Is your AI strategy teaching employees to stop thinking?

Uncontrolled generative AI use leads to organization-wide knowledge decay

Generative AI creates content faster than most companies can verify it. That is the core risk. The problem is not AI use itself. It is the growing volume of AI-generated work that enters business processes without enough human review.

Matthias Holweg, professor at the University of Oxford’s Saïd Business School, and Thomas H. Davenport, analyst, call the resulting problem “knowledge decay.” In their Harvard Business Review post, they write: “When slopification happens at scale and in sequence across a business’s processes, those processes themselves, and their outputs, start to deteriorate.”

This matters because one weak AI output rarely stays isolated. A generated summary can inform a report. That report can feed a management presentation. The presentation can influence a decision. If employees assume earlier material was checked when it was not, errors and unsupported claims can move through the company and acquire false credibility.

The deeper constraint is therefore not access to AI. It is the organization’s capacity to preserve reliable knowledge while AI increases the speed and volume of content production. Employees still need to know what information is authoritative, where it came from, and which conclusions involved real human judgment. Without those controls, faster content production can create more verification work rather than more useful output.

Holweg and Davenport also identify a behavioral risk. Heavy dependence on generated answers can reduce the critical thinking employees apply to routine work. Over time, people may become less willing or less able to challenge the information moving through a process. As the authors state: “Eventually, people start to lose trust in the processes that they rely on to do their jobs.”

For executives, this changes the AI governance question. Measuring adoption rates, generated content, or time saved is not enough. Leaders also need to measure whether AI-assisted work remains accurate, traceable, and useful. AI should reduce the cost of producing reliable work, not simply reduce the cost of producing content.

The opportunity remains substantial. Companies can use AI to accelerate work while protecting institutional knowledge. But that requires clear rules about where AI can generate information, where humans must verify it, and which source data must remain available. The objective should be controlled augmentation: higher output without lower confidence in the knowledge behind it.

Verification costs can erase AI’s expected productivity gains

AI can make the first version of a document much cheaper to produce. That does not mean it makes the complete process cheaper. The economic question is how much work remains after generation.

Verification is one of those hidden costs. Holweg and Davenport define the problem as separating authentic human work from AI-generated material that may contain significant errors. Checking that material takes time, subject expertise, independent research, and critical judgment. In some workflows, they argue, this additional effort can cancel the efficiency gained from using AI in the first place.

Recruitment shows the problem clearly. Candidates can use generative AI to improve resumes and CVs. They can also tailor material to automated ranking systems. More advanced use can extend into interviews, with candidates using AI tools to generate answers to questions in near real time.

The result is an information-quality problem for recruiters. A polished application or fluent interview answer is no longer strong evidence that the candidate personally possesses the knowledge reflected in it. Holweg and Davenport argue that organizations can end up advancing candidates who are subpar or simply unsuitable for the role. Recruiters may then need more controlled assessment, including on-site interviews where candidates cannot easily access AI.

This shifts rather than eliminates work. Candidates save time producing applications. Recruiters may spend more time establishing whether those applications represent genuine capability. A process can therefore become more automated at one stage while becoming more expensive at another.

Executives should apply this principle beyond hiring. The relevant metric is net process productivity. If AI saves 30 minutes during content creation but requires 40 minutes of expert verification, there is no productivity gain. That 30-versus-40-minute example is illustrative, not data reported by Holweg and Davenport, but it captures the measurement executives need: generation time saved minus review, correction, escalation, and downstream rework.

The best response is not to verify every AI output with the same intensity. Companies should design controls around consequence and uncertainty. Low-risk drafting can tolerate lighter review. Decisions involving hiring, financial commitments, customers, compliance, or other material consequences require stronger evidence and accountability.

Recruitment can also be redesigned around information that is harder to manufacture. Holweg and Davenport recommend structured applications that request specific facts, such as roles held, projects completed, team members involved, suppliers served, and budgets managed. The objective is to gather evidence that can be checked rather than reward candidates for producing persuasive prose.

