Introduction

The recent U.S. Census Bureau data revealed an unexpected dip in AI tool usage among large enterprises. Specifically, the biweekly Business Trends and Outlook Survey (BTOS) found that the share of companies with over 250 employees using AI dropped from roughly 14% in mid-June to under 12% by late Augustitpro.com. This marks the sharpest decline since the survey began tracking AI adoption in 2023itpro.com. In a climate where AI has been hyped as transformative, this development has prompted questions about whether the initial “AI boom” is entering a cooling phase. In this report, we delve deeply into the data and its context, examine expert explanations for the slowdown, relate it to the broader AI hype cycle, compare it with alternative adoption metrics (from sources like Ramp and UBS), and assess the future outlook from enterprise, policy, and hype-cycle perspectives. All findings are backed by sources and visualized where appropriate.

U.S. Census BTOS Data on AI Adoption (June–August)

AI adoption rates by firm size (1–4 to 250+ employees) from Nov 2023 to Aug 2025. Large firms (red line) saw a peak in mid-2025 followed by a notable dropapolloacademy.comapolloacademy.com.

The Business Trends and Outlook Survey (BTOS) is a biweekly Census Bureau survey covering ~1.2 million U.S. employer businesses (all nonfarm sectors)census.govtomshardware.com. Starting in late 2022, BTOS added questions on AI, asking firms if they had used any artificial intelligence tools (e.g. machine learning, natural language processing, virtual agents, or voice recognition) to help produce goods or services in the last two weeksapolloacademy.com. This broad definition captures AI use in core business operations (production of goods/services), and responses are weighted to be representative of the entire U.S. economycensus.gov. The survey’s large sample and biweekly frequency make it a valuable barometer of AI adoption trends in near real-timecensus.gov.

Timeline of Adoption Rates: From late 2023 through mid-2025, overall business AI adoption was steadily rising before the recent dip. As of Oct–Nov 2023, only about 3.8–3.9% of U.S. businesses reported using AI in productioncensus.gov. By mid-2024 that share had climbed above 5%tomshardware.com, and continued upward into 2025 – reaching roughly 9%–10% by early summer 2025itpro.com. Large enterprises led the charge: companies with ≥250 employees were the most likely to use AI, with their adoption rate peaking just shy of 14% in June 2025itpro.com. However, in the subsequent biweekly surveys, large-firm adoption slipped back to ~12% by late August 2025, a notable pullback after consistent growthitpro.com. Other size categories saw mixed trends; many mid-size and small businesses had lower overall usage and some experienced plateaus or modest declines around the same periodtomshardware.comapolloacademy.com. The drop among large firms – the first significant decline since AI tracking began – suggests that even the most AI-forward companies hit a moment of hesitation in the summer of 2025.

Industry Breakdown: The BTOS data also show wide disparities by industry in AI adoption. Sectors like Information Technology and related services report the highest usage – well into the double digits – whereas many traditional industries remain in the low single digits. For example, in late 2023 the Information sector (software, data hosting, media, etc.) already had about 13.8% of businesses using AI, far above the national averagecensus.gov. By early 2024 this sector’s adoption climbed to roughly 18% (the highest of any sector) according to Census researchapnews.com. Professional, Scientific and Technical Services also showed above-average AI uptake (~9% of firms by late 2023)census.gov. In contrast, industries such as Construction, Agriculture, and Accommodation/Food Services lagged behind – for instance, only ~1–2% of businesses in construction or hospitality were using AI as of early 2024apnews.com. These gaps reflect that AI tools have so far been more easily adopted in data-intensive and tech-centric fields, whereas sectors with less digital infrastructure or narrower margins have seen slower uptake. Notably, firm size and industry often interact: larger firms (which are over-represented in tech/finance sectors) use AI more than small firms, though an interesting finding was that the very smallest businesses (micro-firms) slightly outpaced mid-sized firms in some adoption metricsapnews.com – perhaps because small startups are experimenting with AI, whereas some mid-tier firms are more cautious. Overall, the June–August 2025 period showed a slight downtick nearly across the board, but especially for large enterprises, even as total adoption (all firms) hovered around ~9–10% nationallyitpro.com. The Census Bureau cautions that the biweekly data can have some sampling variance and “fluctuations in statistical accuracy,” but the trend of a pullback in summer 2025 appears robusttomshardware.com.

Methodological Notes: Each biweekly BTOS sample is large, but individual firms only respond once every 12 weeks (the sample is split into rotating panels)census.gov. This means the “current use of AI” question captures different slices of businesses over time, and the results are smoothed by the Census (e.g. Apollo’s analysis applied a six-survey moving average)apolloacademy.com. The question wording – focusing on AI use in producing goods/services – might undercount companies using AI solely in back-office or non-production functions (more on this in Section 4)itpro.com. Still, the overall trend is clear: after a surge in late 2023 and early 2024 (when AI usage nationally jumped from ~3–4% to ~6% then ~9%tomshardware.comitpro.com), the trajectory leveled off in mid-2025. The dip among large firms in particular has raised eyebrows, given that this cohort had previously been steadily increasing its adoption and is often seen as a bellwether for enterprise technology trendsitpro.com. This data provides the foundation for our investigation – next, we explore why this slowdown might be happening.

Why the Slowdown? Expert Perspectives on AI Adoption Challenges

Experts and industry observers have proposed several, often overlapping, reasons for the recent cooling in AI uptake among big companies. These range from economic considerations to organizational and technical hurdles. We have grouped the commentary into five thematic categories:

1. Difficulty Proving ROI

One of the most cited reasons for caution is the challenge in demonstrating a clear return on investment (ROI) for AI initiatives. After an initial frenzy of AI pilot projects, many firms are finding that the promised efficiency gains or revenue boosts are not materializing – at least not quickly enough. A striking statistic from an MIT report in mid-2025 found that 95% of organizations reported essentially no measurable financial return from their AI investments so farmlq.ai. In other words, only a tiny 5% of AI pilots were “extracting millions in value” while the vast majority had “no measurable P&L impact,” essentially a failure to deliver tangible ROImlq.ai. Similarly, S&P Global data (cited by The Economist) noted that the share of companies abandoning most of their generative AI pilot projects rose to 42% in 2025, up from just 17% the year beforetechcrunch.com. This suggests a wave of disillusionment as early experiments failed to justify their costs. Nicole Kobie, writing for IT Pro, observed that some large enterprises are “growing frustrated at poor returns on investment” from AI, which is contributing to hesitationitpro.com. In fact, the Census data drop itself is interpreted as a sign that initial enthusiasm is giving way to ROI skepticism – companies that rushed to deploy AI are reevaluating if those deployments are actually yielding benefits.

