While some callers were asking about delivery times or return policies, others wanted help designing their homes. No chatbot was going to solve that.

Ingka Group, the largest IKEA franchisee, deployed its AI assistant, Billie, across customer service channels that year. By 2023, the chatbot was handling roughly 47% of all inbound inquiries, about 3.2 million conversations covering product information and recommendations. The operating savings came to nearly 13 million euros (about $15 million).

Reading the Unresolved Queue

IKEA then looked at the other 53%.

Customers were reaching out for help designing rooms, problems that required taste and contextual judgment to solve. Rather than deprioritizing these requests, Ingka treated them as a demand signal.

The company launched a reskilling program, converting roughly 8,500 call center employees into remote interior design consultants. Workers were retrained in digital retail sales, room planning and relationship management. Billie handled the operational load. The reskilled workforce handled conversations that required human judgment.

A Cost Center Becomes a Revenue Line

The remote customer meeting channel generated 1.3 billion euros (about $1.5 billion) in revenue by the end of fiscal year 2022, representing 3.3% of Ingka Group’s total sales. Ingka set a target of growing that share to 10% by 2028, in part as a strategy to attract young customers, Reuters reported at the time.

The chatbot savings was predictable. The revenue line was not. It came from reading customer service data as a demand map rather than a cost metric.

Ingka reported 41.5 billion euros (about $48 billion) in total revenue for fiscal year 2025, with retail sales of 39 billion euros (about $45 billion), as it absorbed cost-of-living pressures across its major markets. Online visits rose 4.6%, and units sold grew 1.6%. The company helped over 73,000 customers with remote furniture and kitchen planning expertise.

A Different Workforce Model

Most companies haven’t built a comparable program. The PYMNTS Intelligence report “No Roadmap, No Problem: How Enterprises Are Reinventing the AI Workforce,” based on a study of chief financial officers at large firms in the United States, found that half expect AI to create new roles requiring new skills, and 47% expect it to significantly reduce headcount. Only 12% said their organizations feel very prepared to manage the shift.

CFOs reported investing in AI tools faster than they are adjusting workforce strategies, leaving most companies without a clear plan for retraining, redeployment or job redesign.

Ingka built the program. Beyond Billie, the company launched an AI literacy initiative targeting 30,000 workers, with more than 4,000 trained during fiscal year 2024.

Bain & Company found that “beyond trade” activities, or services adjacent to core retail transactions, accounted for 15% of sales and 25% of profit at a typical U.S. or European retailer in 2024, up from 10% in both cases in 2021.

Ingka’s next test is whether the channel reaches 10% of revenue by 2028.

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

Artificial intelligence models are now beginning to edge into parts of that relationship, turning conversational AI into another destination where consumers can monitor spending, analyze transactions and manage financial decisions.

The shift does not mean AI models would replace aggregators outright. It does, however, introduce the possibility of new competition around who controls the consumer-facing layer of financial engagement.

Last week, OpenAI introduced personal finance capabilities inside ChatGPT that allow users to connect financial accounts and receive responses grounded in transaction activity and account information, linking the accounts via Plaid. In April, Perplexity expanded its partnership with Plaid to let users connect checking accounts, credit cards and loans directly into its platform for spending analysis and financial management tools.

These developments move conversational AI closer to the territory traditionally associated with financial data aggregators. The announcements spotlight the control of the interface where consumers increasingly ask financial questions.

If consumers begin asking ChatGPT or Perplexity questions such as, “Why was my checking account balance lower this month?” or “Which subscriptions are costing me the most money?,” the AI platform becomes the place where financial analysis and decision-making begin. Consumers may spend less time inside bank applications or personal finance dashboards and more time inside conversational systems interpreting financial activity in real time.

The Connective Infrastructure

AI models can summarize transactions, identify spending patterns and explain cash flow changes, but those functions still depend on access to financial data in the first place. Aggregators continue to provide the infrastructure responsible for account authentication, permissions management, transaction normalization and identity verification across thousands of financial institutions.

OpenAI said its financial account connectivity is supported through Plaid’s network of more than 12,000 financial institutions, while Intuit is expected to support the feature as well. Perplexity’s expanded finance tools similarly rely on Plaid’s infrastructure to connect users to accounts and transaction histories.

