Generative AI adoption has moved rapidly from curiosity to routine workplace use. According to the 2025 Wharton Human-AI Research and GBK Collective study, 82% of surveyed U.S. enterprise decision-makers used generative AI at least weekly, while 46% used it daily. In 2023, only 37% reported weekly usage. The study surveyed approximately 800 senior decision-makers working at U.S. commercial enterprises with at least 1,000 employees and more than $50 million in annual revenue. The more important development, however, is not simply that more people are using AI. Enterprises are beginning to impose financial and operational discipline on that usage. Seventy-two percent of respondents said their organizations formally measured generative AI returns through business-linked metrics such as productivity, profitability, and throughput. Roughly three-quarters believed their investments were already generating positive returns, and 88% expected generative AI spending to increase during the following year. This does not mean enterprise AI has been fully solved. Adoption remains uneven across industries, departments, seniority levels, and company sizes. Many organizations use generative AI frequently without having redesigned the workflows around it. Others report positive returns but cannot yet isolate exactly how much value comes from the technology rather than from broader process improvements. The central challenge has shifted from access to execution.
Enterprise AI success now depends on five capabilities:
Selecting high-value problems rather than accumulating disconnected use cases. Redesigning complete workflows instead of adding AI to inefficient processes. Measuring financial, operational, customer, and risk outcomes. Building employee competence through role-specific training and supervised practice. Creating governance strong enough to manage risk without making legitimate use impossible. The Wharton study found that 89% of leaders believed generative AI enhanced employee skills, but 43% also saw a risk of declining skill proficiency. Leadership involvement and governance were increasing, yet investment in workforce training had softened even as organizations continued to report technical skill shortages. This creates a strategic contradiction. Companies are spending more on AI while potentially underinvesting in the people responsible for turning it into economic value. The next phase will therefore not be won by deploying the most chatbots. It will be won by organizations that can connect AI to real decisions, reliable data, accountable owners, secure systems, trained employees, and measurable business outcomes.
From Experimentation to Accountable Acceleration The first stage of generative AI adoption was driven by discovery. Employees experimented with public chatbots. Marketing teams generated drafts. Analysts summarized documents. Developers asked AI systems to explain code. Executives attended demonstrations and authorized pilots. Many companies created task forces whose principal responsibility was to determine what generative AI could do. The second stage was driven by expansion. Organizations purchased enterprise licenses, added AI features to existing software, launched internal assistants, tested retrieval systems, and funded multiple pilots. Usage spread across departments. The discussion moved from whether generative AI was relevant to where it could be deployed. The third stage is now emerging: accountable acceleration. In this stage, companies do not stop expanding AI. They expand it under stricter conditions.
Management wants to know:
Which applications create genuine value? Which pilots should be scaled? Which experiments should be discontinued? What expenses are being reduced? What revenue is being created or protected? What work is becoming faster? Are quality and customer satisfaction improving? What new risks have been introduced? Who is accountable when AI-supported work fails? Which human skills must be preserved? What infrastructure is needed to operate AI reliably? This is a healthier stage of technological adoption. It replaces vague excitement with operating discipline.
Wharton’s three-year research illustrates this progression. Weekly usage among surveyed enterprise decision-makers increased from 37% in 2023 to 72% in 2024 and 82% in 2025. The study characterizes the three waves as exploration, experimentation, and accountable acceleration. That progression matters because adoption alone is a weak measure of transformation. A company can issue AI licenses to thousands of employees without materially changing its economics. Employees may use the tools for email editing, meeting summaries, brainstorming, and document preparation. These activities can save time, but they do not automatically create a durable advantage. The competitive value appears when organizations redesign how work moves through the company. For example, an insurer may use generative AI to summarize a claim file. That is useful. A deeper transformation would connect intake, document classification, policy interpretation, fraud signals, adjuster recommendations, customer communication, escalation rules, and audit records within one governed workflow. A retailer may use AI to write product descriptions. A deeper transformation would integrate demand forecasts, customer behavior, inventory availability, merchandising rules, promotional planning, localized content, and performance measurement. A bank may use AI to draft internal reports. A deeper transformation would combine research, document retrieval, policy checks, risk analysis, approval routing, recordkeeping, and human review.
The distinction is fundamental:
Tool adoption improves individual tasks. Workflow transformation improves organizational performance.
What the 2025 Wharton Study Actually Measures The Wharton and GBK Collective report provides a valuable view into enterprise AI adoption, but its scope must be understood correctly. The 2025 wave was based on an online survey conducted between June 26 and July 11, 2025. It included approximately 800 U.S.-based senior decision-makers in functions such as information technology, legal, marketing and sales, operations, finance, human resources, procurement, product and engineering, and general management. Respondents worked for commercial organizations with at least 1,000 employees and more than $50 million in annual revenue. This means the results do not represent every employee, small business, public institution, or global market. They reflect the perspectives of enterprise decision-makers in larger U.S. organizations. That distinction helps explain why the reported usage levels may be higher than those observed in broader workforce surveys. Senior decision-makers are more likely to have access to premium tools, strategic initiatives, internal pilots, and executive-sponsored programs. The survey is also repeated cross-sectional research rather than a panel tracking exactly the same individuals every year. It compares similarly defined groups across the three waves, offering a view of directional change rather than individual behavioral evolution. These limitations do not make the research less useful. They make interpretation more precise.
