Large companies may be losing patience with AI pilots, but the available evidence does not show a broad retreat from AI. It points to a more selective phase: organizations continue to experiment and invest, while many struggle to move projects into production, redesign workflows, and prove financial returns.
What the reported decline does—and does not—show
An analysis published by ITPro on September 9, 2025, cited a US Census-based measure showing that the share of businesses with more than 250 employees reporting AI use fell from just under 14% to about 12% during summer 2025. The measure asked whether a business had used AI to produce goods or services in the previous two weeks. That short-term decline is a signal worth watching, not proof that large enterprises have stopped adopting AI. ITPro’s account of the Census measure also reported that overall business AI use rose from 6.3% at the end of 2024 to 9.7% in the latest survey it cited.
The measure captures recent use under a specific definition. It does not directly measure strategic plans, spending, paid deployments, production systems, or whether a company is earning revenue or cutting costs from AI. Its “more than 250 employees” group is also not synonymous with the world’s largest multinational corporations.
Short survey windows can fluctuate, and respondents may interpret “AI” differently. A company might stop counting a pilot as active, cancel a weak experiment while launching a stronger one, or use AI for administrative work that does not fit a question framed around producing goods or services. Employee use of consumer tools and AI features embedded in existing software can be invisible, too. A drop in this measure could indicate hesitation, but it cannot establish a lasting reversal on its own.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute#1 Best Overall
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
Use is spreading faster than production scale
Other surveys show the gap between having AI somewhere in an organization and deploying it deeply across operations. Their figures are not directly comparable: they cover different respondents and ask different questions. Taken together, however, they suggest that experimentation is common while enterprise-wide impact is harder to achieve.
| Indicator | Reported finding | What it indicates |
|---|---|---|
| Regular use | 88% of respondents in McKinsey’s 2025 survey said their organizations regularly used AI in at least one business function. | Use in at least one function is widespread among respondents; it does not mean the whole company has scaled AI. |
| Scaling AI programs | About one-third of McKinsey respondents said their organizations had begun scaling AI programs. | Most respondents remained in experimentation or pilot stages. |
| Pilot-to-production conversion | In Deloitte’s 2026 enterprise survey, 25% of respondents said their organizations had moved at least 40% of AI pilots into production; 54% expected to reach that threshold within three to six months. | Many organizations had not yet converted a substantial share of pilots, though the expected figure is a forecast, not an achieved result. |
| Expected return | 25% of CEOs surveyed by IBM in 2025 said their AI initiatives had delivered expected ROI. | Financial returns were not yet meeting expectations for most surveyed CEOs. |
| Enterprise-wide scale | 16% of CEOs in the same IBM study said their AI initiatives had scaled enterprise-wide. | Broad deployment remained uncommon among respondents. |
| Enterprise-wide EBIT impact | 39% of McKinsey respondents attributed some level of enterprise-wide EBIT impact to AI; most of those said it was less than 5% of EBIT. | Reported financial impact was often limited rather than transformative. |
Sources: McKinsey’s 2025 State of AI survey, Deloitte’s 2026 enterprise survey, and IBM’s 2025 CEO study.
These findings describe different stages, not contradictory verdicts. A worker using an assistant, a business unit running a pilot, a production workflow, and an enterprise-wide operating model are not equivalent forms of adoption. Nor does access to a tool prove that it has changed a process or improved the bottom line.
Rank #2
- High-Performance AI Processing: The MX3 is designed to handle the most demanding AI computer vision workloads, delivering exceptional performance and efficiency.
- Flexible Integration: The MX3 can be easily integrated into your existing systems via its M.2 M-key form factor and support for Linux operating systems.
- Energy Efficient: The MX3 is designed to provide high performance while minimizing power consumption.
- Comprehensive Software Development Kit (SDK): The MX3 is supported by a comprehensive SDK that simplifies development and deployment.