That underlying data layer becomes increasingly valuable as AI systems attempt to generate useful financial guidance.

Deposit account activity, including payroll deposits, recurring bill payments, subscription spending and debit card transactions, often provides a more updated picture of a consumer’s financial condition than traditional credit bureau data alone. Checking account information also allows AI systems to identify spending trends, forecast cash balances and flag unusual account activity in ways that static financial snapshots cannot.

The operational challenge is that financial data remains fragmented across banks, credit unions, card issuers and payment systems. Aggregators built much of the connective framework required to unify that information.

Plaid said its systems support nearly 1 million new account connections. It plans to expand connectivity into areas like cryptocurrency wallets and property-related financial data.

Beyond the current interdependence, the competitive pressure could intensify if AI firms eventually pursue direct bank integrations or proprietary data-sharing systems in a bid to control the customer relationship and, by extension, monetization opportunities attached to it.

Data Sharing and Security Questions

The expansion of conversational finance also introduces new concerns around consumer consent, fraud exposure and operational oversight.

Consumers may be comfortable using AI assistants to review subscriptions or summarize spending categories. Giving autonomous systems authority to move money, authorize transactions or execute financial actions presents a different level of risk.

Only about 1 in 5 consumers would allow an autonomous AI agent to manage banking activity on their behalf, according to PYMNTS Intelligence data. The hesitation reflects broader concerns around unauthorized access, data misuse and the concentration of sensitive financial information inside conversational platforms.

In separate coverage on the specter of bank runs, PYMNTS reported Friday (May 15) that AI-driven fraud and automated account attacks could place additional pressure on financial institutions if compromised systems trigger false account activity, blocked payment access or wider operational disruptions.

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

The May Agentic AI Report from PYMNTS Intelligence, “The New AI Handshake: Data Shows When Consumers Want Help and When They Want Control,” found that AI is moving from an occasional research tool to something closer to a daily household assistant, one that helps consumers plan, compare and spend.

Based on a survey of 2,111 adults in the United States in March, the report covered AI use across nine activity categories and 54 individual activities. It revealed that AI use for household tasks is accelerating. Finding discount codes or deals grew 2.9 percentage points from the October-November average. Managing household logistics rose 2.8 points. Meal planning and grocery lists climbed 2.7 points.

However, a separate finding complicates the story. Even as more consumers encounter AI embedded in merchant apps and third-party websites, fewer are calling those tools the most helpful. Embedded AI lost ground as the top-rated tool in six of eight task categories. The steepest drops came in learning and self-improvement, down 12.1 percentage points, and health and wellness, down 7.1 points.

Dedicated AI platforms filled the gap. They gained ground in health and wellness, up 6.2 points, everyday planning, up 3.9 points, shopping and purchasing, up 3.5 points, and finance and banking, up 3.1 points.

The report offered a clear explanation. A merchant app may know its own inventory well. But consumers often begin with discovery, comparison and context before they narrow their choice to a specific product or service. Dedicated platforms offer that breadth. Embedded tools, by design, do not.

For payments and commerce companies, the implication is direct. Putting AI inside an app does not guarantee consumers will find it more useful. It only guarantees they will encounter it. The helpfulness credit, and the trust that follows from it, may go somewhere else entirely.

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.

In an interview with Bloomberg TV, Friar said OpenAI may do so because of the need for compute. She said the company is seeing “a vertical wall of demand,” but there is “not a lot of compute in 2026,” so the company is looking into how to get the compute it will need.

“It will depend somewhat on matching the demand that we’re seeing, the revenue that we’ll create, and hence the cash flows, and then that gap to wanting to buy compute out into future years,” Friar said. “So, I’m not averse to raising more capital, although right now that $122 billion gives us a lot of optionality, and I think derisks a lot of what we need to go do in the market at the moment.”

Asked about reports that OpenAI had not met some of the stretch goals it had set for itself in terms of customer acquisition, Friar said that on the consumer side, the company is “delighted with how the consumer platform is going,” and that on the enterprise side, “our sales team has just run ragged at the moment” due to companies wanting to transform their businesses.

“We just see our business going from strength to strength right now,” Friar said.

Addressing reports that OpenAI’s partnership with Apple has frayed to the point that OpenAI is considering a lawsuit, Friar said she cannot comment on legal action.