The report is especially valuable for understanding:
How enterprise leaders perceive adoption. Where AI is entering business functions. Whether organizations are increasing spending. How companies are approaching ROI measurement. Which organizational barriers are becoming more important. How executives view AI’s effect on work and skills. Where adoption differs across industries and company sizes. Its central message is credible and significant: generative AI has moved far beyond isolated experimentation in large U.S. enterprises.
Everyday AI Is Becoming Normal Business Software One of the strongest signals in the report is the normalization of AI. Nearly half of surveyed decision-makers said they used generative AI every day. More than four-fifths used it at least weekly. Respondents increasingly described themselves as competent or expert users, while usage among older professionals was also rising. This suggests generative AI is beginning to resemble other general-purpose workplace technologies. Email, search engines, spreadsheets, collaboration platforms, cloud storage, and video conferencing were once separate innovations. Over time, they became ordinary components of business operations. Employees stopped describing themselves as “email adopters” or “cloud users.” The technologies simply became part of work. Generative AI is moving in the same direction, although its risk profile is more complicated. Unlike a spreadsheet or conventional search tool, a generative model may produce an authoritative-sounding answer that is incomplete, incorrect, fabricated, biased, or inappropriate for the context. It may transform unstructured information in powerful ways, but it does not possess institutional responsibility or human judgment. Therefore, normalization cannot mean uncritical acceptance. The goal is not to teach employees that AI is always right. It is to teach them when AI is useful, what evidence must be checked, which information may be entered, when human approval is mandatory, and how outputs should be documented.
Everyday AI requires everyday competence.
That competence includes more than prompt writing. It includes:
Understanding the limits of the selected model. Recognizing unreliable or unsupported claims. Protecting confidential information. Choosing authoritative internal sources. Reviewing outputs in proportion to the risk. Understanding applicable legal and regulatory obligations. Escalating uncertain or high-impact decisions. Recording material AI involvement when required. Knowing when not to use AI. A company that provides broad access without developing these habits may increase activity without increasing capability.
The Most Valuable Early Use Cases Share Common Characteristics Enterprise AI adoption tends to spread first through work that is text-heavy, repetitive, information-intensive, and easy to review. The Wharton study identified broad use in activities such as data analysis, document summarization, editing, writing, and research. It also found more specialized functional applications, including code generation in information technology, recruitment and onboarding in human resources, and contract generation in legal departments. These use cases have several characteristics in common.
1. They begin with existing information
Generative AI works particularly well when it transforms material that already exists.
Examples include:
Summarizing a long document. Reformatting information. Extracting important clauses. Comparing two versions. Classifying support requests. Converting meeting notes into action items. Drafting a response based on approved source material. Translating technical language into a simpler explanation. This is generally safer than asking a model to create unsupported factual conclusions from nothing.
2. Humans can review the output
Many early use cases keep a knowledgeable employee in the approval loop. An attorney reviews the contract draft. A developer reviews the generated code. A recruiter reviews the candidate summary. A financial analyst reviews the commentary. A customer service representative reviews the proposed response. Human review does not eliminate risk, but it creates a practical control while the organization develops confidence.
3. The work occurs frequently
A small time saving becomes economically meaningful when multiplied across thousands or millions of transactions. Saving five minutes on a task performed ten times a year is unlikely to justify a major investment. Saving five minutes on a task performed 20,000 times per month may be transformative.
4. Success can be measured
High-value use cases usually have observable baselines.
A company may know:
How long the task currently takes. How many people perform it. How much rework occurs. How frequently errors appear. How many cases are completed. How customers rate the experience. How much the process costs. How much delay it creates elsewhere. Without a baseline, an AI pilot may look impressive without proving improvement.
5. Failure is manageable
Organizations often begin with tasks where mistakes can be caught before causing irreversible harm. This is why drafting, summarizing, searching, and recommending tend to precede unsupervised decision-making. The appropriate level of autonomy should depend on the consequence of failure. An AI system that proposes three alternative marketing headlines presents a different risk from one that rejects an insurance claim, approves a loan, modifies production equipment, dispenses clinical advice, or transfers money.