Why is ROI proving elusive? Often AI projects incur significant up-front costs in data preparation, integration, training, and change management, while the gains (like labor savings or new revenue) are uncertain or slow. A biweekly MIT Sloan survey described an “AI paradox”: big firms led in launching AI pilots, but few pilots scaled to full production because they didn’t show quick winsmlq.aimlq.ai. Many executives likely overestimated AI’s short-term benefits. A Gartner analyst pointed out that in the real world, “proper due diligence” and extensive work are needed to make AI reliable and effective, rather than expecting it to “magically…everything happens” after feeding it datacio.com. When those magic results didn’t instantly appear – and in some cases AI even underperformed (e.g. producing errors or “hallucinations”) – business leaders pulled back, unwilling to invest further without a clear business case. Surveys indicate that fewer than 30% of AI leaders report their CEO is satisfied with the returns on AI spending so fargartner.com. This ROI pressure is leading to more scrutiny: projects now need to prove value or risk cancellation. In summary, the hype outpaced the economic reality, and the recent slowdown reflects a correction as companies demand that AI prove its worth before further rollouts.

2. High Integration Costs and Technical Complexity

Another barrier is the cost and complexity of integrating AI systems into existing operations. Many organizations underestimated how difficult and expensive it is to implement AI at scale. Initial pilot programs — often using off-the-shelf models or APIs — can be set up quickly, but scaling them company-wide or embedding them into core business workflows is much harder. Gartner’s Birgi Tamersoy noted that “AI systems inevitably make mistakes” and achieving robustness requires building a lot of surrounding infrastructure and safeguardscio.com. This includes data engineering, model monitoring, error handling, security measures, etc., which can be technically challenging and expensive. As generative AI models tackle more complex tasks, computing and energy costs skyrocket; running advanced AI can cost millions in cloud infrastructure and electricity, which directly impacts the bottom linecio.com. For example, companies that enthusiastically deployed large language models have reported eye-opening bills for cloud GPU usage. At some point, organizations must ask if the benefit outweighs these costscio.com – and some are concluding it doesn’t, leading them to scale back or optimize their AI usage.

Integration with legacy systems is another issue: enterprise IT stacks are often outdated or not designed to work seamlessly with AI tools. A 2025 analysis identified “Integration with legacy systems” as a top-7 challenge in enterprise AI adoption, noting that many companies’ existing software and databases are “incompatible with AI workflows”, requiring substantial investment in new infrastructure or middlewarestack-ai.com. Without this, AI projects remain siloed experiments. Data quality problems also surface – models perform poorly if fed fragmented or biased data, yet cleaning and unifying corporate data is a monumental task for large firmsstack-ai.com. All of these integration hurdles contribute to delays and disappointments. Indeed, Apollo’s chief economist Torsten Sløk commented that the Census data likely reflects companies “tapping the brakes” on AI after encountering these practical roadblocksitpro.com. In concrete terms, some firms that raced to implement AI found themselves needing to spend “huge sums on data centers” and IT upgrades to support these systemsitpro.com. If budgets blow out or timelines slip due to integration challenges, executives may pause further adoption until a clearer, cost-effective path emerges. Thus, high implementation costs and technical complexity are key factors tempering the initial rush to AI.

3. Organizational Culture and Change Resistance

The human element inside firms is another decisive factor. Introducing AI often requires significant changes in workflows and can generate employee resistance or even fear. Many organizations lack a culture that fully embraces fast-paced technological change, especially one as potentially disruptive as AI. Employees may worry about job security (e.g. “Will AI replace my role?”) or simply be reluctant to trust and use AI tools in their daily work. This can lead to low adoption even when the technology is available – e.g. a company might license an AI software, but employees under-utilize it or managers don’t mandate its use due to skepticism.

Analysts have indeed flagged “organizational resistance” as a top challenge: workers often “fear change, don’t adopt tools, or resist new AI-driven processes,” which can seriously hinder an enterprise’s AI effortsstack-ai.com. Ingrained processes and middle-management inertia can slow down the integration of AI into business decisions. Additionally, some companies lack “AI literacy” among staff and leadership, as Gartner observes – if people don’t fully understand AI capabilities, they can’t effectively implement or trust themgartner.com. Training employees and redesigning roles to work alongside AI is a non-trivial task that many firms are only beginning to tackle. The initial hype often glossed over the fact that AI adoption is as much a people challenge as a tech challenge. As reality sets in, organizations are recognizing that they need to invest in change management: communicating a vision for how AI will augment (not just replace) staff, providing upskilling opportunities, and creating an internal culture of experimentation rather than fear. Companies with a rigid or risk-averse culture likely put the brakes on AI projects when they encountered internal pushback or low uptake, contributing to the observed slowdown. Going forward, addressing cultural readiness is seen as key: a Thomson Reuters study found firms with a clear AI strategy (often implying better organizational alignment) were 3.5× more likely to realize benefits than those without onethomsonreuters.comthomsonreuters.com, highlighting how important the human factor is in successful adoption.

4. Regulatory Uncertainty and Risk

Regulatory and ethical uncertainties around AI are also casting a shadow and causing some firms to hesitate. In mid-2025, the policy landscape for AI is in flux: governments are scrambling to craft rules on data privacy, AI transparency, bias, intellectual property, and liability. This uncertainty can make companies nervous about deploying AI at scale, especially in heavily regulated industries (like finance, healthcare) or for use cases involving sensitive data. Business leaders don’t want to charge ahead with AI projects only to find out later that they run afoul of new regulations or expose the firm to legal risks.