“Right now, in terms of our relationship with Apple, it’s like every other large distribution partner for consumers: We want to make it work, and we know that consumers want our technology,” Friar said. “So, we think that there’s a way.”

Friar also mentioned that AI models are in limited release because they can find vulnerabilities in software code at an unprecedented level.

“I’ve been spending a bunch of time with banks and financial institutions; I have not talked to a single CEO of a bank that does not have this as their top priority currently,” Friar said.

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

Three months after raising $30 billion at a $350 billion valuation, the company is on the verge of nearly tripling that figure, the Financial Times reported Friday (May 15), leapfrogging rival OpenAI in the process.

For payments and FinTech professionals, the numbers signal that enterprise AI spending is accelerating fast enough to reshape capital markets in real time.

Anthropic agreed terms on a new $30 billion funding round that would value the company at $900 billion, not including the new capital, the report said. The round is expected to close as soon as this month.

Dragoneer, Greenoaks, Sequoia Capital and Altimeter Capital agreed to co-lead the raise, with each firm expected to invest $2 billion or more, according to the report. Anthropic is still in discussions with additional investors to fill out the remainder of the round, and final terms could change before a formal announcement.

The round came together quickly, per the report.

“Anthropic was approached by investors last month, and chief financial officer Krishna Rao initiated talks with prospective backers in the past two weeks,” the report said.

Anthropic’s annualized revenues are expected to cross $45 billion imminently, a fivefold increase from $9 billion at the end of last year, according to the report. Anthropic’s revenues appear to have surpassed OpenAI’s, although the two companies use different accounting methods.

Big Tech is not expected to participate in the current round, although Amazon and Google have previously provided tens of billions of dollars in backing, the report said. Three of the four lead investors, Dragoneer, Sequoia and Altimeter, are also backers of OpenAI, valued in March at $852 billion.

Earlier this month, Anthropic debuted 10 new financial services-focused AI agents designed to automate tasks, including know your customer (KYC) checks and pitchbook creation, while expanding partnerships with Dun & Bradstreet, Verisk and Moody’s.

Also this month, the company launched Claude for Small Business, which wires Claude into QuickBooks, PayPal and HubSpot to handle invoicing, payroll planning and contract routing.

In March, Anthropic debuted the Claude Marketplace, which positions the company as a central procurement hub for enterprise AI tools built on its models.

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

Chatbots would redefine service, robo-advisors would democratize wealth management, and sleek digital assistants would become the new interface between banks and their customers.

New insights from the May edition of The Enterprise AI Benchmark Report by PYMNTS Intelligence, however, revealed a different story emerging. Financial firms are not just experimenting with AI; they’re operationalizing it at scale, and in the least visible parts of the enterprise, including the core systems that determine how work gets done.

The report highlighted an inflection point around the transition from isolated use cases to integrated systems. Financial institutions are not adopting AI more broadly; they are adopting it more deeply. The emphasis is on back-office functions such as compliance, underwriting, fraud detection and operational workflows, where data is structured, outcomes are measurable and the return on investment is easier to quantify.

In that sense, the AI race is no longer just about technology. It is increasingly about execution, integration and the ability to turn potential into performance.

The Back Office Is Becoming AI’s New Proving Ground

It is tempting to view the back office as a secondary domain, far removed from innovation. In practice, it is precisely where the most consequential changes are taking place. Financial services firms have long operated in environments defined by regulatory scrutiny, risk management and data intensity. These conditions make the back office uniquely suited for AI deployment.

Once AI becomes embedded at this level, it begins to behave less like a tool and more like infrastructure. Decisions that were once episodic become continuous. Processes that required human intervention become self-adjusting systems. Over time, the distinction between using AI and running on AI starts to collapse.

What emerges is a feedback loop. As AI systems improve operational efficiency, they generate more data. That data, in turn, refines the models, further improving performance. Over time, this compounding effect creates a widening gap between firms that have successfully integrated AI into their core operations and those that have not.

Structured datasets enable more reliable model training. High-frequency decision environments create natural opportunities for automation. The cost of inefficiency, whether in fraud detection or compliance, is high enough for firms to justify sustained investment.