Adoption Is Broad, but It Is Not Even A high enterprise-wide usage percentage can hide substantial variation. The Wharton report found stronger usage in technology and telecommunications, banking and finance, and professional services. Manufacturing and retail showed slower progress. Information technology and procurement were ahead in frequency and confidence, while marketing and sales and operations lagged in parts of the study. At first glance, some of these differences may appear surprising. Retail, for example, contains many potential applications involving merchandising, pricing, marketing, customer service, labor planning, supply chains, and product information. The explanation is not simply that some industries understand AI while others do not. Adoption difficulty depends on the nature of the underlying work. Digital work is easier to transform Professional services and technology companies produce large amounts of digital information. Their workflows often involve documents, code, analysis, communication, and knowledge retrieval. Generative AI can enter these environments through software without requiring physical infrastructure changes. Physical operations introduce additional complexity Manufacturing and retail operate across stores, warehouses, factories, transportation systems, equipment, and distributed workforces. A generative AI assistant may improve administrative work relatively quickly, but transforming the core operating model may require integration with sensors, enterprise resource planning systems, inventory systems, robotics, quality controls, and real-time operational data.
Regulation changes the deployment model Financial services and legal organizations may have high-value information use cases, but they also face stricter requirements around privacy, documentation, fairness, explainability, and accountability. These companies may invest heavily while imposing stronger controls. Data maturity matters Organizations with accessible, reliable, well-governed information can build useful internal AI systems more easily. Companies with fragmented databases, inconsistent records, outdated systems, weak metadata, and unclear ownership may discover that their AI problem is actually a data-management problem. Culture determines whether access becomes usage Two companies can purchase the same technology and produce different outcomes. In one organization, employees may be encouraged to experiment within clear rules. Managers may share successful workflows, and employees may receive training and time to practice. In another, policies may be vague, employees may fear punishment, and managers may not understand the tools. Some workers will use unapproved systems privately while others avoid AI entirely. The technology is similar. The organizational environment is not.
Positive ROI Is Encouraging, but It Requires Careful Interpretation One of the most important findings from the Wharton study is that approximately three-quarters of surveyed leaders believed their generative AI investments were delivering positive returns. Seventy-two percent reported using structured metrics connected to business outcomes, while four in five expected positive returns within roughly two to three years. This represents a major change from the first wave of adoption, when many investments were justified by experimentation, strategic fear, or pressure to avoid falling behind. Still, reported positive ROI should not be confused with audited, company-wide financial transformation. Several measurement challenges remain. Time saved is not automatically money saved Suppose an AI tool reduces the time required to prepare a report from four hours to two. The company has created two hours of capacity. Whether that capacity becomes economic value depends on what happens next.
Value may be realized when:
The employee produces more revenue-generating work. The company completes more cases without adding staff. Customer response times improve. Overtime declines. Hiring needs are avoided. A bottleneck elsewhere is removed. Work quality improves. Employees redirect time toward higher-value decisions. If the employee simply finishes earlier but the organization changes nothing, the time saving may not appear in financial results. Benefits can be counted twice A finance department may report reduced labor costs while an operations team reports increased productivity from the same improvement. A rigorous measurement system should establish ownership and prevent overlapping claims.
Quality may decline while speed improves An AI-supported workflow may process more work but generate additional corrections, customer complaints, legal exposure, or rework. A proper ROI calculation must include both productivity and quality. Costs extend beyond software licenses
The full cost of enterprise AI may include:
Model usage. Cloud computing. Data storage and retrieval. Integration engineering. Cybersecurity. Governance. Legal review. Evaluation and testing. Employee training. Process redesign. Monitoring. Vendor management.
Change management. Human oversight. Incident response. A pilot can appear inexpensive because many of these supporting costs have not yet been included. Pilot economics may not survive scale A small team may obtain excellent results using carefully selected data and enthusiastic participants. Enterprise deployment introduces greater variation in users, data quality, languages, systems, risk levels, and business conditions. Scaling can either improve economics through reuse or weaken them through complexity. This is why the shift toward business-linked measurement is so important. The question is no longer whether employees enjoy using AI. The question is whether AI improves the operating and financial performance of the organization.
A Practical Enterprise AI ROI Framework Companies need a common model for evaluating use cases across departments. A useful framework contains six layers. Layer 1: Adoption Measures whether the intended people actually use the system.
Possible metrics include:
Eligible users. Activated users. Weekly active users. Repeat usage. Feature adoption. Abandonment rate. Number of workflows supported. Adoption is necessary, but it is not proof of value. Layer 2: Productivity Measures how AI changes the effort required to complete work.
Possible metrics include:
Time per task. Tasks completed per employee. Cycle time. Queue length. Response time. Automation rate. Number of manual handoffs. Time spent searching for information. Layer 3: Quality Measures whether the output is accurate, useful, and compliant.
Possible metrics include:
Error rate. Rework rate. First-contact resolution. Escalation rate. Factual accuracy. Policy compliance. Defect rate. Customer satisfaction. Expert approval rate. Layer 4: Financial impact Measures changes in revenue, cost, margin, or capital requirements.
Possible metrics include:
Labor cost avoided. Overtime reduced. Cost per transaction. Revenue per employee. Conversion improvement. Customer retention. Loss prevention. Fraud reduction. Working-capital improvement. Gross margin contribution. Layer 5: Risk Measures the negative exposure introduced or reduced by the system.