For instance, concerns about data privacy and compliance are significant – AI often requires large datasets (which may include personal or customer data), raising questions under laws like GDPR in Europe or various U.S. privacy laws. A survey of AI adoption challenges noted that “privacy, security, and compliance” issues are a major hurdle, as AI systems can “raise risks around sensitive data and regulatory compliance”stack-ai.com. Until clear guidelines are established, many companies proceed cautiously to avoid reputational or legal damage from a misuse of AI. Additionally, high-profile controversies (e.g. AI algorithms exhibiting bias or making unethical decisions) have made executives aware of the ethical governance needed – and not every organization is prepared to implement AI ethics boards or rigorous oversight, so some choose to slow down deployment.

On the regulatory front, concrete frameworks are emerging but not yet fully settled. In the EU, the forthcoming AI Act is a comprehensive regulation that will impose requirements on AI systems (e.g. transparency, human oversight, and risk management for “high-risk” AI). Parts of it start taking effect in 2025, with enforcement in 2026–2027eversheds-sutherland.com. European businesses (and U.S. multinationals operating in Europe) are now preparing for compliance – for example, providers of general-purpose AI models will face new obligations as of August 2025, and by 2027 even AI systems already on the market must complyeversheds-sutherland.com. This creates a moving target for companies: some may delay certain AI deployments until they see the final regulations to ensure they don’t invest in a direction that becomes non-compliant. In the United States, regulatory action has been slower, but there are executive initiatives and agency guidelines being developed. Sector-specific rules (like FDA guidance on AI in medical devices, or FTC warnings about AI in consumer protection) add to the patchwork of considerations. Additionally, the legal environment around AI-generated content (IP ownership, copyright) and AI-related liability is still being tested. All of this uncertainty imposes a “wait-and-see” mindset for more risk-averse firms. Gartner’s 2025 analysis noted that “government regulations…may impede [some] GenAI applications” and that organizations are grappling with governance challenges like bias and transparencygartner.com. In sum, the current regulatory limbo makes AI a compliance risk – and many enterprises have hit pause until rules of the road are clearer.

5. Strategic Re-evaluation after the Hype

Finally, there is a broader strategic re-evaluation happening in boardrooms regarding AI. The past couple of years (2022–2023) saw what many call an “AI arms race,” where companies felt pressure not to miss out on the next big thing. This led to huge investments – IT Pro notes “trillions in AI capex” were announced or spent by large firms in a bid to ride the waveitpro.com. But now that the initial hype is cooling, some of those aggressive bets are being reconsidered. Arpit Gupta, a finance professor at NYU, remarked upon seeing the Census adoption downturn that “trillions in AI capex should probably be reconsidered”itpro.com. This encapsulates a shifting sentiment: instead of “AI at any cost,” executives are adopting a more measured approach, integrating AI into a longer-term strategy rather than as a quick win.

In practice, strategic re-evaluation means companies are pivoting from experimentation to consolidation. Many are taking stock of which AI pilot projects actually showed promise and which did not. Rather than continuing to scale pilots blindly, they may shelve a number of them (as evidenced by the 42% abandonment rate mentioned earlier) and double-down on the few that align with core strategy or demonstrated ROI. The era of “just try AI everywhere” is giving way to “let’s focus on where it truly adds value.” For example, a bank that experimented with dozens of AI use cases might determine that only fraud detection and customer chatbots yielded positive results, and thus channel investment to those areas while cutting others.

Additionally, macroeconomic conditions cannot be ignored – 2023–2024 were marked by rising interest rates and cost pressures, leading companies to prioritize cost-effective investments. If AI projects are not yielding near-term results, CFOs might curtail their budgets until a clearer business case is presented. This strategic pragmatism is part of normalizing the technology beyond the hype. In interviews, some tech leaders have admitted that the hype led to unrealistic expectations, and now there’s a need to “reset” and educate stakeholders on what AI can and cannot do in the near termcio.comcio.com. The slowdown in adoption can thus be seen as a strategic pause – companies taking a breath after the frenetic hype peak, refining their AI roadmaps, and ensuring that future AI deployments are tightly aligned with business goals and come with proper change management. This more cautious, strategy-driven approach contrasts with the earlier FOMO-driven rush, and is likely healthier in the long run, but it does show up in the data as a plateau or dip in adoption as the frenzy cools.

The AI Hype Cycle: From Peak of Inflated Expectations to the Trough of Disillusionment

The trajectory of AI adoption in enterprises right now closely maps to what Gartner’s famous Hype Cycle model would predict. Over the last two years, AI – especially Generative AI – shot up to what Gartner calls the “Peak of Inflated Expectations,” and is now sliding down into the “Trough of Disillusionment.” This conceptual framework helps contextualize the recent pullback in adoption: it’s a classic case of a technology moving past the hype peak and into a phase of realism (and sometimes pessimism) before genuine productivity gains are eventually realized.

Gartner’s 2025 Hype Cycle for Artificial Intelligence explicitly places Generative AI on the descent. According to Gartner analysts, “Gen AI enters the Trough of Disillusionment” in 2025 as organizations develop a more sober understanding of its capabilities and limitsgartner.com. In 2023, generative AI (think ChatGPT and similar tools) was at the absolute peak of hype – touted as game-changing for every industry, attracting massive investments, and yielding plenty of media buzz and pilot projects. By mid-2024 and into 2025, however, the inflated expectations began to deflate. The reasons mirror what we discussed earlier: inconsistent results (e.g. AI chatbots that sometimes err or “hallucinate”), lack of reliability for critical tasks, and the significant work needed to maintain and refine these systems. Gartner notes that many organizations “have run into problems with [GenAI’s] robustness and reliability” and discovered that the “hype…downplayed much of the work needed to reap its benefits.”cio.comcio.com. In essence, the shiny promise met messy reality – leading to disappointment.