The executive priority is clear. Do not calculate AI ROI from the speed of generation alone. Measure the full workflow. AI creates economic value only when the total cost and time required to reach a trusted result decline.

Organizations must prove where human expertise adds value in AI-assisted work

Generative AI has made standard business content cheap to produce. Reports, presentation slides, summaries, and routine analysis can now be created in minutes. This changes the economics of knowledge work. Clients and executives will increasingly question what they are paying for when much of the visible deliverable can be generated automatically.

Holweg and Davenport define this as a knowledge-validation problem. The issue is not whether AI was involved. It is whether the organization can identify the valuable human contribution within an AI-assisted workflow.

Consulting provides a clear example. A firm can use generative AI to produce standard written reports and PowerPoint slides. But clients do not primarily pay consultants to create documents. They pay for expertise, judgment, contextual knowledge, and recommendations based on evidence. If AI creates much of the final document, the firm must still demonstrate where its professionals supplied those capabilities.

As Holweg and Davenport write: “Human experts now have to justify not only the quality of the output submitted, but also that actual human intellectual work has produced it.”

This creates an important distinction for executives. Output quality and human value are not the same metric. A polished document can be useful even when AI produced most of it. But in work sold as expert advice, the organization also needs to establish who selected the evidence, challenged assumptions, interpreted uncertain information, and accepted responsibility for the conclusions.

The same principle applies internally. If employees use AI for analysis, management needs to know which conclusions came from verified company information and which were generated by a model. This becomes more important as AI enters finance, strategy, recruitment, performance management, and other workflows where decisions have material consequences.

The management response should not be to require humans to manually produce work that AI can perform efficiently. That would preserve effort rather than value. Instead, companies should move human capacity toward tasks where judgment matters: defining the problem, supplying reliable evidence, evaluating options, challenging weak conclusions, and making accountable decisions.

This also changes how professional work should be measured. Page counts, slide production, drafting hours, and similar output measures become less meaningful when AI can produce them at very low marginal cost. Executives should place more weight on decision quality, verified insights, customer outcomes, and the quality of the evidence supporting recommendations.

The opportunity is to separate expertise from document production. AI can reduce the effort required to package information. Human professionals can spend more time creating the knowledge and judgment that make that information valuable.

Repeated AI processing can cause knowledge entropy and weaken ground truth

Information quality can decline when organizations repeatedly use generative AI to transform existing material. Holweg and Davenport call this “knowledge entropy.” Each round of summarizing, rewriting, combining, or regenerating content can move the result further from the original source.

The technical reason matters. Large language models are probabilistic systems. They generate outputs by predicting tokens based on learned patterns and the context supplied to them. They do not provide an inherent guarantee that every generated statement is factually correct. Holweg and Davenport describe LLMs as “context-agnostic” statistical models with “no conception of fact or truth and simply predict the most likely outputs.”

That description should not be interpreted to mean LLMs cannot produce accurate or context-sensitive answers. Modern models can process extensive context and can perform many knowledge tasks effectively. The business constraint is narrower: generated text should not be treated as authoritative evidence merely because it is fluent, detailed, or consistent with earlier generated material.

Repeated transformation increases this concern. As Holweg and Davenport state: “The greater the number of iterations of content through an LLM, the more it will depart from the original.” A customer interview may first become an AI summary. That summary may feed another report, which may then be condensed for management. If later stages rely only on generated versions, important qualifications, facts, and contextual details from the original source can disappear or change.

Executives therefore need to distinguish source information from generated representations of that information. AI summaries are useful for speed and accessibility, but the original evidence should remain available for verification. Important decisions should be traceable back to source documents, transactions, interviews, measurements, or other authoritative records.

A related technical concern appears when synthetic data, information generated by models rather than collected from original sources, is used to train other models. The article identifies the resulting degradation risk as “generative inbreeding” or model collapse. Repeated dependence on manufactured training data can reduce model accuracy and variability.

