The integrations include Spotify, Uber, Uber Eats, Instacart, Intuit TurboTax, Intuit Credit Karma, TripAdvisor, Booking.com, AllTrails, Audible, Resy, StubHub, Taskrabbit, Thumbtack and Viator. Since launching in July, the Claude connector directory has grown to over 200 integrations, with users frequently connecting multiple apps and using them together inside a single conversation, according to Anthropic.

The move brings Claude into direct competition with OpenAI and Google. As CNET reported, both Gemini and ChatGPT already allow third-party app connections, with ChatGPT linking to tools like Canva and Zillow. Historically, Claude’s connectors have been work-focused. This launch is geared toward daily consumer use.

The architecture of Claude’s connector system differs from how competitors have approached third-party integrations. Gemini connects primarily to Google’s own app ecosystem. ChatGPT relies on users selecting tools manually.

In contrast, Claude now suggests the right app based on what the user is doing inside the conversation, whether finding a reservation, adding to a grocery cart or identifying a flight, working from the user’s stated preferences and conversational context, according to Anthropic. The user doesn’t browse or select. When more than one connected app could help, Claude shows both and lets the user choose.

Claude is ad-free and plans to stay that way. There are no paid placements or sponsored answers in conversations, and when two connectors could help, both are shown ranked by usefulness to the user, the company said. Before completing any booking or purchase, Claude is designed to check with the user first.

That confirmation-before-action design matters for enterprise and consumer trust alike. The model recommends. The user decides. The app executes.

AI as a Transaction and Commerce Layer

The connector model turns Claude into an orchestration layer above individual apps. A conversation about weekend plans can produce a hiking recommendation from AllTrails, a restaurant booking from Resy, a grocery order from Instacart, and a ride from Uber without the user switching between apps. Each step represents a potential transaction.

The grocery use case illustrates the complexity that makes this harder than it appears. Instacart’s CTO Anirban Kundu described the challenge in a company post published alongside the launch. Grocery involves tens of thousands of SKUs per store, inventory availability that changes by the minute, and deeply personal household preferences. Most AI grocery experiences fail because they surface plausible-sounding results that don’t reflect what’s actually in stock, what things cost at a specific store, or what a household actually buys.

What makes Instacart’s integration viable inside Claude is its underlying infrastructure: data pipelines into more than 2,200 retail banners, a product catalog with more than 2 billion items, and a personalization engine called Smart Shop that tracks how preferences shift by season, occasion and household. When a user builds a grocery cart inside Claude, it connects to real inventory from nearby stores. Any cart built in the conversation syncs automatically to the user’s Instacart account.

Competitive Landscape and Platform Implications

The connector expansion puts Anthropic on the same consumer terrain as OpenAI and Google at a moment when all three are competing to become the default layer for everyday decisions. The execution model is where Anthropic is making its bet. Suggestion-driven discovery inside a conversation is a different distribution model than a tool store where users choose integrations manually.

The more apps a user connects, the more Claude can do. Claude suggests relevant connectors as the user works, which can be installed with a click on desktop or a few taps on mobile, according to Anthropic. Once connected, a service is available in every conversation.

Connectors are available across all Claude plans. Mobile access is currently in beta. Anthropic said it will continue expanding the directory and invited developers to submit their products for inclusion.

For all PYMNTS AI and digital transformation coverage, subscribe to the daily AI and Digital Transformation Newsletters.

This edition of a new series, “The ABCs of AI Credit: A Playbook for Issuers,” shows that AI agents are emerging as the operating layer for this shift. Rather than acting as gatekeepers, these agents function as a cognitive layer embedded directly into payment flows, evaluating transactions in milliseconds using behavioral signals and real-time data. This enables more adaptive, context-aware transaction intelligence.

The impact on issuer performance is immediate. Traditional fraud controls often block legitimate transactions that fall outside predefined patterns, leading to false declines and lost revenue. AI agents analyze a broader set of signals to distinguish genuine behavioral changes from high-risk activity, improving authorization rates while keeping fraud in check.

Just as important, they introduce flexibility into transaction logic. Instead of a binary “yes” or “no,” issuers can trigger real-time alerts, adjust limits dynamically or step up verification when needed—balancing security with a smoother customer experience.