Possible metrics include:
Privacy incidents. Unsupported claims. Security events. Biased outcomes. Regulatory violations. Unauthorized tool usage. Intellectual-property exposure. Model failure frequency. Human override frequency. Financial losses associated with errors. Layer 6: Strategic capability Measures whether the company is building an advantage that competitors cannot easily reproduce.
Possible metrics include:
Proprietary data assets improved. Reusable AI components created. Internal knowledge made accessible. New services launched. Time to market. Process flexibility. Experimentation speed. Cross-functional reuse. Customer experience differentiation. Organizational learning. This final layer is often neglected. A project may deliver moderate short-term savings while building a valuable long-term capability. For example, creating a governed enterprise knowledge layer may initially support a customer service assistant. Later, the same foundation could support sales, legal, compliance, onboarding, engineering, and procurement.
The first use case should not bear the entire cost of infrastructure that will serve many future applications.
The Difference Between Productivity and Transformation Most early enterprise AI projects focus on productivity because productivity is visible and relatively easy to measure. Employees can draft faster. Documents can be summarized. Research can be accelerated. Code can be generated. Customer responses can be proposed. Information can be retrieved through natural language. These gains matter. But they are usually incremental. Transformation begins when the company changes the structure of work itself.
Consider a conventional procurement process:
A department identifies a need. Employees search for vendors. Requirements are documented. Proposals are collected. Terms are compared manually. Legal and security teams review documents. Approvals are requested. A contract is negotiated. The vendor is onboarded. Performance is monitored. An incremental AI approach might summarize proposals and draft emails.
A transformative approach might create an integrated procurement intelligence system that:
Converts business needs into structured requirements. Searches an approved vendor universe. Compares proposals. Identifies contractual differences. Checks security documentation. Flags risk. Recommends negotiation positions. Routes approvals. Generates audit records. Tracks commitments and supplier performance. Humans still make consequential decisions, but the workflow becomes faster, more consistent, and more measurable. Transformation therefore requires more than a model. It requires process ownership, system integration, permissions, data access, evaluation, exception handling, governance, and employee redesign.
This helps explain why enterprise AI can spread rapidly at the individual level while organizational transformation remains slower. Using AI is easy. Rebuilding a company around it is hard.
The Human Capital Paradox The Wharton research presents a particularly important contradiction. Eighty-nine percent of surveyed leaders agreed that generative AI enhances employee skills. At the same time, 43% saw a risk that skill proficiency could decline. Nearly half of organizations reported technical skill gaps, yet investment in training had softened, and confidence in training as the primary path to fluency had fallen. Recruiting advanced generative AI talent also remained difficult. This is the human capital paradox of AI. The technology can make workers more capable while simultaneously making some underlying abilities weaker. A junior analyst who uses AI to structure research may learn faster. But if the analyst never verifies sources, constructs an argument independently, or performs foundational analysis, professional judgment may fail to develop. A developer may become more productive with code generation. But excessive dependence may weaken debugging ability or understanding of system architecture. A recruiter may evaluate more candidates with AI assistance. But overreliance may reduce careful reading and reinforce patterns hidden inside historical data. An attorney may review documents more quickly. But if generated clauses are accepted without deep review, expertise may gradually become superficial. The correct response is not to prohibit AI. It is to redesign learning. Employees need to practice both AI-assisted and independent work. They need to understand the reasoning behind the output, not merely accept the finished product.
A mature workforce strategy should distinguish between:
Skills AI should automate. Skills AI should augment. Skills humans must continue practicing. Skills that become more important because AI exists. Completely new skills required to supervise AI systems. The most valuable future employee may not be the person who performs every task manually. Nor will it necessarily be the person who accepts every AI output. It will be the person who can define the problem, direct the system, evaluate the result, detect failure, apply context, and take responsibility.
Why Generic AI Training Often Fails Many organizations respond to the skills challenge by offering a short seminar on generative AI. Employees learn what a large language model is, receive several example prompts, review a list of prohibited activities, and return to work. This creates awareness, but not operational competence. Real capability is role-specific. A finance professional needs different training from a software engineer. A customer service representative needs different controls from a corporate attorney. A marketing employee needs different review standards from a clinical professional. A procurement specialist needs different data access from a human resources manager. Effective enterprise training should contain several layers. Foundational literacy
Employees should understand:
What generative AI can and cannot do. Why outputs may be unreliable. How organizational data may be exposed. Which tools are approved. Which uses are prohibited. When disclosure or documentation is required. Role-specific workflows Employees should learn exactly how AI applies to their work. Examples should use realistic documents, systems, risks, and decision points. Verification skills Training should teach employees how to inspect sources, compare outputs, challenge assumptions, detect fabrication, and escalate uncertainty. Practice environments
Employees need safe spaces to test workflows without exposing confidential data or affecting real customers. Manager education Managers must know how to set expectations, measure outcomes, redesign roles, and judge whether employees are using AI responsibly. Continuous updates AI products change rapidly. A one-time training course will become outdated. Organizations need updated playbooks, internal communities, office hours, use-case libraries, and feedback channels. Training is not an accessory to deployment. It is part of the operating system.