The article does not provide numerical findings or cite a specific model-collapse study, so executives should not infer a quantified level of risk from it. The underlying issue is nevertheless relevant to enterprise architecture: provenance matters. Organizations need to know whether important information comes from an original source, a human interpretation, an AI transformation, or another model’s output.

The practical response is to preserve data lineage. High-value information should carry sufficient metadata to identify its source and major transformations. When AI summarizes customer interviews, research, financial information, or operational records, the generated output should remain connected to the underlying material.

This does not require abandoning iterative AI workflows. It requires designing them so that convenience does not replace evidence. Companies that preserve ground truth can use AI aggressively while retaining the ability to audit outputs, correct errors, and make decisions from authoritative information rather than accumulated generations.

Enterprises should restrict AI use to workflows where it creates clear value

The first AI governance question should be simple: does AI improve the result or reduce the total cost of producing it? If the answer is unclear, there is little reason to add AI to the workflow. Adoption by itself is not a business outcome.

Matthias Holweg, professor at the University of Oxford’s Saïd Business School, and Thomas H. Davenport, analyst, argue for explicit limits on employee AI use. Their position is that generative AI should be applied where it creates identifiable value, rather than becoming the default tool for every knowledge task.

Recruitment shows why. Candidates can use generative AI to optimize CVs, tailor applications to automated screening systems, and prepare answers. This makes polished writing a weaker indicator of an applicant’s actual experience or capability. A recruitment process designed around documents that AI can easily optimize may therefore become less informative.

Holweg and Davenport propose changing the input rather than adding more AI detection. Employers can use structured applications that require concrete facts: a candidate’s role on a project, work completed, team members involved, suppliers served, or budgets managed. Recruiters can then verify those claims against interviews, references, and other evidence.

The broader principle is useful well beyond recruitment. Companies should decide which information must come directly from a person, which tasks AI may support, and which activities AI can perform with limited supervision. Those rules should reflect the consequence of an error. A draft internal email does not require the same controls as financial reporting, hiring decisions, legal analysis, or customer commitments.

Complete prohibition is also unlikely to be the right default. AI can create substantial value in appropriate tasks, and some AI capability is already embedded in standard workplace software. The stronger approach is controlled use with a defined purpose, approved data access, clear accountability, and review requirements based on risk.

Disclosure is part of that model. As Holweg and Davenport write: “Content does not need to be entirely human-created, but if AI is being used, be clear why and how.” That gives managers and downstream users essential context about how an output was produced.

Executives should therefore govern AI around business outcomes, not tool usage. Useful measures include total cycle time, error and rework rates, decision quality, verification cost, and customer outcomes. An AI system that makes one task faster but creates additional checking downstream has not necessarily improved productivity.

The objective is selective automation. Give AI work where its speed, scale, or synthesis capability improves the complete process. Preserve stronger human controls where authenticity, accountability, and verified expertise determine the value of the result.

AI creates the most value when it synthesizes strong human input or removes low-value production work

Generative AI does not need to originate knowledge to be useful. In many enterprise workflows, its strongest role is transforming reliable information that already exists. It can summarize, organize, reformat, compare, and produce alternative versions while people remain responsible for the underlying facts and judgment.

Holweg and Davenport point to office productivity software as one example. AI capabilities in tools such as Microsoft Copilot and Google Gemini can quickly produce different versions of reports and PowerPoint slides. The authors argue that this makes it “virtually pointless” for employees to manually create additional versions of the same material.

That is a meaningful productivity opportunity because document production is not necessarily where scarce human expertise delivers the highest return. If AI handles formatting, restructuring, summarization, and first drafts, employees can devote more time to deciding what the document should say and whether its conclusions are supported.

Performance evaluations provide a more consequential example. A manager could ask AI to generate a standard review and receive polished but generic bullet points. That saves drafting time but may add little useful information.

Holweg and Davenport recommend a different workflow. Managers first collect specific feedback from team members and customers. That material supplies direct observations about the employee’s work. AI can then synthesize those inputs into a structured evaluation. The machine handles consolidation; people supply the evidence and remain accountable for the assessment.