We are “exactly at that inflection point” on the curve. Gartner’s expert Birgi Tamersoy says the excitement in the enterprise has “passed its peak” and now we need better use cases and more accurate results to “renew the enthusiasm”cio.com. Other analysts have started invoking the term “AI winter” – referring to past periods where AI development slowed after hype cycles. For instance, by late 2025 some financial analysts warned of a potential new AI winter if investment enthusiasm wanes; Fortune magazine noted these warnings as AI stock prices wobbled, and Tom’s Hardware highlighted that narrative in coverage of AI’s boom-to-chilltomshardware.com. While an “AI winter” (a prolonged trough) isn’t a certainty, the language shows a clear sentiment shift from unbridled optimism to caution and even pessimism in some quarters.

However, the Hype Cycle also suggests what comes next: after the trough of disillusionment, the technology (if fundamentally value-adding) enters a “Slope of Enlightenment” and eventually a “Plateau of Productivity.” Gartner predicts that generative AI will “take 2 to 5 years to climb out of the trough” and reach a stable, productive stagecio.com. In practical terms, this means we can expect the current slowdown to persist in the near term – with continued skepticism and slower growth in adoption – but over the next few years, as best practices are learned and the tech matures, adoption could pick up again in a more sustainable way. Importantly, Gartner identifies trends that will aid this maturation: for example, AI engineering and ModelOps (model operations) are being emphasized as foundational practices to reliably and scalably deploy AIgartner.comgartner.com. These are the kinds of “boring” but necessary capabilities that signal moving past hype into real productivity. Likewise, emerging solutions to AI’s current limitations (like “composite AI” that combines multiple techniques to overcome single-model weaknessescio.comcio.com) show that the industry is actively working through the trough.

In short, the AI hype cycle narrative aligns well with what we observe: 2023 saw the “Peak of Inflated Expectations” (everyone jumping on AI with perhaps unrealistic hopes), and now in 2025 we’re in the “Trough of Disillusionment” (some firms pulling back, citing difficulties and unmet promises). The Census Bureau data drop among large firms is a quantitative symptom of this qualitative cycle. The saving grace is that beyond the trough, a more measured and realistic growth in AI adoption is likely. As one expert put it, “we have definitely not overestimated the medium- and long-term implications of LLMs [large language models]” even if we “overestimated [their] potential in the near term”*cio.com. The task now is to climb out of the trough by solving the real challenges – something that will be discussed in the next sections on alternative data signals and future outlooks.

Contrasting Data: Ramp and UBS – Is AI Adoption Still Booming or Leveling Off?

While the Census BTOS offers one view of AI uptake (based on self-reported usage in production), other data sources paint a nuanced picture. Notably, financial and investment analyses from firms like Ramp and UBS suggest that AI adoption, in some respects, continues to grow robustly – especially when measured by spending or across particular sectors. These perspectives provide a counterpoint to the narrative of a broad slowdown, indicating that the reality may differ by data source, metric, or segment of the market.

Ramp’s AI Adoption Index: Ramp, a fintech company, uses anonymized corporate credit card and bill payment data from 40,000+ businesses to track how many firms are paying for AI products or services. This Ramp AI Index captures actual spending on AI tools (e.g. subscriptions to AI software, API usage charges, etc.), which is an indirect but concrete measure of adoption. According to TechCrunch reporting on Ramp’s data, the proportion of businesses using (i.e. spending on) AI reached about 41% in May 2025techcrunch.com. Strikingly, Ramp’s methodology found much higher adoption rates than the Census survey – e.g., 49% of large businesses had some AI spend, 44% of mid-sized, and 37% of small companies by Maytechcrunch.com. (This is versus Census figures in single digits overall, because Ramp counts any AI-related usage, not just in product production. Also Ramp likely captures many instances of AI used in support functions or via third-party services that a survey respondent might not think to report.) Ramp’s data thus suggests AI penetration is already quite widespread – over a third of U.S. businesses were paying for AI tools by early 2025ramp.com. Moreover, spending was surging: Ramp noted AI-related expenditures among its customers grew nearly 4× year-on-year, and that “over a third of Ramp customers now pay for at least one AI tool, compared to 21% one year ago”ramp.com. This indicates rapid growth through 2024 and early 2025.

However, even Ramp saw signs of leveling off by mid-2025. After ~10 months of steady increase, their index plateaued at ~41% in Maytechcrunch.com. In other words, the share of companies adopting AI (by this measure) stopped growing for the first time in almost a year. This parallels the Census finding of a summer plateau/decline. Ramp’s economist Ara Kharazian suggested that businesses might be hitting a natural adoption ceiling in the short term, or at least a pause as companies digest their AI investments. The alignment is notable: both a direct usage survey (Census) and a spending-based measure (Ramp) indicate that late spring/summer 2025 was an inflection point where explosive growth gave way to a more hesitant trend. Ramp’s data also highlights a potential reason for the pause – firms realizing “there’s a limit to what today’s AI can do” and encountering failures. They cite, for example, the fintech company Klarna’s experience: it attempted to replace support agents with AI but had to backtrack and rehire staff due to “lower quality” customer service outcomestechcrunch.com. Many companies likely experienced similar setbacks. Additionally, 42% of companies in an S&P Global survey said they have now shelved most of their GenAI pilots (as noted earlier)techcrunch.com. These real-world checks could explain why, despite more businesses than ever having tried AI (per Ramp’s high adoption percentage), the growth rate of new adopters is slowing.

UBS’s Analysis: Meanwhile, UBS’s Wealth Management research team presented an optimistic macro view in June 2025. In a report titled “Overall AI adoption is far from a peak,” UBS analysts interpreted the Census data up to Q2 2025 quite differentlyubs.com. They noted that “the latest survey showed another step-up” in usage: national AI adoption rose to 9.2% of firms in Q2 2025, up from 7.4% in Q1 and 5.7% in Q4 2024ubs.com. From UBS’s perspective, this was evidence of a robust secular trend, with AI diffusion accelerating. They even compared it to the trajectory of e-commerce, noting that “crossing the 10% threshold” (which was imminent for AI) took e-commerce 24 years to achieve, whereas AI might do it in a couple of yearsubs.comubs.com. UBS argued that “overall AI adoption is far from a peak”, pointing out that some industries already report 25–30% adoption with tangible use casesubs.com. For example, they cited real-world success stories: tech companies using AI coding assistants to save hundreds of millions of dollars, or heavy AI usage in customer service (PayPal automating ~80% of support interactions with AI)ubs.com. These anecdotes illustrate that in certain leading firms and sectors, AI implementation continues to deepen and yield benefits.