Leadership Is Moving into the C-Suite The Wharton report found that executive leadership in generative AI adoption had increased and that Chief AI Officer roles were present in approximately 60% of surveyed enterprises. This reflects an important shift. During early experimentation, AI programs were often owned by innovation laboratories, data science groups, or technology departments. These teams could prove technical feasibility, but they did not always control business processes, budgets, incentives, or workforce decisions. As AI becomes strategically important, accountability is moving upward. However, appointing a Chief AI Officer does not automatically solve the ownership problem.
An enterprise AI operating model must clarify the responsibilities of:
The board. The chief executive. The Chief AI Officer. The chief information officer. The chief technology officer. The chief data officer. The chief information security officer. Legal and compliance teams. Human resources. Business-unit leaders. Model owners. Process owners.
Frontline managers. Individual users. Central leadership is needed for standards, infrastructure, security, vendor strategy, and governance. Business units are needed to identify problems, own outcomes, redesign workflows, and manage adoption. This usually leads to a federated model.
A central AI function may provide:
Approved models and tools. Shared infrastructure. Identity and access controls. Data connectors. Evaluation systems. Governance standards. Vendor management. Security policies. Reusable components. Training resources.
Business units may provide:
Use-case priorities. Domain experts. Workflow ownership. Baseline metrics. Human review. Change management. Financial accountability. Continuous improvement. Too much centralization can slow adoption. Too much decentralization can produce duplicated spending, inconsistent controls, and incompatible systems. The goal is not centralization or decentralization in isolation. It is coordinated autonomy.
Governance Should Accelerate Safe Use, Not Merely Restrict It Governance is sometimes treated as the department that says no. That approach can produce the opposite of safety. When official tools are too difficult to access or policies are too vague, employees may use public systems privately. This creates shadow AI: unapproved and poorly monitored use outside corporate controls. The Wharton study found that more organizations were adopting data security policies and employee awareness programs as access expanded. It also found increasing use of AI in fraud detection and risk management, suggesting that AI was becoming both an object of governance and a tool used for governance. A mature governance program should make low-risk, high-value use easy while applying stronger controls to consequential applications. A practical classification model may include four tiers. Tier 1: Low-risk assistance
Examples:
Brainstorming. Rewriting nonsensitive material. Meeting summaries. Formatting. General research. Drafting internal communications. Controls may include approved tools, basic privacy rules, and employee review. Tier 2: Internal operational support
Examples:
Enterprise search. Internal policy assistance. Financial commentary. Vendor analysis. Code generation. Employee support. Controls may include authenticated access, approved data sources, logging, output evaluation, and role-based permissions. Tier 3: Customer-facing or regulated support
Examples:
Customer service responses. Contract analysis. Credit recommendations. Claims processing. Hiring support. Medical administration. Controls may include mandatory human approval, documented testing, bias assessment, audit trails, legal review, and ongoing monitoring. Tier 4: High-impact autonomous action
Examples:
Transferring funds. Making binding eligibility decisions. Executing high-value purchases. Changing critical infrastructure. Taking disciplinary action. Issuing clinical decisions. Entering legal commitments. These applications require stringent authorization, technical safeguards, independent validation, limited autonomy, and clear human accountability. Some may not be appropriate for autonomous deployment at all. The U.S. National Institute of Standards and Technology’s Generative AI Profile extends its AI Risk Management Framework to risks specific to generative systems. It is designed to help organizations incorporate trustworthiness into the design, development, use, and evaluation of AI systems. The important principle is proportionality. Not every AI-generated sentence requires a governance committee. Not every automated decision should be treated casually. Controls should correspond to impact.
Build, Buy, or Combine? The Wharton research found that enterprises were allocating roughly one-third of generative AI technology budgets to internal research and development, indicating interest in customized capabilities rather than exclusive reliance on standard software. This does not necessarily mean enterprises are building foundation models from scratch.
Internal AI investment may include:
Connecting commercial models to company data. Building secure retrieval systems. Developing domain-specific assistants. Creating model-routing layers. Evaluating different providers. Constructing agentic workflows. Building proprietary interfaces. Integrating AI with enterprise applications. Establishing monitoring and governance systems. Fine-tuning smaller models for specific tasks. Creating internal datasets and benchmarks. Most companies will use a hybrid strategy.
They will buy general capabilities while building the layers that reflect their unique data, processes, controls, and customer experience. A useful decision framework asks four questions.
1. Is the capability strategically differentiating?
Commodity functions such as generic summarization may be purchased. A capability central to pricing, risk, product design, or customer experience may justify greater internal ownership.
2. Is proprietary data essential?
When performance depends on confidential organizational knowledge, the company may need a controlled internal architecture even when using external models.