This shift also depends on infrastructure. Application programming interface (API)-first, real-time processing environments provide the foundation for activating intelligence at the point of transaction. Without that layer, AI-driven intelligence cannot operate at scale.

Ultimately, the playbook reframes how credit operates in live payment environments. Decisions no longer act as isolated checks—they now determine whether customers can access and use their available funds in the moment. In an always-on economy, performance is shaped continuously, one transaction at a time.

About “The ABCs of AI Credit: A Playbook for Issuers”

The inaugural edition of the “ABCs of AI Credit” Playbook, a PYMNTS Intelligence collaboration with Thredd, provides issuers with a practical framework for deploying AI agents in transaction flows, outlining how to move from static rules to real-time transaction intelligence while improving approvals, reducing friction and protecting customer liquidity.

Artificial intelligence is moving deeper into payments, and at an unusual pace. It’s helping companies spot fraud, improve approvals, personalize offers, support compliance, manage risk and shape customer experiences in real time. That much is clear.

What’s less settled is the part that now matters most: Who is governing these systems, how that governance works in practice and what happens when AI begins influencing decisions faster than an organization can explain them.

Across the essays in this eBook, one message comes through with force. In payments, AI is no longer an experiment. Governance is the differentiator.

That’s because the payments environment is one of the hardest places to get AI wrong. These systems do not live in a lab. They operate inside live transaction flows, fraud programs, onboarding journeys, identity checks, credit decisions and customer service interactions.

A weak model can create problems. A weak governance structure can multiply them. At scale.

The risk is not only that AI makes a flawed call. It’s that nobody is fully accountable for the outcome, nobody can trace the logic and nobody spots the drift until customers, regulators or partners do first.

The executives in this collection return again and again to a handful of hard truths. Governance tends to break down in the gaps between teams. Product may own the feature, engineering may own the model, compliance may own the policy and operations may own the day-to-day consequences, but the end-to-end accountability is often blurred.

At the same time, many AI systems depend on third-party models, vendors, data providers and external platforms. This means companies are being asked to govern not only what they build, but also what they rent, ingest and rely on. In that environment, governance can’t be treated as a meeting, a checklist or a slide deck. It has to become an operating discipline.

Another theme runs through these pages: Speed is seductive, but speed without structure is expensive. Payments companies are under pressure to automate more, move faster and show returns quickly. Yet several contributors make the same point from different angles. The real work is not simply deploying AI. It’s building the data foundations, oversight mechanisms, fallback plans, audit trails and human review points that let a company move quickly without losing control. In other words, the organizations that benefit most from AI may not be the ones that rush first, but the ones that prepare best.

Readers should come away from this eBook with more than a warning. They should come away with a playbook. These essays offer a practical look at where governance breaks down, what strong organizations are doing earlier, which questions boards and CEOs should be asking now and how leading payments executives are thinking about explainability, accountability and trust in an AI-driven market. For leaders across banking, payments and FinTech, that has real value. It can help sharpen internal conversations, expose blind spots in current governance models, and frame AI not as a race to adopt the newest tool, but as a long-term test of institutional discipline.

AI may be powering the next era of payments. But governance will decide which companies can scale it with confidence, which ones can defend it under scrutiny, and which ones can turn automation into a durable advantage. That is the conversation this eBook begins.

According to the company’s announcement, the update is being made available to a limited group of customers through its Frontier program, which gives early access to new AI tools.

Copilot Cowork is designed to handle work that takes more than a single prompt. Instead of asking the assistant for one answer at a time, users can assign a task and have Copilot complete it across apps like Outlook, Teams and Excel. The system can gather information from emails, documents and chats, organize it and produce a finished output.

Microsoft said the feature is meant to help users manage ongoing work by keeping track of context across tasks. It allows Copilot to continue working on assignments over time, rather than treating each request as a separate interaction.

Alongside Cowork, Microsoft introduced updates that allow Copilot to use more than one AI model when generating responses. With “Critique,” the program can produce an answer using one model and then review it using another before showing it to the user.

The goal is to improve accuracy and reduce errors by adding a second layer of review. Microsoft also introduced a feature called “Council,” which lets users compare responses from different models side by side.

