But the ambition the bank has described points toward something more integrated over time. For now, BBVA is claiming a position inside the interface before the interface becomes the bank.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.

The $2.2 billion deal is aimed at making Publicis a “leader in data co-creation, an important capability in the age of artificial intelligence and an enabler of agentic business transformation,” the companies said in a Sunday (May 17) news release.

As the release noted, LiveRamp is a global data collaboration platform that let companies “unify, manage, and activate” data across the digital space, connecting more than 25,000 publisher domains and 500+ technology and data partners in 14 markets. It also allows brands, retailers, media platforms and data providers to safely and effectively collaborate and connect data.

A report by The Wall Street Journal (WSJ) about the deal characterized the acquisition as Publicis trying to tap a rising demand from companies that want to transform their businesses by deploying AI agents that can complete tasks autonomously.

“We did not need LiveRamp to win in the marketing space,” Publicis CEO and Chairman Arthur Sadoun told WSJ. “Where LiveRamp plus Publicis is going to make a difference is in the agentic space, in this new market where there is huge opportunity because there is a huge barrier created by data.”

LiveRamp allows companies in different industries to scan data across different sources and transform them into actionable data assets, the report added.

“There is no way you can win with agents if you don’t have the right and differentiated data,” Sadoun said. “For agents to be competitive and to work, they have to run on good data, data that is unique, actionable, connected.”

In other agentic AI news, PYMNTS wrote Monday about the technology’s use in the banking world, following Fiserv’s launch of agentOS, an operating system that lets financial institutions deploy and manage AI agents across core banking, payments and servicing workflows.

The infrastructure here, that report added, is “moving faster than the rules,” with the Financial Data Exchange launching an initiative focused on what happens when AI agents handle consumer financial data autonomously. 

“The problem it is trying to solve is structural. When a consumer connects a bank account to a third-party app, the consent is visible and deliberate,” PYMNTS added.

“When an AI agent does the same thing on a consumer’s behalf, the questions multiply: who authorized the agent, what data can it access, how is that permission tracked and who is liable when something goes wrong. The standards that govern consumer financial data sharing today were not written for that scenario.”

 

China entered the AI era with super apps already embedded into daily life. Those platforms did not need to find users for AI. That entry point is defining how quickly AI moves from feature to infrastructure.

WeChat combines messaging, payments and commerce for 1.3 billion monthly active users. Taobao, Meituan and Douyin each run their own integrated discovery and payment layers. Inserting AI into those environments means plugging into habits that already exist, CNBC reported.

The scale of what has already deployed is significant. Alibaba’s Qwen AI assistant reached 300 million monthly active users across Taobao, Tmall and Alipay by early 2026. Roughly 140 million first-time AI shopping experiences were logged during a single Chinese New Year campaign, Let’s Data Science reported. Transactions were completed through Alipay. The AI steps back only for final user confirmation. The loop never leaves the app.

ByteDance upgraded its Doubao AI chatbot to autonomously handle tasks such as ticket bookings through Douyin’s commerce layer. Tencent is building equivalent capabilities directly into WeChat. Each platform is racing to make AI the operating layer of an app consumers already depend on, Sinolytics noted. The goal is not to launch a new AI product. It is to make AI invisible inside one that already has a billion users.

What Western Platforms Are Still Building Toward

ChatGPT remains largely a standalone interface. It can answer questions. It can browse the web. Users who want to transact typically leave to do it somewhere else. The transaction layer is not yet attached.

That gap is visible in consumer behavior data. Product link discovery has the highest adoption rate of any personal AI task measured across a five-month study. Shopping is where Western consumers are learning to trust AI, PYMNTS reported. It is high frequency, low stakes and requires no demographic preconditions. A wrong result costs nothing. That makes it the entry point. In China, that entry point was already inside the super app before AI arrived.