How do we reconcile UBS’s bullish view with the recent slowdown data? One key is timing and scope: UBS published their take in June 2025, using quarterly aggregated data (up to Q2) – which indeed showed strong growth from 2024 to early 2025ubs.com. The more granular biweekly data picked up the softening in late Q2 into Q3 (June–August) that a quarterly view might gloss over. It’s possible that from UBS’s vantage in June, the broader trend still looked clearly upward (and they projected it forward confidently), whereas by August the picture had changed modestly. Additionally, UBS and Ramp highlight that adoption is not uniform – even if large firms paused, smaller firms or specific industries might still be increasing their use. The Census data itself showed overall usage tick up from 8.8% to 9.7% around Augustitpro.com, implying that while large firms’ share dipped, some smaller ones may have continued to adopt, keeping the national number roughly flat-to-slightly-up. UBS’s point about some industries being as high as 30% indicates pockets of very high adoption (information/tech, finance, etc.) which could continue to rise even if the average across all businesses stagnates temporarily.

There’s also a methodological difference: surveys like Census may undercount casual or indirect AI use, whereas spending data might overcount instances (e.g. a firm that spends $50 on an AI SaaS tool once is marked as an “AI adopter” by Ramp). Ramp itself notes its metric might miss free AI usage but also might count companies that are merely experimentingramp.comramp.com. The truth likely lies between these measures. What’s clear is that large enterprises (250+ employees) are the vanguard – they reached the highest adoption levels first and are the first to show a plateau. Smaller firms lag in percentage adopting, but that also means many have room to grow and could continue to increase adoption even as large firms regroup. In fact, Bloomberg reported in mid-2025 that “AI usage is spreading among small businesses” even as big firms temper expectationstomshardware.com (Tom’s Hardware cited a June survey of 1,500 small businesses that saw AI usage dipping among them too, but other data suggests small firms are still catching up).

Convergence or Contradiction? In summary, these alternative data points both converge with and diverge from the Census trend. They converge in suggesting that summer 2025 saw a leveling off after a period of rapid growth. They diverge in magnitude and framing: Ramp shows a much higher absolute adoption rate (41% vs ~10%) by counting any AI spend – implying many firms have at least dabbled in AI – whereas Census’s stricter usage definition shows a lower core adoption. UBS remains optimistic that we haven’t come close to a saturation point for AI in business, highlighting ongoing growth stories and forecasting continued exponential investment (they estimate global AI capital expenditure will grow another 33% in 2026, reaching $480 billion after a 60% jump in 2025)ubs.com. The Census data doesn’t contradict that long-term growth – it may simply reflect a short-term pause. In fact, one could interpret the pause as a healthy breather; adoption curves are rarely straight lines upward. Even during the rise of cloud computing or e-commerce, there were moments of plateau before re-acceleration.

What all sources agree on is that AI adoption is highest in large companies and certain tech-forward sectors, and that we are still in early days relative to AI’s potential. The disillusionment phase might be a temporary hurdle on an otherwise upward trajectory, as UBS suggests. The next section will look ahead to how enterprises and policymakers are responding, and where AI adoption might go from here, balancing the cautious signals with the continuing momentum in the background.

Future Outlook and Next Steps

Despite the recent slowdown in adoption growth, AI is far from yesterday’s news. The question on everyone’s mind: What comes next? In this section, we assess the future from three angles – how enterprises plan to proceed with AI investments, how policy and regulation might shape the playing field, and where we stand on the hype cycle (and what that implies for the near future).

Enterprise Strategies Going Forward

For enterprises, the brief chill in AI enthusiasm does not mean they are giving up on AI – rather, many are recalibrating their AI investment strategies to be more strategic, value-driven, and sustainable. We can expect a few key trends in how companies approach AI in the coming years:

• From Experimentation to Execution: Companies will shift from the frenzy of running dozens of pilots to focusing on a handful of high-impact, feasible use cases. The “fail-fast” experimentation ethos is giving way to a “scale what works” mentality. In practical terms, firms will allocate budget to areas where AI has proven its ROI or clear competitive advantage (e.g. automating routine customer inquiries, predictive maintenance in manufacturing, AI-assisted coding in software firms) and pull back from more speculative projects. This focused approach ties into having a coherent AI strategy. As noted earlier, only ~22% of organizations had a defined AI strategy in 2025thomsonreuters.com – but those that do are seeing better results. We anticipate more companies will develop formal AI roadmaps and governance structures to guide investments, rather than the ad-hoc adoption of the past. AI will be treated as one tool in the digital transformation toolkit, used where it aligns with business priorities, rather than as an end in itself.
• Investing in Foundations (Data, Talent, Processes): To get AI out of the trough of disillusionment, enterprises are recognizing the need to invest in the foundational enablers of AI. This includes improving data quality and accessibility (since AI is only as good as the data fed into it), building the right data infrastructure (cloud, data lakes, etc.), and upskilling the workforce or hiring AI talent. Gartner emphasizes “AI engineering” and “ModelOps” as crucial practices going forward – meaning companies will put effort into the engineering discipline of reliably deploying and managing AI models at scalegartner.comgartner.com. We’ll likely see the rise of internal AI platforms, better tooling for model monitoring and versioning, and integration of AI systems with core IT systems. Essentially, enterprises will move from the ad-hoc pilot stage to embedding AI in their operational backbone (but to do so, they must lay a lot of groundwork). This period may involve heavy lifting with less glamorous work – cleaning data, establishing governance committees, defining ethical guidelines, etc. – but it’s necessary for long-term success.
• ROI and Productivity Focus: Economic pressures will ensure that AI projects are judged by strict performance metrics. After learning hard lessons on ROI, companies will set clearer KPIs for AI deployments (e.g. reduction in processing time by X%, increase in sales conversions by Y%, etc.). There will be an emphasis on augmenting employees rather than outright replacing them in most cases – i.e. using AI to boost productivity. Many firms are optimistic that such productivity gains are real: a Thomson Reuters survey of professionals estimated AI could save ~5 hours per week per employee within a year, up from 4 hours a week predicted earlierthomsonreuters.comthomsonreuters.com. Those hours translate into economic value (they estimated ~$19k annual value per employee on average if AI is utilized fullythomsonreuters.com). Enterprise AI strategies will thus focus on integrating AI into workflows to capture these efficiency gains (for example, using AI to draft documents or summarize data, thereby freeing employee time for higher-value tasks). The goal is to harvest low-hanging fruit – where can AI quickly reduce costs or improve outputs? We already see a pivot: whereas 2023 was about grand visions (e.g. “transform our entire customer experience with AI!”), 2025–2026 will be about iterative improvements (e.g. “use AI to automate data entry in our finance department to cut costs by 10%”).
• Continued Investment, but Disciplined: On the financial side, we will still see significant investment in AI, but likely more disciplined. Big Tech companies (Google, Microsoft, Amazon, etc.) and AI firms will continue pouring money into R&D (indeed, their AI capex is enormous and still increasing). For other enterprises, the investment will continue but might shift form – for instance, rather than building everything in-house, many will opt for AI-as-a-service and cloud AI platforms to reduce complexity. This means partnering with vendors who provide pre-trained models or industry-specific AI solutions. Such partnerships can accelerate adoption for companies that lack deep in-house AI expertise. The flip side is vendor risk: businesses will have to vet AI providers for reliability and compliance. In any case, the spending trajectory is still up: recall that UBS projects global AI capital spending to grow another 33% in 2026ubs.com. Enterprises are likely to spend, but the spending will target enabling technologies (data platforms, cloud services, security) and proven applications, rather than moonshot experiments.
• Competitive Dynamics: Finally, enterprises will consider competitive pressure – if AI truly delivers productivity gains, no firm will want to be left behind. The initial hype may have led some companies to adopt AI because “everyone is doing it”; the next wave might see adoption because specific competitors are gaining an edge with AI. For example, if one bank’s AI fraud detection drastically lowers fraud losses, other banks will race to match that. This competitive adoption could drive a second wind of AI uptake once the effective use cases are clearer. In Thomson Reuters’ 2025 report, 80% of professionals believed AI would have a high or transformational impact on their industry in 5 yearsthomsonreuters.com – the awareness of AI’s long-term potential remains high, even if near-term expectations dipped. So enterprises that emerge from the trough successfully will likely inspire others, kicking off new adoption cycles.

In summary, enterprise AI investment strategies post-hype will be characterized by a pragmatic, value-oriented approach. The wild west days are ending; a more mature phase is beginning where AI is one tool among many – important, yes, but needing to prove itself within a broader digital strategy. Companies that navigate this well (with strong data foundations, clear ROI metrics, and workforce buy-in) are poised to lead in the eventual “plateau of productivity” phase.

Policy and Regulation: Shaping the Playing Field

The policy environment for AI is evolving rapidly, and it will have a significant influence on how adoption progresses. Governments and regulators around the world are introducing frameworks to ensure AI is used responsibly, safely, and ethically. For businesses, these frameworks are double-edged: in some ways they add compliance burdens (possibly slowing adoption or increasing costs), but they can also create clearer rules that reduce uncertainty and enable wider adoption under well-understood guardrails.

Some key developments and expectations on the policy front include:

• EU AI Act and Global Standards: The European Union’s AI Act is one of the most comprehensive regulatory efforts. It will impose strict requirements on AI systems, especially those deemed “high-risk” (like AI in healthcare, finance, employment decisions, etc.). Providers of AI will need to conduct risk assessments, ensure transparency (users should know when they’re interacting with AI), and implement oversight for critical applications. The Act is expected to be finalized in 2025, with phased implementation: by mid-2025 voluntary codes and some obligations for foundation model providers took effecteversheds-sutherland.comeversheds-sutherland.com, and full enforcement for high-risk use cases will likely kick in 2026. As noted, businesses deploying general-purpose AI in the EU must start complying with new obligations from Aug 2025 (with an adaptation period until 2026–27)eversheds-sutherland.com. This means companies need to invest in compliance – e.g. keeping documentation of their AI models, implementing monitoring to prevent misuse or bias, etc. European companies (and multinationals operating there) may have slowed AI rollouts in anticipation of these rules, waiting to ensure compliance. On the other hand, once rules are set, companies might accelerate adoption knowing what the “rules of the road” are. The EU is also promoting standards like the General Purpose AI Code of Practice (a voluntary set of guidelines for AI developers to align with the Act)eversheds-sutherland.com. Such standards, if widely adopted, could make it easier for companies to trust third-party AI systems that certify compliance – potentially boosting adoption in regulated areas by providing trust and accountability.
• U.S. Policy and Self-Regulation: In the U.S., there isn’t yet an AI-specific federal law akin to the EU’s, but there are important moves. The Biden Administration has issued directives and facilitated commitments from AI companies on issues like safety testing, watermarking AI-generated content, and sharing best practices. Agencies like NIST have released an AI Risk Management Framework (a voluntary framework to help companies manage AI risks regarding bias, transparency, security, etc.). We might see more concrete sectoral guidelines (for example, the FDA has been updating guidance on AI in medical devices, the CFPB looks at AI in lending, etc.). Additionally, discussions about intellectual property (who owns AI-generated content?) and data rights (using scraped data for training AI) are ongoing. For companies, this patchwork means staying agile – many are creating internal AI governance committees to preempt regulatory issues by setting their own policies. For instance, some organizations now have internal rules on which AI tools employees can use with sensitive dataeversheds-sutherland.comeversheds-sutherland.com. We expect more self-regulation in the short term: companies voluntarily implementing ethical AI principles, bias audits, etc., both to prepare for future laws and to build customer trust. This proactive stance can actually encourage adoption – if a company is confident in its ethical guardrails, it may deploy AI more broadly without fear of scandal.
• Privacy and Data Protection: Globally, data privacy laws (GDPR in EU, CCPA/CPRA in California, etc.) indirectly impact AI, since AI often needs lots of data. Enforcement of these laws is becoming stricter. The use of personal data in AI (like training customer service AI on chat logs) must respect consent and purpose limitations. Companies are exploring Privacy-Enhancing Technologies (PETs) – like federated learning or differential privacy – to use data for AI without violating privacy. Singapore, for example, has launched a PETs guide and sandbox to help firms adopt AI while preserving privacyeversheds-sutherland.com. We will likely see regulators pushing such approaches. For adoption, this means firms that manage data well will have an easier time scaling AI; those that don’t will hit legal roadblocks. Also, regulators could mandate transparency to users (e.g. disclosing AI usage in decisions) – companies will incorporate that in design.
• AI Accountability and Liability: A big open question is who is liable when AI goes wrong (e.g. flawed decisions, accidents with autonomous systems, etc.). Regulations may start addressing this (the EU Act does to some extent). Clarity here will affect adoption – for example, if laws shield companies from certain liabilities when using certified AI tools, companies might be more willing to use them. Conversely, if using AI exposes a company to new liabilities, they might be cautious. We’re starting to see movement: the EU is considering adjustments to product liability law to cover AI, and some jurisdictions talk of an AI insurance or licensing regime for high-stakes AI. The balance of risk will factor into CIOs’ decisions on deploying AI widely.
• Geopolitical and Cross-border Issues: On an international scale, differing regulations could complicate adoption for global companies. They may have to tailor AI systems or even hold back features in certain markets. Over time, there might be efforts for harmonization – e.g. the EU’s push for a global “Convention on AI” to align principles across countrieseversheds-sutherland.com. If successful, that could simplify compliance and encourage more universal adoption of best practices.