These additions are intended to give users more visibility into how answers are generated and to provide a way to check results before using them.

The Copilot Cowork feature is currently limited to customers in Microsoft’s Frontier program. This program allows select users to test new AI tools and provide feedback before they are released more widely.

Microsoft said it plans to expand access based on user feedback and testing results. The company has been steadily adding new features to Copilot as it looks to increase usage of its AI tools across its workplace software. The release follows Anthropic’s introduction of its own “Cowork” capability, which allows its Claude AI system to act as a collaborator across tasks, as PYMNTS reported in January.

For all PYMNTS digital transformation coverage, subscribe to the daily Digital Transformation Newsletter.

While CEOs talk about AI’s impact across automated outreach, predictive pipelines and agents that can draft proposals in seconds, what is vanishing isn’t friction but the B2B sales process itself.

Procurement leaders inside large organizations are faced with a declining volume of formal request for proposals (RFPs) and vendor calls that are happening later in the process, if at all. It’s not that decisions are speeding up for procurement leaders, but the purchasing decision itself is frequently already made before the sale details ever reach their desk.

Procurement teams equipped with AI-driven tools can now independently conduct market scans, compare vendors, evaluate pricing benchmarks and even simulate negotiation scenarios. What once required weeks of vendor engagement can now be executed in hours without a single sales call.

This transition from seller-led narratives to model-led decisions may fundamentally alter the role of sales.

Read also: It’s Contract Renewal Season, Are Your B2B Payment Strategies Ready?

The Rise of Zero-Touch B2B Buying

For decades, procurement functioned as the institutional backbone of B2B commerce. It imposed order on a messy landscape of vendors, claims and pricing models by creating checkpoints that ensured decisions were deliberate and defensible. Just as importantly, it created access.

Sales teams knew where to enter the organization, who controlled budgets, and how decisions would unfold. Procurement was not just a control mechanism; it was a map.

That map is now out of date. The conditions that made procurement and sales checkpoints indispensable, like limited information, opaque pricing and high uncertainty, have been increasingly eroded by digital tools.

PYMNTS Intelligence data in the report, “The Investment Impact of GenAI Operating Standards on Enterprise Adoption,” a PYMNTS Intelligence study produced in collaboration with Coupa, shows that 75% of companies are now considering using AI in procurement.

In this emerging procurement landscape, access is determined less by relationships and more by data compatibility and system integration. Procurement AI relies on structured, reliable and comparable data. Vendors that provide clear pricing models, standardized specifications, performance metrics and interoperable data formats are far more likely to be included in automated evaluations.

“The friction points tend to be in the operational environment,” Rene Stynen, senior vice president, EMEA, B2B Payments at Boost Payment Solutions, told PYMNTS. “Supplier enablement, onboarding complexity and the amount of data gathering required. Those things can become quite complex.”

“There are expectations both on buyer sides and supplier sides for things to become a little bit more digital and automated,” Stynen added.

See also: Agentic Commerce Finds Traction in B2B Payments 

The End of Information Asymmetry

A central pillar of traditional B2B sales has been information asymmetry. Sellers often possessed deeper knowledge about their products, pricing structures and market alternatives than buyers, enabling them to guide, and sometimes control, the decision-making process.

AI effectively eliminates this asymmetry. Procurement teams now have access to aggregated market intelligence, real-time pricing benchmarks, peer reviews and predictive analytics. They can test assumptions, validate claims and explore alternatives without relying on vendor input.

As a result, the informational advantage that once justified extensive sales engagement has eroded. Vendors are no longer primary sources of insight; they are data points within a broader analytical framework. Tools that aggregate vendor performance data, simulate ROI scenarios and flag implementation risks are becoming embedded in everyday workflows. Instead of asking vendors for information, buyers are querying systems that already contain it.

And as routine decisions become increasingly automated or self-directed, the role of sales may shift toward exceptions—complex deployments, cross-functional alignment and bespoke solutions that cannot be easily standardized. The center of gravity moves from selling to enabling.

Additionally, on the buyer side, by the time procurement is involved, its role is often reduced to validation: checking compliance, finalizing contracts and processing payment.

Still, AI in procurement is in its early innings, and there remains a lot of time on the game clock.