More like this: The Data Behind AI’s Shift to Everyday Consumer Use

Over half of U.S. adults now use AI tools in their daily lives, up five percentage points in January alone. The habit is forming. The infrastructure to close transactions inside that habit is still being built.

The deeper gap is not about features. It is about where the data lives. Every AI-assisted transaction inside a super app trains the model on real purchase behavior, real payment data and real fulfillment outcomes. All of it stays inside a closed system. Western AI companies working through open browser interfaces collect a thinner signal.

That asymmetry compounds over time. The platform with the most complete transaction data builds the most accurate model. China’s super apps entered 2026 already holding that data. Western AI platforms are still negotiating access to the transaction layer.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.

Inside Fiserv’s Bet on a Single AI Operating System for Banks

Banks have spent years running disconnected AI pilots with no common infrastructure underneath them. Fiserv launched agentOS, an operating system designed to let financial institutions deploy and manage AI agents across core banking, payments and servicing workflows from a single governed environment. Six banks helped build it. Two are in beta today. OpenAI and AWS joined as collaborators.

The use cases already in motion are narrow but concrete. First Interstate Bank is piloting an agent for commercial loan onboarding, a process that currently spans multiple systems and requires significant manual hours. Boulder Dam Credit Union is running a daily operational analysis agent that compressed report generation from ten minutes to seconds. Fiserv noted the platform includes kill switches, human-in-the-loop controls and audit trails designed to meet bank-grade regulatory requirements.

Fiserv Co-president Dhivya Suryadevara said that every bank client the company has spoken with is facing the same pressures: cost, deposit competition and a retiring workforce. AgentOS is Fiserv’s answer to all three at once. Whether that answer holds at the scale of thousands of institutions, across systems Fiserv does not control, is what the next 12 months will reveal.

When AI Agents Access Your Bank Account, Who Is Responsible?

The infrastructure is moving faster than the rules. The Financial Data Exchange launched an initiative specifically focused on what happens when AI agents handle consumer financial data autonomously. FDX is a non-profit standards body representing approximately 200 organizations, with more than 114 million customer accounts connected through application programming interfaces (APIs) aligned with its technical standards.

The problem it is trying to solve is structural. When a consumer connects a bank account to a third-party app, the consent is visible and deliberate. When an AI agent does the same thing on a consumer’s behalf, the questions multiply: who authorized the agent, what data can it access, how is that permission tracked and who is liable when something goes wrong. The standards that govern consumer financial data sharing today were not written for that scenario.

FDX CEO Kevin Feltes said that broad industry collaboration will be critical in the months ahead to ensure connections are built in a way that protects consumers. The call for input is open through May 29. The answers will shape what banks and FinTechs can actually deploy and how much regulatory exposure they carry when they do.

Banks Spent a Decade on Data Infrastructure. Now Comes the Hard Part.

The backdrop for both announcements comes from an analysis Camunda published, drawing on research from Datos Insights. The argument is direct: financial institutions spent a decade building data infrastructure: APIs, consent frameworks, secure sharing protocols. Agentic AI is now the layer that can act on that infrastructure rather than simply move data through it.

Many banks treated that infrastructure build as a compliance exercise, Datos Insights found. They built the APIs and stopped. Data flows, but very little happens with it. Agentic AI changes the calculus: an agent can reason across multiple systems, sequence a set of actions and execute a transaction without waiting for a human to initiate each step.

The open question the analysis does not resolve is who captures the value. Banks that built data infrastructure but never activated it have a window now. So do the FinTechs that have been working with that data for years. AgentOS is one answer to how institutions get there. FDX is working out what the guardrails look like when they do. Neither question has a settled answer yet.

This collaboration, the data/tech company announced Friday (May 15), is designed to help businesses make better decisions with the help of autonomous AI agents.

“Through this partnership, autonomous AI agents can gain the ability to act faster, and more consistently, starting with employee onboarding, third-party risk management and model life cycle governance use cases,” Experian said in a news release.