In essence, policy is moving towards more oversight of AI, but also towards providing frameworks that make AI usage more sustainable. In the near term (the next 1–2 years), regulatory uncertainty could still slow some projects (as noted earlier, uncertainty causes hesitation). But within, say, 3–5 years, we should have clearer rules in major economies. Paradoxically, clear regulation can be an enabler because it builds public trust and sets a level playing field. For example, if regulations ensure AI in hiring is audited for bias, companies might adopt it more once they know how to do so without discrimination. So, we might see a dip during the transition to regulated AI, then a surge as compliant, trustworthy AI becomes the norm. Businesses that are proactive – engaging with regulatory sandboxes, adopting voluntary codes, etc. – will likely navigate this period best and could even help shape policies to be innovation-friendly.

Hype Cycle Positioning: Where Are We and What’s Next?

As discussed, it appears we are currently in the “Trough of Disillusionment” for enterprise AI, particularly for generative AI. Understanding this positioning helps set expectations for what comes next on the hype cycle curve:

• Current Position – Trough of Disillusionment: At this stage, the inflated hype has died down. Many organizations are disappointed with early results, the media narrative has become more balanced (covering failures as well as successes), and the technology is no longer viewed as a silver bullet for everything. This is exactly where AI finds itself in late 2025. The Census data decline among large firms and commentary about AI projects being abandonedtechcrunch.com exemplify this disillusionment. Gartner’s assessment that gen AI “reached its peak of inflated expectations last year” and is now firmly in the trough confirms this positioningcio.com. Importantly, being in the trough doesn’t mean AI isn’t useful – it means the expectations vs. reality gap is at its widest, and negative press or sentiment is common. Indeed, we’ve seen rising concerns (hallucinations, lack of trust in AI outputs, etc.) and a retreat by some early adopters, which is classic trough behavior.
• Duration of the Trough: Gartner estimates it may take 2 to 5 years for generative AI to emerge from the trough and climb the “Slope of Enlightenment” toward the “Plateau of Productivity”cio.com. This suggests that through 2026 and 2027, we will gradually see improvements and lessons learned that restore confidence. In this time frame, there might not be another explosive hype spike, but rather a slow build of adoption as the technology improves and finds its true niche. It’s worth noting that different AI sub-technologies can be at different cycle phases – for example, older, more established AI techniques (like classic machine learning for supply chain optimization) might already be on the Plateau of Productivity in some industries, whereas bleeding-edge tech like autonomous AI agents are just now at the Peak and will hit a trough latercio.com. But focusing on the current star (gen AI), we’re likely near the bottom of the trough now. Sentiment: We should expect continued skeptical headlines (“AI wasn’t all it was cracked up to be,” etc.), companies being quieter about AI in earnings calls unless they have solid numbers to share, and investors scrutinizing AI startups more closely on fundamentals (as opposed to just concept). The market already reflected this: AI-heavy stock valuations saw some cooling in late 2025, and there’s talk of an “AI bubble” normalizingitpro.com.
• Climbing Out – Slope of Enlightenment: As the industry digests the lessons of early failures, the Slope of Enlightenment phase will see practical, narrower applications of AI gaining traction. We’ll start hearing success stories of AI delivering consistent value in specific contexts. Knowledge about best practices will spread – for instance, case studies of how Company X solved the reliability issue and achieved a 20% cost reduction using a combination of AI + process change. Those kinds of stories will rebuild enthusiasm, but in a tempered way. New generations of AI models (perhaps more efficient, less data-hungry, more transparent models) will emerge from R&D, addressing some current pain points. For example, there’s heavy research into reducing hallucinations and making AI outputs more explainable; any breakthroughs there will boost confidence. Timeline wise, perhaps by 2026 we’ll see early signs of this enlightenment phase: some companies that quietly kept investing will demonstrate notable gains, and the narrative will shift to “here’s how to do AI right.” This may correspond with broader adoption of the enabling technologies we mentioned (AI engineering pipelines, composite AI, etc. – which themselves are rising on their own hype cyclesgartner.com). So essentially, expect a gradual upturn where the conversation moves from “AI failed to deliver” to “AI is delivering in these areas if done correctly.”
• Plateau of Productivity: Looking further out, once AI is truly robust, well-understood, and widely deployed, it enters the Plateau of Productivity – meaning it becomes a normal, even mundane, part of business operations that consistently yields benefits. For perspective, technologies like cloud computing or mobile payments are in such a plateau now – everyone uses them and they provide clear value, even if they no longer make sensational headlines. For AI, reaching this plateau could mean, say, by late 2020s, using AI in business is as routine as using databases. What’s Next (Plateau): in that world, many AI tools will be off-the-shelf, integrated into software suites, with most employees having an AI assistant at their disposal (and trusting it). Productivity boosts accumulate and are reflected in macroeconomic productivity stats (one thing skeptics and economists are watching for). Gartner seems to hint that many AI techniques will reach plateau in a matter of a few years to a decade depending on the tech. For generative AI specifically, we might guess the late 2020s as the plateau period – once issues are ironed out. At that point, adoption rates could again accelerate in late majority firms (the ones who wait for tech to prove itself). So paradoxically, the broad base adoption might actually occur during the plateau, not at the peak of hype. We’ll know we’re at the plateau when companies talk about AI as a given in operations (like “of course we use AI in our supply chain, it’s standard”) rather than a novelty.