“If you’re going to experiment with agentic AI or any type of AI solutions, you want to focus on two things. One is the area where you’re most likely to have success. And two, is there going to be a good return on that investment?” WEX Chief Digital Officer Karen Stroup told PYMNTS in an earlier interview.

For all PYMNTS AI and B2B coverage, subscribe to the daily AI and B2B Newsletters.

A central finding is that progress now depends on solving practical problems that sit between technical capability and real-world adoption. Acquirers say merchants still face real obstacles, including integration costs, older systems and the work required to connect new tools to existing operations. The report also finds that trust remains a critical issue. As commerce becomes more automated, acquirers see fraud controls, identity verification and clear rules around responsibility as essential to wider adoption. Their message: While the opportunity is significant, scaling agentic commerce safely will require the payments ecosystem to align around standards, governance and risk management.

In “How Acquirers Prepare for Agentic Commerce,” learn how:

• Acquirers view the current payment infrastructure as a starting point for agentic commerce. The study shows that many acquirers believe existing rails can support AI-enabled transactions. What comes next is the work of adapting that foundation to handle new forms of authorization, oversight and control.
• Merchant-side barriers could slow the market even as interest grows. Many acquirers say the challenge is no longer just about consumer demand or technical possibility. Merchants also need simpler deployment paths, lower implementation burdens and stronger support for connecting new capabilities to existing systems.
• Trust and accountability will shape how quickly agent-led payments scale. Acquirers see secure credentials, identity checks and fraud prevention as core requirements for broader agentic commerce adoption. They also make clear that clearer rules around liability and governance will help determine how confidently businesses move forward.

About the Report

“How Acquirers Prepare for Agentic Commerce,” a PYMNTS Intelligence report in collaboration with Visa Acceptance Solutions, examines the critical factors that drive agentic commerce readiness, drawing on insights from a survey of 75 acquirers across Brazil, the UAE, and the U.S.

Comprising information collected from Jan. 16, 2026, to Jan. 30, 2026, the report provides a roadmap for acquirers (and ecosystem partners) seeking to enhance customer relationships and gain a competitive edge in the agentic commerce arena.

This research was independently designed, fielded, analyzed and written by PYMNTS Intelligence. Research partners provided funding support but exercised no control over methodology, data collection, findings or conclusions.

Huang said Monday (March 16) during his keynote address at the Nvidia GTC AI Conference & Expo that tokens are becoming a recruiting tool in Silicon Valley, according to the report. Tokens are units of data used by AI systems, the report said.

“I could totally imagine in the future every single engineer in our company will need an annual token budget,” Huang said during his keynote. “They’re going to make a few hundred thousand dollars a year, their base pay, I’m going to give them probably half of that on top of it as tokens so that they could be amplified 10x. Of course we would. It is now one of the recruiting tools in Silicon Valley — ‘How many tokens comes along with my job?’”

Huang told CNBC in February that a growing number of AI agents will work alongside Nvidia employees.

“I have 42,000 biological employees, and I’m going to have hundreds of thousands of digital employees,” he said, per the report.

Huang also told CNBC in February that AI agents will have a beneficial impact on the software industry. Rather than reducing demand for the industry’s products, AI agents will become its customers as they use programs, tools and computing resources, he said.

“The number of C-compilers that we use, the number of Python programs that we have, the number of instances, are growing very, very fast — because the number of agents we have that use these tools are going up,” Huang said, per the report.

PYMNTS reported Thursday (March 19) that Huang said during his keynote address that companies will shift from software that enables employees to do work, to software that does the work itself, autonomously, through AI agents executing tasks without continuous human input.

Nvidia said in reports released earlier this month that AI is delivering measurable financial gains for businesses. Eighty-eight percent of organizations say AI has increased their annual revenue, and 87% say it has reduced their costs. In terms of AI adoption, 64% of companies are actively using AI, while 28% are evaluating potential deployments.

The company introduced Wednesday (March 11) Personal Computer, a system designed to allow AI agents to interact with files, applications and online services rather than simply responding to prompts in a chat interface. The system runs on Apple’s Mac mini and can operate around the clock while connecting to a user’s local applications and Perplexity’s servers.