Read the report: Financial Services Pulls Ahead in the Enterprise AI Race

Yet the report also underscored a less comfortable reality. AI adoption is uneven. While some firms are scaling AI across dozens of use cases, others remain stuck in what might be called pilot purgatory, representing a cycle of experimentation without meaningful deployment.

It is no longer sufficient to say that an organization is investing in AI. The relevant question is whether that investment is translating into operational change.

The barriers are familiar but persistent. Data fragmentation limits the effectiveness of models. Organizational silos slow down implementation. Talent shortages constrain the ability to move from prototype to production. Cultural resistance, which is often underestimated, can stall even well-funded initiatives.

Operationalizing AI requires more than technical capability. It demands changes in process, governance and organizational design. It requires aligning incentives, rethinking workflows and building trust in automated systems. These are not trivial challenges, and they cannot be solved through technology alone.

Financial services offers a preview of this future. By focusing on high-impact, data-rich use cases, the industry has accelerated the transition from experimentation to scale. It has demonstrated that the real value of AI lies not in isolated applications, but in the transformation of systems.

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.

To find out, the grocer built an artificial intelligence tool that uses computer vision and Google’s Gemini models to grade fresh produce inside its distribution centers, replacing the inspector-by-inspector judgments that have long determined what clears the quality threshold.

The tool, called Intelligent Quality Control, launched in select Albertsons distribution centers this month. A quality inspector feeds an image of the produce into the tool, which evaluates visual characteristics against Albertsons’ internal grading standards and returns a rating and recommendation, according to a Wednesday (May 13) press release. The inspector makes the final call. It’s live now for strawberries and red and green grapes, with the full berry section next in line and a nationwide rollout planned.

The Inconsistency Problem

Produce quality inspection has always been a human problem with a human-shaped flaw. The same item might grade differently depending on the inspector, the shift, the warehouse or the hour. Across a network like Albertsons’ 22 distribution centers and 2,244 stores, small inconsistencies can compound.

Produce accounts for the largest share of surplus food generated by retailers in the United States, at 33.1% of the roughly 4 million tons that went unsold in 2024, ReFED reported. That’s not all traceable to grading decisions, but those decisions sit at the front of every freshness outcome that follows.

Albertsons Executive Vice President and Chief Supply Officer Evan Rainwater said in the release that early results showed the tool had reduced variability in quality ratings. The company didn’t release specific figures but said the system produces faster decisions and captures more granular quality data per inspection than the manual process allowed.

“This is just the latest advancement in how we are using AI within our multibillion-dollar supply chain to improve operational efficiencies, improve product quality, and ultimately enhance customer satisfaction,” Rainwater said in the release.

Standardizing the Eye

The case for computer vision here isn’t about replacing inspectors. It’s about giving them a consistent baseline. The AI applies the same visual criteria to every piece of fruit that moves through the system. Inspectors still approve the final rating.

The AI in food safety and quality control market was valued at $2.7 billion in 2024 and is expected to reach $13.7 billion by 2029, at a compound annual growth rate of 30.9%, according to BCC Research. Most deployments in that market run on manufacturing and processing lines. Albertsons built its tool in-house, positioned it at the distribution center level, and designed it around its own proprietary grading standards rather than a generic quality model.

That distinction matters for how the system performs. Generic visual inspection tools can identify obvious defects. A tool trained on a retailer’s internal standards grades against the same criteria the chain uses to make buying and markdown decisions, keeping the inspection layer consistent with the commercial layer.

Where the Data Goes Next

The system captures granular quality measures per inspection, building a data layer that didn’t exist when inspectors logged grades manually. That record can show how produce quality varies by supplier, origin or other variables. Consistent grading is what makes that analysis usable.

The Intelligent Quality Control tool is the latest product of Albertsons’ partnership with Google Cloud. In 2025, Albertsons was among the first grocers to launch Google Cloud’s Conversational Commerce agent as a customer-facing shopping assistant. The supply chain tool follows that deployment with AI applied to the distribution layer.

Albertsons plans to expand the system across more fresh products, the release said.

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

When engineers can move from 100 to 200 lines of code per day to thousands—a 10x leap enabled by AI tools—the effects don’t stop at the pull request.

Product design, testing, governance, budgeting and approval workflows were built for human-speed development. Those workflows are straining to operate at machine speed.