3. How serious is vendor dependence?
Companies should consider model pricing, service availability, data terms, regional hosting, product changes, and switching costs.
4. Can the organization maintain what it builds?
Custom AI systems require continuous evaluation, security, integration, and updating. Building creates control but also creates long-term responsibility. The best answer is rarely “build everything” or “buy everything.” It is to buy the commodity, build the differentiation, and preserve the ability to change providers.
The Data Foundation Determines the AI Ceiling Many AI initiatives fail for reasons that have little to do with the model.
The organization’s knowledge may be scattered across:
Email. Shared drives. Customer relationship management systems. Enterprise resource planning systems. Collaboration platforms. Contract repositories. Data warehouses. Ticketing systems. Local employee files. Old databases. Scanned documents. Vendor portals.
Documents may be outdated, duplicated, inconsistent, inaccessible, or owned by nobody. An AI assistant connected to poor information can produce poor answers more efficiently.
Successful enterprise AI therefore requires:
Clear data ownership. Source-of-truth identification. Access permissions. Retention policies. Metadata. Version control. Data quality monitoring. Confidentiality classification. Retrieval testing. Content lifecycle management.
This creates an important strategic insight:
AI adoption often becomes a forcing mechanism for long-delayed organizational cleanup. Companies may begin by trying to build an intelligent assistant and discover that no one knows which policy is current. They may attempt automated analysis and find that business units use incompatible definitions. They may try to connect customer data and discover fragmented consent records. These are not reasons to abandon AI. They are reasons to treat information architecture as part of the AI program.
The Rise of Agentic Systems Raises the Stakes The next stage of enterprise AI extends beyond generating content. Agentic systems can plan steps, call tools, retrieve data, interact with software, and take actions within defined limits. A conventional assistant may draft an email. An agentic system may determine that an email is required, collect the relevant information, prepare the message, request approval, send it, record the action, and schedule a follow-up. A conventional assistant may summarize a vendor proposal. An agentic procurement system may request missing documents, compare bids, check policies, calculate risk, recommend a supplier, and route the decision. This can create far more value, but it also expands the potential impact of failure. An inaccurate answer is one problem. An inaccurate action is another.
Before granting an AI system authority to act, organizations need controls involving:
Identity. Authentication. Authorization. Transaction limits. Tool permissions. Human approval. Logging. Reversibility. Separation of duties. Exception handling. Continuous monitoring. Emergency shutdown.
Liability and accountability. The future enterprise AI architecture will therefore resemble a digital workforce management system as much as a software platform.
Companies will need to know:
Which agent performed an action. Which model was used. Which data was accessed. Which instructions were active. Which person authorized the activity. What changed in the system. Whether the result can be reversed. Who is responsible for review. Accountable acceleration becomes even more important as AI moves from communication into execution.
A 12-Step Enterprise AI Transformation Roadmap Organizations can translate these findings into a practical operating plan. Step 1: Establish the business thesis Define why AI matters to the organization. The thesis should connect AI to strategic priorities such as growth, cost efficiency, customer experience, speed, risk reduction, innovation, or resilience. Step 2: Map workflows, not just ideas Document how important processes currently operate. Identify bottlenecks, waiting time, manual review, repeated information retrieval, errors, and customer pain points. Step 3: Create a use-case portfolio
Classify opportunities by:
Expected value. Feasibility. Data readiness. implementation cost. Risk. Time to impact. Reusability. Strategic importance. Avoid selecting projects only because they are visually impressive. Step 4: Establish baselines Measure existing cost, time, quality, volume, and risk before deployment. Without a baseline, the organization cannot prove improvement.
Step 5: Assign accountable business owners Every important use case should have a business owner responsible for outcomes. Technology teams may enable the project, but the owner of the workflow must own the value. Step 6: Build a shared technology foundation Create approved access to models, data, retrieval, identity, monitoring, evaluation, and logging. Shared foundations reduce duplicated work. Step 7: Define risk tiers Classify use cases according to the consequence of failure and apply proportional controls. Step 8: Train by role Move beyond generic AI awareness. Teach employees the workflows, limitations, verification practices, and policies relevant to their responsibilities. Step 9: Pilot inside real operations
Test with realistic data, actual users, and measurable business conditions. A laboratory demonstration is not proof of operational performance. Step 10: Measure multidimensional value Track adoption, productivity, quality, financial outcomes, risk, and strategic capability. Do not rely on one metric. Step 11: Scale reusable components When a use case succeeds, identify which data connectors, prompts, evaluations, interfaces, policies, and workflow components can support other applications. Step 12: Retire weak projects An accountable AI portfolio should discontinue projects that fail to create enough value. Stopping a weak pilot is not failure. It is disciplined capital allocation.
What Small and Mid-Sized Companies Can Learn Although the Wharton survey focused on large U.S. enterprises, its lessons also apply to smaller companies.