Perplexity described the platform as a kind of digital proxy that works continuously on a user’s behalf. Instead of waiting for instructions each time a task appears, the system can coordinate tools, files and workflows across devices while carrying out tasks in the background.

The launch expands the power of Perplexity Computer, which debuted last month, “across personal workflows, enterprise software, developer platforms and finance,” Perplexity said in a statement. “The throughline is the same in each case: a system that can understand the goal, gather the right context, use the right tools, and carry the work forward. Everything is Computer.”

Personal Computer operates within a controlled environment designed to limit risk. Sensitive actions require user approval, each session generates an audit trail, and a kill switch allows users to immediately stop activity.

The system is initially being rolled out through a limited waitlist and is expected to be offered through a subscription model priced at about $200 per month, Decoder reported Friday (March 13).

Beyond individual users, the company is also positioning the technology for businesses through a version called Computer for Enterprise. In internal testing of more than 16,000 queries, measured against institutional benchmarks used by organizations such as McKinsey, Harvard, MIT and Boston Consulting Group, Perplexity said the system completed what it estimated to be 3.25 years of work in four weeks, saving roughly $1.6 million in labor costs.

The enterprise system connects directly to business software through application connectors, allowing teams to query platforms such as Snowflake, Salesforce and HubSpot while simultaneously analyzing other internal or external data sources.

A financial analyst, for example, could ask the system to retrieve revenue data from Snowflake while combining it with market analysis, while a sales team could pull CRM data alongside competitive intelligence.

The platform runs on a secure foundation that includes SOC 2 Type II compliance, SAML single sign-on and audit logs, with each query executed inside its own isolated environment.

Developers Turn to Mac Mini to Run AI Agents

Perplexity’s system runs on a dedicated Mac mini, although the hardware itself is not essential to the concept and could eventually be replaced by other always-on machines designed to host AI agents.

Mac minis have become increasingly popular among developers who want to run AI agents locally rather than relying entirely on cloud infrastructure, Tech Rader reported Feb. 17. Mac minis have become harder to obtain in certain markets as programmers buy the machines specifically to run AI agent systems.

Developer communities are increasingly recommending Mac minis for running AI agents because they combine strong computing performance with relatively low energy consumption, open-source vector database Milvus said in a blog post. Developers often treat the device less like a personal computer and more like a small server dedicated to running AI workloads.

The growing use of Mac minis to run AI agents reflects a broader shift in how AI software operates. As new agent frameworks emerge, computers may evolve from tools used occasionally into environments where AI agents run continuously, managing tasks and interacting with digital systems throughout the day.

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

The company pushed the release to at least May from a planned debut this month, The New York Times reported Thursday (March 12).

The delay follows internal testing that showed the model trailing leading systems from Google, OpenAI and Anthropic in key areas, including logical reasoning, programming and writing, the report said.

Avocado outperformed Meta’s previous generation models and some earlier competing systems, The Information reported Feb. 4, but it failed to match the performance of Google’s newest Gemini models, per the NYT report.

The gap matters because Avocado was not intended as an incremental upgrade. The model was designed to compete directly with frontier systems from OpenAI and Google and to serve as the centerpiece of Meta’s next phase of AI development.

Internally, the model had been framed as a major leap forward. Meta Superintelligence Labs Product Manager Megan Fu described Avocado in an internal memo as the company’s most capable base model yet and suggested it had the potential to outperform rival systems once additional post-training improvements were applied, per the report from The Information.

The production version has not yet met that expectation.

The delay also follows a difficult year for Meta’s AI efforts. The company’s Llama 4 release last year failed to generate strong developer enthusiasm, prompting an internal restructuring across parts of the AI organization, The Information reported Aug. 15.

Avocado had been positioned internally as the reset.

The Gemini Question and What It Signals

The more revealing development may be what Meta is considering in the interim.

The company’s leadership discussed temporarily licensing Google’s Gemini technology to power certain Meta products while Avocado is brought up to competitive performance, the NYT reported. No decision has been confirmed.

Meta has spent years positioning open-source AI models such as Llama as a strategic alternative to proprietary systems from OpenAI and Google. Relying on Gemini, even temporarily, would invert that narrative and raise questions about how quickly Meta’s internal models can reach the top tier of AI performance.