The traditional software development lifecycle had one foundational premise: shipping code takes time. Increasingly, it doesn’t. That speed gap is exposing a structural fault line between software deployment, which can now move at machine speed, and enterprise finance and budgeting for those same initiatives, which still moves at committee speed. CFOs are finding their financial planning and analysis (FP&A) frameworks were designed for a world where software shipped on quarterly cycles, not daily ones.

Read also: CFOs Turn to AI Harnesses as Agentic Capabilities Scale

Software Velocity Is Outpacing Corporate Planning

The rise of agentic AI changes the economics of software creation in two ways simultaneously. First, it lowers the cost of production. Second, it dramatically increases iteration speed. Historically, software projects required large upfront commitments because development cycles were long and labor-intensive. Finance teams could forecast expenses with relative confidence because milestones unfolded over quarters or years. A product roadmap resembled a capital project: linear, deliberate and heavily gated.

But teams today using agentic coding tools can now prototype multiple product directions simultaneously, abandon failing paths quickly, and scale promising ones almost instantly. The cost of experimentation drops sharply, but the volume of experimentation rises just as fast. A company that once funded five software initiatives per year may now launch fifty micro-projects in the same period. This creates a paradox for CFOs. AI-assisted development can improve efficiency while simultaneously increasing financial volatility.

Traditional FP&A systems are poorly designed for this environment because they optimize for stability and control, not rapid adaptation.

Unlike traditional software, agentic systems generate ongoing operational costs tied to inference, orchestration, model tuning and external API consumption. Those costs can shift dramatically within days depending on usage patterns and product adoption. A feature that unexpectedly gains traction may require instant infrastructure expansion. A new AI workflow may trigger substantial token consumption overnight.

The PYMNTS Intelligence report “Smart Spending: How AI Is Transforming Financial Decision Making” found that more than 8 in 10 CFOs at large companies are either already using AI or considering adopting it.

Read more: The Second Coming of Secondments? FDEs Hit the CFO Office

The Rise of the Adaptive CFO

Smart CFOs are increasingly positioning themselves as mediators between acceleration and control. Rather than resisting AI-driven velocity, they are investing in governance automation that can operate at comparable speed. They are also increasingly open to external help. PYMNTS covered recently how AI providers like OpenAI and Anthropic are reinventing Wall Street-style secondments through forward deployed engineer (FDEs), or company-employed AI specialists embedded inside client companies to customize systems, solve integration issues and speed deployment.

In practical terms, that means FP&A teams are becoming more embedded with product and engineering organizations. Finance is moving closer to the codebase because software velocity now directly influences capital allocation decisions.

The broader implication is that AI is not merely automating tasks. It is compressing corporate time horizons. The history of enterprise technology is filled with examples of operational bottlenecks migrating from one function to another. Manufacturing automation shifted constraints into logistics. Cloud computing shifted constraints into cybersecurity and governance. Agentic AI is now shifting constraints into organizational decision-making itself.

See also: Tech Giants Just Made Every Business Their Business

The companies that benefit most from AI-driven software acceleration will not necessarily be those with the best models or the largest engineering teams. They will be the ones capable of redesigning their internal operating systems fast enough to absorb the new pace of execution.

For CFOs, that means recognizing that finance infrastructure is no longer a back-office support function. It is part of the production environment. When software can evolve in days instead of months, budgeting cycles, approval frameworks and governance structures become strategic differentiators. Enterprises that continue operating with slow financial processes may discover that they are constraining the very productivity gains AI was supposed to unlock.

The question now is which artificial intelligence tools will cement their place as go-tos for consumers and their shopping and spending habits. A landmark study of Iowa corn farmers published decades before the emergence of ChatGPT, Claude and other AI models provides some clues.

The 1962 book-length study showed that on the eve of the Great Depression, a small number of farmers in two Iowa communities began scrapping old planting methods dating back to Native American practices in favor of recently invented hybrid strains. By 1941, after droughts had proved the hybrids resilient, nearly all the farmers had switched. Over roughly a decade, they showed an adoption pattern that followed the shape of the letter S, with slow initial uptake, then rapid acceleration as the majority climbed aboard, then a flattening of the upswing as the remaining holdouts joined in.