Smaller firms may lack large AI departments, but they often possess important advantages:
Fewer legacy systems. Faster decision-making. Shorter approval chains. Greater process flexibility. Closer communication between leadership and frontline staff. Easier experimentation. Less organizational fragmentation. The report found that smaller enterprises continued to perceive themselves as more agile, even as very large companies narrowed the usage gap. A small company should not imitate the bureaucracy of a global enterprise.
It can begin with a focused approach:
Select three high-frequency workflows. Measure the current time and cost. Use approved commercial tools. Protect sensitive information. Keep a human reviewer. Compare results over 60 to 90 days. Expand only where value is proven. A 50-person company may benefit more from automating customer intake, proposal preparation, knowledge retrieval, and reporting than from creating an elaborate AI innovation office. The principle remains the same: value before spectacle.
What AI Vendors and Startups Should Learn The transition toward accountable acceleration changes what enterprise customers will buy. Early buyers may have purchased novelty, demonstrations, and general productivity promises.
Mature buyers will demand:
Clear business cases. Integration with existing systems. Security documentation. Permission controls. Reliable data handling. Evaluation results. Audit logs. Administrative visibility. Usage analytics. Cost controls. Human approval mechanisms. Vendor stability.
Exit and portability options. Measurable implementation outcomes. Startups that sell only model access may struggle as those capabilities become widely available. The stronger opportunity is often to solve a complete business problem. Instead of selling an “AI assistant for finance,” a vendor might reduce month-end reporting time. Instead of selling an “AI legal copilot,” it might shorten contract review while maintaining approved playbooks and audit records. Instead of selling an “AI sales agent,” it might improve qualified pipeline, response speed, and customer relationship management accuracy. Enterprise customers increasingly care less about the novelty of the model and more about the reliability of the outcome.
This favors companies that combine:
Domain expertise. Workflow integration. Proprietary data structures. Evaluation systems. Governance. Implementation services. Change management. Outcome measurement. The defensible product is not merely the AI. It is the operating system around the AI.
The New Competitive Divide The emerging AI divide will not simply separate adopters from non-adopters. Most significant organizations will use AI in some form.
The more meaningful divide will separate:
Casual users from disciplined operators. Individual productivity from workflow transformation. Disconnected pilots from shared infrastructure. Vague enthusiasm from measurable outcomes. Generic training from role-specific capability. Restrictive governance from enabling governance. Model dependence from architectural flexibility. Automation from responsible autonomy. Short-term savings from durable strategic advantage. The Wharton study identified a group of decision-makers using generative AI less than weekly and warned that cultural restrictions, low trust, budget pressure, and slower-adopting environments could widen the gap between AI-enabled organizations and those struggling to keep pace. Yet moving quickly without discipline creates another form of disadvantage. A company can waste money, expose data, weaken skills, create unreliable processes, and accumulate technical debt in the name of acceleration.
The winning position is not maximum speed. It is maximum learning velocity under accountable control.
Key Takeaways
Generative AI has become mainstream among enterprise decision-makers. Weekly usage in the Wharton study rose from 37% in 2023 to 82% in 2025, while daily use reached 46%. The enterprise conversation has shifted from adoption to accountability. Companies increasingly want evidence of productivity, profitability, throughput, quality, and risk reduction. Frequent usage does not equal enterprise transformation. Durable value comes from redesigning complete workflows rather than adding AI to isolated tasks. ROI requires more than counting hours saved. Organizations must determine whether time savings become revenue, capacity, cost reduction, faster service, or better decisions. People and processes are becoming the primary constraints. Technology access is expanding faster than organizational readiness, role redesign, and training. AI can augment skills while also creating skill atrophy. Companies must decide which abilities should be automated and which human capabilities must continue to be practiced. Training must be role-specific and continuous. Generic prompt workshops are insufficient for operational, regulated, or high-impact work. Governance should make responsible use easier. Excessively restrictive or unclear policies can push employees toward unapproved shadow AI. The best enterprise architecture will usually be hybrid. Companies can buy general capabilities while building differentiated workflows, data connections, controls, and user experiences. Data quality determines the value ceiling. AI cannot reliably compensate for outdated, fragmented, contradictory, or inaccessible organizational information.
Agentic AI will require stronger controls than content generation. Systems that act need identities, permissions, limits, logs, human approvals, and reversible actions. The future competitive divide will be based on execution quality. Nearly every large company may adopt AI, but only some will convert it into repeatable economic advantage.
Frequently Asked Questions
What is the main conclusion of the 2025 Wharton AI Adoption Report?
The report concludes that generative AI has moved from enterprise experimentation into widespread, routine use. Organizations are now applying greater financial accountability, expanding governance, and focusing more heavily on workforce readiness and measurable returns.
How many enterprise leaders use generative AI regularly?
In the 2025 Wharton and GBK Collective study, 82% reported using generative AI at least weekly and 46% reported daily usage. The survey represented approximately 800 U.S.-based senior decision-makers in large commercial enterprises.
Are companies achieving a positive return from generative AI?