- Hardware compatability: The MX3 is compatible with the PCI-SIG M.2 M-key 2280 Specification. It can be used with the Raspberry Pi 5 with a M-key 2280 HAT.
Why companies are becoming more selective
Returns can be difficult to demonstrate
Saving time is not automatically the same as reducing costs or increasing revenue. The financial case becomes clearer when saved capacity is used to handle more work, avoid hiring or outsourcing, shorten a revenue-generating cycle, or reduce errors and rework. If no one measures what changed against a baseline, a productivity claim may never become a credible ROI figure. IBM’s 2025 CEO study found that only 25% of surveyed CEOs said AI initiatives had delivered expected ROI, even as those CEOs expected AI investment growth to more than double over the following two years.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesMcKinsey’s 2025 survey found that about 6% of respondents qualified as “AI high performers,” a survey category defined by significant value and at least 5% EBIT impact. That is a much narrower claim than simply using AI or reporting some benefit.
Pilots accumulate without a route to production
A demonstration can work in isolation yet stall when it needs reliable data, system integration, security review, staff training, support, and a production budget. Deloitte identified competing core-business priorities as a source of “pilot fatigue.” It reported that 30% of respondents were redesigning key business processes around AI, while 37% were using AI at a surface level with little or no underlying process change.
Rank #3
- ✅Powered by 26 Tera-Operations Per Second (TOPS) Hailo-8 AI Processor. 2.5W typical power consumption
- ✅Scalable, enabling simultaneous processing of multi-streams & multi-models
- ✅Enabling real-time, low latency and high-efficiency AI inferencing on the edge devices
- ✅Supports TensorFlow, TensorFlow Lite, ONNX, Keras, Pytorch frameworks
- ✅Supports Linux and Windows. Supports the temperature range of -40°C to 85°C
A pilot is more useful when it has a business owner, a defined baseline, a production budget, an integration plan, an adoption target, and a date to scale, revise, or stop. Without those elements, a growing pilot count can signal activity without progress.
Fragmented data and systems raise the cost of scaling
AI applications need appropriate access to useful, well-governed information and dependable connections to the systems where work happens. IBM reported that 50% of surveyed CEOs said rapid investment had left their organizations with disconnected, piecemeal technology; 68% considered an integrated, enterprise-wide data architecture critical to cross-functional collaboration. That helps explain why adding another tool can be easier than making it work across a large company.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Workflow change matters more than a tool launch
Putting a chatbot beside an existing process is not the same as redesigning procurement, claims handling, customer support, software delivery, or finance operations around AI. McKinsey found that high-performing organizations were more likely to redesign workflows and pursue transformative objectives rather than focus only on efficiency. The difficult work includes deciding which tasks should change, who reviews outputs, and how responsibility shifts when a system makes a mistake.
Rank #4
- 48GB AI graphics accelerator
Governance and liability grow with autonomy
Enterprises must account for inaccurate outputs, privacy, intellectual-property exposure, cybersecurity, auditability, human accountability, discrimination, and changing models or vendors. These concerns become more acute when software can take actions rather than draft text for a person to review. Deloitte found that nearly three-quarters of surveyed companies planned to deploy agentic AI within two years, but only 21% of those planning deployment said they had mature agent-governance models. An agent that changes records, sends customer communications, executes payments, or modifies production systems needs authorization boundaries, monitoring, and a way to stop or reverse harmful actions.
Vendor dependence is an architectural risk
IBM’s 2026 AI sovereignty study found that 71% of surveyed executives said switching their primary AI vendor or model would be difficult, and 68% considered data-residency and sovereignty requirements challenging. A company can support AI while hesitating to make a critical workflow dependent on one provider, cloud, or proprietary data architecture. Multiple vendors may provide options, but they can also add cost and operational complexity when each business unit chooses independently.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What is still moving forward
Evidence of hesitation at the scaling stage is not evidence of a universal freeze. McKinsey’s regular-use finding and the Census-based rise in overall business use point to continued adoption at some level. IBM’s CEO study found that 61% of surveyed CEOs were already adopting AI agents and preparing to implement them at scale, while Deloitte reported strong planned interest in agentic AI. Plans and intentions are not the same as deployed systems, but they show that strategic attention and experimentation remain active.