AI tools like Google Gemini aren’t popcorn. But they’re starting to show a pattern of acceptance similar to the one described by the influential sociologist Everett Rogers in his Iowa farmers study. Rogers argued that his “diffusion of innovations” model was a signature of how any new idea moves through a population. That’s why we’re talking here about a vegetable. With AI agents that carry out shopping orders autonomously estimated to handle 15% to 25% of all U.S. eCommerce purchases by 2030, according to JPMorganChase, how consumers adopt AI now and in the coming years has big stakes for the payments industry.

An April PYMNTS Intelligence report, “The AI On-Ramp: Data Shows How Everyday Tasks Build Consumer Habits,” suggested that consumer use of AI is following the same S-shaped curve as the Iowa corn farmers, crossing from early adopters (the small group trying hybrid seeds) into the majority. In Rogers’ framework, that transition is a highly consequential moment in a technology’s commercial life because it’s when the niche goes mass market and trials and experiments crystallize into lasting habits. It’s also when technologies that establish themselves as the default (hybrid corn seeds) for key tasks can gain structural advantages.

Seed of an Idea

Rogers argued that an innovation takes off with individuals when it has a clear advantage, is compatible with existing practices, is easy to understand, can be tried at low cost (financial and time- and asset-wise) and produces visible results. One of the leading scholars of how innovations take hold, he famously segmented his farmers into Innovators, Early Adopters, Early Majority, Late Majority and Laggards, which the PYMNTS report partially mirrored. Rogers found that with his farmers, the S-shaped curve took off when 10% to 25% of the population adopted a new technology as a result of “interpersonal networks” (“Hey! This seed works!”) becoming “activated.”

In that vein, the PYMNTS report found that AI has taken off with young consumers. Generation Z adults, the oldest now 29, are firmly in the majority of AI users. Roughly 70% of Gen Z adults are already using the technology for tasks like finding product links, writing resumes and editing personal writing, looking up medical symptoms and health information, learning new skills, crafting social media content, and sourcing financial guidance. Boomers and seniors are another matter. More than 1 in 3 American consumers aged 60 and older don’t use AI tools, consistent with the “laggard” profile in Rogers’ thesis about older farmers resisting hybrid seeds.

The PYMNTS report also revealed that AI tools become an “on-ramp” to wide acceptance when they have certain features. One of them is delivering value immediately, such as allowing a consumer to find relevant product links faster through, say, Google Gemini, than through a conventional Google search or Amazon. That captures Rogers’ theory of relative advantage, meaning the degree to which an innovation is perceived as better than what it replaces. And it mirrors his observation that severe droughts across the Midwest in 1934, 1936 and 1939-40 convinced farmers that drought-resistant hybrid seeds were better than open-field pollinated ones.

Another feature for an on-ramp posited by the PYMNTS report was the ability to try AI without losing too much if it makes a mistake. If ChatGPT shows, say, purple king-sized 300-thread count duvet covers when you want a lavender gray, queen-sized 600 count one, you’ve wasted a bit of time but nothing more.

The PYMNTS report also partially echoed Rogers on another front. It used data to argue that tasks that require no prerequisite knowledge and no specialist context were the ones that spread most broadly. That’s another way of saying they’re compatible with the widest range of existing habits and values. Rogers argued that hybrid corn seeds took off in part because they were “perceived as consistent with the existing values, past experiences, and needs of potential adopters. An idea that is not compatible with the prevalent values and norms of a social system will not be adopted as rapidly as an innovation that is compatible.” Similarly, the PYMNTS report said that the two highest-adoption tasks with generative AI—finding product links and editing personal writing—succeed because they require no demographic preconditions and thus are compatible with broad societal norms. Among AI users of all ages and income levels, 30% use AI models to look for product links and writing tasks.

A ChatGPT prompt that lists peer-reviewed medical studies on Lyme disease can be a lifesaver. An agent that autonomously books a vacation on your behalf according to your specific requirements can be a time- and wallet-saver. Still, beneficial innovations don’t sell themselves. As Rogers wrote in a later edition of his original study, the British Navy took nearly 50 years to adopt rations of oranges, lemons and other citrus fruits. (He blamed competing alternatives, a negative signal from a credible source in Captain Cook’s Pacific voyage reports, and the low status of the British researcher who demonstrated the solution.)