A major challenge for companies adopting agentic AI is achieving scale, the release added, with a lack of trusted data often holding back deployments, something industry research shows is the chief barrier for most organizations.

“By connecting trusted intelligence directly into enterprise workflows, this partnership enables agentic AI to scale well beyond pilot deployments,” the release said. 

Connecting the Experian Ascend Platform to the ServiceNow AI Platform lets AI agents access Experian’s insights and decisioning capabilities directly within existing workflows, giving customers the opportunity to automate intelligence at scale.

Keith Little, president of Experian Software Solutions, said the partnership comes as agentic AI is changing how intelligent services are provided.

“By connecting our intelligence and decisioning capabilities in Ascend directly into ServiceNow’s workflow, businesses can operate with confidence at scale, while extending the impact of our capabilities into new industries and enterprise workflows,” Little said.

The companies say their collaboration will support a range of use cases for businesses in highly regulated environments, beginning with third-party risk management — including fraud and identity verification for businesses, employee onboarding and model risk management.

Meanwhile, PYMNTS wrote earlier this month that ServiceNow is part of a group of companies that are “drawing new lines around the customer data stored inside their platforms” as external AI agents start to “erode the per-seat pricing model that has defined enterprise software for two decades.”

This came after ServiceNow introduced Action Fabric, a new integration layer that external AI agents must pass through to access data and execute workflows within its platform.

AI agents, that report said, break the traditional alignment found in software-as-a-service (SaaS) pricing, where each license was tied to an employee, a department and a cost center.

“AI agents break that alignment. A single agent can trigger thousands of API calls in a day while adding no new seats,” the report said.

As covered here, enterprise AI is replacing “predictable per-seat billing with consumption models that behave less like subscriptions and more like utility invoices, leaving finance teams to manage spend that fluctuates with model activity rather than headcount.”

In an interview with Bloomberg News’ Odd Lots podcast released Saturday (May 16), Collison described agentic commerce as an extension of earlier efforts to reduce eCommerce friction.

“When you find the product at the very end, do you really then want to go and be filling out all these web form fields and things like that?” asked Collinson whose company debuted its own agentic commerce offering last year. “Or do you want to just say, ‘Yeah, that sounds good, buy it for me in this size’?”

He argued people will want the low friction option, and that the history of technology shows that lower-friction options tend to succeed.

“And so we think that clearly will happen and is happening to some degree already,” said Collison, who founded Stripe with his brother, CEO Patrick Collison.

Collison added that he distinguishes between mundane tasks that AI agents should handle (like purchasing recipe ingredients or travel adapters) and “fun” activities that humans want to keep, such as scrolling through online clothing selections or planning a vacation.

In general, Collison said, people don’t want robots to take over their “scrolling jobs.” He also touched on the idea that traditional keyword search is outdated for complex shopping like furniture or clothing.

“It’s ridiculous that we got to the year 2026 relying on keyword search where that makes sense for buying a book or a DVD, where you know the title, but that’s about the limit of keyword search,” he said.

Instead, artificial intelligence allows for textual, constraint-based research, such as finding furniture that fits specific dimensions.

This shift may benefit smaller, niche brands because AI models — which have “read the whole internet,” as Collison put it — uncover high-quality products that aren’t necessarily at the top of traditional aggregator pages or SEO-driven lists.

As PYMNTS wrote last week, agentic commerce has come a long way in the past year. In January, Google debuted its Universal Commerce Protocol, and has since introduced agentic commerce through several high-profile retailers. 

Still, there are some large, lingering concerns, the report added. Agentic commerce involves turning a lot of information over to the AI platforms and payment companies supporting those transactions. 

“For agents to be able to make purchases, the Visas and American Expresses of the world need to know a lot about the people they’re acting on behalf of,” PYMNTS wrote. 

“Shoppers have their doubts. Almost all (95%) consumers have at least one concern about agentic commerce. Those worries range from simple mistakes like buying the wrong item to higher-stakes issues like identity theft.”

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.