Best Value
- Professional AI & Creator Workstation: AMD Radeon AI PRO R9700 GPU with 32GB GDDR6 is engineered for AI development, professional content creation, and compute-intensive workloads.
- Massive 32GB Memory Capacity: 32GB of GDDR6 memory on a 256-bit bus provides ample bandwidth for large AI models, 8K video editing, and complex 3D rendering.
- Advanced RDNA 4 with AI Accelerators: 64 Compute Units with 3rd Gen Ray Tracing and dedicated 2nd Gen AI Accelerators for groundbreaking AI performance and visual computing.
- Professional Blower Cooling: Efficient single blower design exhausts heat directly out of the chassis, ideal for multi-GPU workstation and server configurations.
- Enterprise-Grade Thermal Solution: Vapor chamber heatsink with industrial Honeywell PTM7950 thermal interface material ensures reliable cooling under sustained professional loads.
Reported value is more credible when tied to a defined workflow and a measurable outcome. McKinsey respondents most often reported cost benefits in software engineering, manufacturing, and IT, and revenue benefits in marketing and sales, strategy and corporate finance, and product or service development. Those survey results do not guarantee that a particular company will get the same result; they do point toward a more useful starting question than “Where can we add AI?”: “Which costly, high-volume process can we improve, and how will we know?”
How to tell whether a company should scale, pause, or stop
A disciplined decision starts with the business problem, not the novelty of the model. Before moving a pilot into production, leaders can test it against these criteria:
- Outcome: Identify a specific cost, revenue, quality, risk, or cycle-time measure the project is meant to change.
- Ownership: Name the business leader accountable for that outcome, not only the technical team running the pilot.
- Baseline: Record current performance before deployment so any improvement can be assessed.
- Workflow fit: Confirm that the process has enough volume to justify integration and that staff can adopt the new way of working.
- Controls: Define the data the system may access, when human review is required, how outputs are audited, and who is accountable for errors.
- Resilience: Understand what happens during a vendor outage, a model change, or a price increase, and whether the company can switch or roll back.
- Decision point: Set a date and explicit evidence threshold for scaling, revising, or ending the project.
Good measures depend on the workflow. They can include cost per completed transaction, average handling time, first-contact resolution, defect or rework rate, conversion rate, software deployment frequency, incident rate, time to resolve incidents, percentage of outputs needing human correction, and total cost per successful task. Pilot counts, prompts, tokens, and licenses describe activity; they do not prove business value.
Pause or cancel when no one owns the process, current performance cannot be measured, the business case depends on near-perfect accuracy, integration costs exceed the task’s value, sensitive data access is uncontrolled, or employees are expected to change their work without training or incentives. A project can also create real value that is recorded outside the AI budget—or not recorded at all—so leaders should decide in advance where benefits such as avoided hiring, reduced outsourcing, or better customer retention will appear.
How the evidence should be read
The surveys cited here use different populations, definitions, and questions; their percentages should not be combined into a single adoption rate. Self-reported regular use, plans to deploy, pilot conversion, expected ROI, and measured financial impact each answer a different question. For example, the Census-based measure described by ITPro is a recent-use indicator, while McKinsey, Deloitte, and IBM report on respondent assessments of organizational use, scaling, plans, and outcomes.
The frequently repeated claim that 95% of AI pilots fail is not used here as a general failure rate: without a clear definition of “pilot” and “failure,” and clarity on whether the measure concerns revenue, profit-and-loss impact, or any return, it cannot settle whether enterprise AI is working. The more supportable reading of the evidence is that adoption is broadening in some forms while many organizations still struggle to turn experiments into governed, integrated production systems with measurable financial results.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