In other words, it’s not the actual quality of an innovation that drives its diffusion. It’s what people perceive, based on what their peers tell them and show them. An AI model can have all the bells and whistles, but as the PYMNTS data showed, consumers just want to find the right product and information.

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

However, instead of law firms sending junior associates over to their banking and institutional clients for six months, the artificial intelligence era is putting a new spin on how and why firms are embedding their own employees client-side.

This time, the professionals arriving on-site are not lawyers or management consultants. They are forward deployed engineers, or FDEs, technical specialists employed by AI vendors who embed directly within customer organizations to customize models, integrate enterprise systems, troubleshoot adoption barriers, and accelerate operational deployment.

As generative AI shifts from experimentation toward enterprise-scale implementation, the rise of FDEs reflects a growing recognition that buying AI software is the easy part. Making it work reliably inside complex organizations is something else entirely.

For chief financial officers, that can mean AI strategy is evolving from a procurement issue into an operational capability question.

See also: Tech Giants Just Made Every Business Their Business

The Human Layer Behind Enterprise AI

The rise of FDEs underscores a paradox emerging across the enterprise AI market. Despite the automation rhetoric surrounding AI, successful deployment remains human-intensive.

OpenAI announced Monday (May 11) that it plans to acquire applied AI consulting and engineering firm Tomoro, which will bring about 150 experienced FDEs and deployment specialists to OpenAI Deployment Company, a $4 billion partnership between OpenAI and 19 global investment firms, consultancies and system integrators. The company is designed to help the AI provider gain greater share in the winner-take-most enterprise AI market.

OpenAI rival Anthropic also announced last week the launch of its own new venture focused on selling AI tools to enterprise companies, in partnership with Goldman Sachs, investment firm Blackstone and private equity group Hellman & Friedman. The Anthropic initiative will help companies embed Anthropic’s Claude AI model into their businesses.

As FIS Head of Product Management, Payment Networks Mladen Vladic wrote in a new PYMNTS eBook, “AI Runs Payments. Governance Decides What Happens Next,” integration is key to ensuring effective AI governance.

Organizational readiness is the most cited barrier to AI adoption at large companies. More than 71% of executives at companies with at least $1 billion in yearly revenue named it as the chief limit on AI performance, according to research by PYMNTS Intelligence. Just 11% said the technology itself is the main obstacle.

After all, the underlying models are rapidly commoditizing. Multiple providers now offer highly capable generative AI systems. What increasingly separates winners from laggards is the ability to integrate those tools responsibly into repeatable workflows that employees actually use.

Read also: CFOs Turn to AI Harnesses as Agentic Capabilities Scale

Why Finance Teams Are Becoming AI’s First Enterprise Battleground

Historically, enterprise software transformations often began in marketing, customer service or IT. But AI’s early enterprise value proposition is landing squarely inside finance operations.

Accounts payable automation, procurement workflows, forecasting, treasury management, audit preparation, compliance reporting and FP&A analysis are all becoming prime candidates for AI augmentation. Finance organizations are rich in structured data, governed by repeatable processes, and under constant pressure to improve efficiency.

The PYMNTS Intelligence report “Smart Spending: How AI Is Transforming Financial Decision Making” found that more than 8 in 10 CFOs at large companies are either already using AI or considering adopting it.

Still, as AI pushes further into the back office, many CFOs are discovering that enterprise AI deployment is less analogous to installing SaaS software and more akin to undertaking a systems integration project layered atop organizational transformation. Technology vendors increasingly recognize this reality, which helps explain why some AI firms are staffing aggressively around deployment and customer engineering roles rather than relying exclusively on product-led adoption models.

The emergence of FDEs also reflects a broader shift in enterprise buying behavior. Companies no longer want merely a software license; they want operational outcomes. Vendors are being asked to share accountability for adoption, integration success and measurable ROI.

First, embedded engineers can shorten deployment timelines by addressing integration problems in real time. Second, they can improve employee adoption by tailoring AI systems to actual workflow needs rather than idealized product assumptions. Third, they may reduce governance risk by helping organizations establish controls around outputs, permissions and auditability from the outset.

Ultimately, FDEs may represent something larger than a staffing trend. They are evidence that enterprise AI still requires substantial institutional translation between software capability and operational execution.

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