The episode also underscores a broader strategic shift underway inside the company. Meta’s earlier AI strategy emphasized open development and ecosystem adoption through models such as Llama. Avocado represents a move toward more tightly controlled, commercial-grade systems that could power proprietary products.

That transition was already complex. A performance gap only complicates it further.

Meta’s broader AI roadmap remains active. A next-generation model codenamed Watermelon is under development, the NYT reported, along with an image and video generation system known internally as Mango.

However, the timeline now appears less certain than the ambition.

Spending Without a Cloud Business

Beyond the technical delay, Avocado highlights a deeper financial question surrounding Meta’s AI strategy.

The company outlined capital spending plans of between $115 billion and $135 billion for 2026 as it races to build data centers, chips and infrastructure capable of supporting increasingly powerful AI models.

That level of spending places Meta alongside the largest investors in AI infrastructure, including Amazon, Microsoft and Google.

The difference is that those companies operate cloud businesses that directly monetize the computing power they build. Training models for internal use is only one part of the equation. The same infrastructure can also be rented to thousands of enterprise customers.

Meta does not have that outlet. Instead, the company’s AI monetization strategy is expected to run primarily through improvements to its existing platforms. AI is already being used to refine ad targeting, generate content recommendations and power assistant tools across Facebook, Instagram and WhatsApp.

Those applications could drive revenue growth by increasing engagement and advertising efficiency. However, they do not produce the same direct revenue stream as selling AI compute or model access through the cloud. During Meta’s fourth-quarter earnings call, CEO Mark Zuckerberg positioned AI as the company’s next long-term growth engine. Avocado was supposed to provide the near-term proof.

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

That shift was at the center of “How AI Becomes the Place Consumers Start Everything,” the December installment of PYMNTS Intelligence’s Agentic AI Report series. The report was based on a survey of 2,113 U.S. adult consumers and tracked adoption across 54 personal-use tasks in nine areas of daily life, including shopping, finances, health, education and travel.

What stood out was not just that people are using AI, but how they are using it. Consumers are turning to dedicated AI platforms to plan, learn, shop and decide, compressing the familiar search-to-purchase sequence into a more conversational flow where intent is stated once and then refined through follow-up prompts.

Key findings from the report include:

• More than 6 in 10 consumers used dedicated AI platforms in the past year.
• More than one-third of Generation Z consumers and Power Users (consumers who perform 25 or more distinct tasks) turned to a dedicated AI platform as their first tool when tackling personal tasks.
• Among consumers who primarily use dedicated AI platforms, 43% reported fully replacing older methods, and 42% said they now use search less.

The task-level story helps explain why the “AI-first” pattern is taking hold, the report said. People start with jobs where a conversational interface saves time, and the downside of a bad answer is limited. The report pointed to personal tasks as the main driver of use, and it described Power Users as extending AI across shopping discovery, planning, learning and wellness.

Those are practical, repeatable activities, like drafting and polishing writing, mapping out an itinerary, summarizing information, narrowing product choices, or building a simple plan for a goal. Over time, those small wins create what the report called behavioral “muscle memory,” which can make AI-mediated decisioning feel normal once trust and user experience improve.

The report also showed why not everyone is moving at the same speed. PYMNTS Intelligence grouped consumers by how many distinct AI tasks they perform and how complex those tasks are, yielding four personas, including holdouts, Light Users, Mainstream Users and Power Users. Power Users represented 11% of consumers, while holdouts represented 43% of consumers. Adoption also varied by age. Gen Z was the least resistant to trying the technology, with 71% having experimented with it.

Where the next wave gets interesting for payments is in the bridge from helping consumers decide to helping them pay.

Consumers signaled that the most credible path is not storing a card inside an AI app. They lean toward familiar protections. One-third prefer linking a digital wallet to an AI platform for safer, easier payments, and the report said wallets could become the trust layer that makes AI-mediated commerce viable at scale.

Trust remains the gating factor. Consumers cited privacy and misunderstandings as key barriers, which is why disclosures, user-controlled permissions and a way to escalate when a model is uncertain matter. The encouraging part is that the everyday tasks are already building momentum. This is how habits change, slowly and then quickly.

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

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.