Approximately three out of four respondents believed their organizations were receiving positive returns. However, reported ROI may include productivity improvements and management estimates rather than fully audited company-wide financial results.
Which business functions are using generative AI?
Applications include data analysis, research, document summarization, writing, editing, software development, recruiting, onboarding, legal drafting, procurement, fraud detection, risk management, marketing, and customer support. The depth of adoption varies considerably by function and industry.
Does AI primarily replace employees or improve their capabilities?
Most respondents believed generative AI enhanced employee skills, although many also believed it could replace portions of existing work. The report suggests that augmentation is currently the dominant leadership perspective, but hiring and role effects remain uncertain.
What is skill atrophy in the context of AI?
Skill atrophy occurs when employees become so dependent on AI-generated work that their underlying ability to research, reason, write, calculate, diagnose, code, or exercise judgment declines.
Why do AI pilots fail to scale?
Common causes include weak data, poor integration, unclear business ownership, insufficient employee training, unreliable outputs, security concerns, difficult change management, and lack of measurable outcomes.
Should companies build their own AI models?
Most companies do not need to build a foundation model. They may gain more value from connecting existing models to proprietary data, developing specialized workflows, creating governance systems, and building applications around unique business processes.
What is the best way to measure enterprise AI ROI?
Companies should measure adoption, productivity, quality, financial impact, risk, and strategic capability. The measurement should compare results against a documented pre-deployment baseline.
What is accountable acceleration?
Accountable acceleration is the expansion of AI usage under stronger business discipline. It combines rapid adoption with measurable outcomes, responsible governance, workforce preparation, and clear ownership.
What is shadow AI?
Shadow AI is employee use of unapproved AI tools or systems outside the organization’s official security, data, procurement, and governance controls.
What should companies do before deploying AI agents?
They should define permissions, identities, transaction limits, approval requirements, monitoring, audit logs, escalation rules, reversibility, and human accountability before allowing systems to take actions.
Will a Chief AI Officer solve enterprise AI adoption challenges?
A Chief AI Officer can improve strategy and coordination, but the role cannot replace business ownership, data governance, technology leadership, workforce development, or executive accountability.
Is enterprise AI adoption still increasing?
The Wharton survey found that 88% of respondents expected generative AI spending to increase during the following 12 months, while 62% anticipated increases of at least 10%.
Conclusion
Generative AI has entered a more serious phase of enterprise adoption. The age of experimentation is not completely over, but experimentation is no longer enough. Senior leaders increasingly expect AI investments to meet the same standards applied to other major business programs. They want measurable gains. They want accountable owners. They want secure deployment. They want trained employees. They want reliable operations. They want evidence that AI is improving the business rather than merely increasing software activity. The Wharton research captures this transition clearly. Usage has become mainstream. Spending remains strong. Positive returns are being reported. Executive leadership is increasing. Governance is becoming more formal. Yet workforce preparation, organizational coordination, and skill development remain unresolved. The companies most likely to benefit will not treat AI as a magical replacement for people or as a collection of disconnected productivity tools. They will treat it as a new operating capability. They will identify valuable problems, redesign workflows, improve their data foundations, measure real outcomes, protect human judgment, establish proportional controls, and create an environment in which employees can learn responsibly.
The central enterprise AI question is therefore changing.
It is no longer:
Can this technology generate useful work? The answer is clearly yes.
The question now is:
Can the organization repeatedly convert that capability into trusted, measurable, and scalable economic value? That is the challenge of accountable acceleration. It is also where the next generation of competitive advantage will be built.
Relevant Articles and Resources
Primary Source 2025 AI Adoption Report: Accountable Acceleration, Gen AI Fast-Tracks into the Enterprise Knowledge at Wharton, Wharton Human-AI Research, and GBK Collective. The report examines three years of changing enterprise generative AI adoption, usage, investment, ROI, workforce readiness, and governance. Additional Research The 2025 AI Index Report Stanford Institute for Human-Centered Artificial Intelligence. A broad, data-driven examination of AI research, investment, adoption, performance, policy, and economic impact. The State of AI: Global Survey 2025 McKinsey & Company. Research on the expansion of enterprise AI, agentic systems, workflow redesign, and the continuing difficulty of scaling AI beyond pilots. Superagency in the Workplace: Empowering People to Unlock AI’s Full Potential at Work McKinsey & Company. Research examining employee readiness, leadership barriers, organizational maturity, and the human side of workplace AI adoption. The State of Generative AI in the Enterprise Deloitte AI Institute. Survey research examining enterprise deployment, measurable value, operational barriers, workforce access, risk, and the emergence of agentic AI.
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile U.S. National Institute of Standards and Technology. A voluntary cross-sector framework for identifying and managing risks associated with generative AI systems. The Learning Organization: How to Accelerate AI Adoption McKinsey & Company. An examination of organizational barriers, employee experimentation, management behavior, and the practices required to move generative AI from individual use to institutional transformation.