In summary, we are at the trough now – a necessary phase where hype meets reality. The next stage should be incremental improvements and rebuilding of trust in AI as kinks are worked out. Stakeholders should not misinterpret the trough as the “end” of AI; rather, it’s a transition. As one expert noted, we did “overestimate AI’s near-term” impact but have “not overestimated its long-term implications”cio.com. That captures the outlook well: the long-term trajectory (over the next decade) still points to AI being transformative across industries – but via a series of smaller, realistic advances, not overnight revolutions. Companies and policymakers that understand this will be well positioned to take advantage of AI’s “Slope of Enlightenment” and beyond, rather than becoming too discouraged during the trough. In practical terms, the advice is to keep iterating and learning from the current deployments, because the knowledge gained now will pay dividends when the technology enters its productive phase.

Conclusion

The recent dip in AI adoption among large U.S. companies – from ~14% to ~12% usage over the summer of 2025 – serves as a valuable reality check. Far from a sign that AI is a failed trend, it signifies a maturation process. The initial exuberance of the AI boom has been tempered by practical challenges: proving ROI, integrating complex systems, overcoming cultural and regulatory hurdles, and aligning AI projects with real business needs. In many ways, this was an expected transition along the innovation journey, analogous to other transformative technologies that experienced hype cycles.

Our deep dive into the data and commentary revealed a consensus that AI’s promise remains immense, but capturing that promise requires hard work and patience. The Census Bureau’s BTOS data and other metrics indicate that adoption is still growing in aggregate (especially compared to a year or two ago), but organizations are becoming more selective in how they implement AI. The “low-hanging fruit” of inflated expectations has been picked (or pruned), and now a period of cultivation is necessary for AI to bear sustainable fruit. Leading firms are taking this time to build the infrastructure, governance, and skills needed to make AI a lasting part of their operations. At the same time, some firms on the sidelines may use this lull to learn from early adopters’ mistakes and then leapfrog ahead when the technology stabilizes – so competitive dynamics will remain in play.

On the broader stage, we situated this moment within the Gartner Hype Cycle, identifying it as the trough of disillusionment for generative AI. The encouraging insight from that model is that after the trough comes enlightenment and productivity. Indeed, the narrative is already subtly shifting from “AI can do anything right now” to “AI can do specific things really well if you do it right.” That is progress. History suggests that technologies which survive the trough and reach plateau end up integral to the economy. All signs point to AI being such a technology. The current slowdown might even be considered a healthy correction, weeding out frivolous applications and forcing focus on viable use cases and responsible practices.

In the coming years, we expect to see a convergence of factors driving a second wave of AI adoption – improved technology (more reliable models, better tools), clearer regulations (providing guardrails and confidence), and refined business strategies (centered on ROI and integration). The “hype” will likely not return in the same frenzied form, but a steadier, enduring growth of AI usage will take its place. As enterprises emerge from this reassessment period, those that have invested wisely in foundational capabilities will start reaping noticeable benefits, spurring others to follow suit in a more evidence-based manner. Policymakers, by then, will hopefully have laid down rules that protect society from AI’s risks without unduly hampering innovation, striking a balance that encourages adoption of trusted AI solutions.

To answer the fundamental question: Has the AI boom gone bust or just hit a speed bump? – the research leans strongly toward the latter. The boom is evolving, not evaporating. The slight decline in adoption rates among big firms is a pause for breath, not a full stop. It’s a chance for organizations to regroup and ensure that when they accelerate again, it will be with clearer vision and stronger execution. In the meantime, AI continues to advance behind the scenes (in R&D labs, in niche deployments delivering value). When the fruits of these advancements are ready and the business world has digested its early lessons, we will likely see AI adoption curve upward again – perhaps not as explosively as before, but more resolutely and pervasively.

In conclusion, we stand at an inflection point where the AI narrative is shifting from hype to pragmatism. The coming phase promises to be about “making AI work” – integrating it wisely, governing it responsibly, and measuring its impact rigorously. Companies that persist through the current headwinds by doing these things will be the pioneers of AI’s plateau of productivity. In a few years, we may look back at 2025’s dip as a mere blip on the radar – a necessary course correction on AI’s long journey to transforming business and society. For now, caution and optimism coexist: AI’s wild ride is entering a new, more mature chapter, and everyone – enterprises, regulators, and workers alike – has a role in shaping what that chapter looks like.

Sources: The analysis above draws on data from the U.S. Census Bureau’s BTOS surveysitpro.comitpro.com and official Census publicationscensus.govcensus.gov, expert commentary from Apollo Global Managementapolloacademy.com, ITProitpro.com, TechCrunchtechcrunch.comtechcrunch.com, and others, industry research (MIT, Gartner)mlq.aigartner.com, and alternative metrics from Ramp and UBS reportstechcrunch.comubs.com, as cited throughout the text. These sources provide a multifaceted, credible foundation for the insights and claims made in this report.