This distinction matters for C-suite leaders because AI output quality depends heavily on the quality of its inputs and the design of the workflow around it. Giving a model weak or generic information and asking it to produce a more sophisticated document does not create missing organizational knowledge. It produces a more refined representation of the information available to the system, potentially supplemented by model-generated claims that require verification.

Executives should therefore focus AI programs on work with high transformation costs but strong underlying data. Customer-feedback synthesis, meeting summarization, document comparison, internal knowledge retrieval, report drafting, and content restructuring can fit this pattern when the source information is reliable and appropriate controls are in place.

Human responsibility remains important. Employees should define the objective, select trustworthy inputs, review material claims, resolve uncertainty, and approve consequential outputs. AI can reduce the manual effort between those stages.

There is also a financial implication. Companies should avoid spending scarce employee time on output production that software can perform at much lower marginal cost. The higher-value investment is human attention directed toward evidence, judgment, customer knowledge, problem definition, and decisions.

Holweg and Davenport’s examples point to a practical AI strategy: automate the production burden before automating the judgment. This allows enterprises to capture efficiency while preserving the human knowledge that gives the final output its business value.

AI governance must cover the full business process, not isolated tasks

AI can make one task faster while making the wider process less efficient. This is the central governance problem. Executives need to measure the performance of the complete workflow, including verification, corrections, handoffs, and downstream decisions.

Matthias Holweg, professor at the University of Oxford’s Saïd Business School, and Thomas H. Davenport, analyst, argue that companies should examine how individual AI use affects the broader process. Their example is an interorganizational revenue cycle. Everyone involved should understand where AI is being used, what it is doing, and what happens to its outputs at later stages.

This matters because business processes depend on connected inputs and outputs. AI-generated information from sales may enter finance, operations, customer service, or compliance systems. A local productivity gain has limited value if another team must spend additional time validating the output or correcting incomplete information.

Executives should therefore distinguish task performance from process performance. Whether AI can perform an individual activity better than a human is only part of the decision. As Holweg and Davenport argue, the more important question is whether AI can take over that activity in a way that makes the broader process more efficient.

This requires clear ownership. For each important AI-enabled workflow, an enterprise should know who owns the source data, who approves AI use, who validates material outputs, and who is accountable when the result moves into another business function. Without this structure, AI can distribute responsibility while leaving no single owner for quality.

Cross-company processes require even stronger coordination. Suppliers, customers, advisers, and service providers may each use different models and controls. If AI-generated information crosses organizational boundaries, participants need agreed rules for disclosure, data quality, verification, and acceptable use. Otherwise, one company’s automation choices can create another company’s verification costs.

The measurement model must change as well. Executives should track end-to-end cycle time, error rates, rework, verification effort, exception rates, and final business outcomes. A reduction in the time required for one activity is not sufficient evidence of a productivity gain if total process cost rises.

This approach does not require central approval for every AI interaction. Governance can be proportional to risk. Low-impact applications can operate under broad policies, while workflows affecting revenue, financial reporting, customers, employees, compliance, or major decisions need stronger controls.

The goal is coordinated automation. AI should remove work from the complete process rather than transfer effort from one employee or department to another. That gives executives a much stronger test for AI investment: measurable improvement in the end-to-end business outcome.

Company-specific AI can deliver more enterprise value than generic public LLMs

Holweg and Davenport take a strong position on public large language models. They argue that generic public LLMs add “little to no real value” because their knowledge is broad and their outputs can contain mistakes. They instead point toward small language models and proprietary models trained on company-specific data as tools that can augment human work more effectively.

The underlying business issue is specificity. Enterprises do not compete primarily on access to generic information. They compete through proprietary data, operating knowledge, customer relationships, specialized processes, intellectual property, and the quality of their decisions. AI becomes more valuable when it can work with that organization-specific context.

A company-specific system can, for example, retrieve approved policies, product documentation, customer information, internal research, or operating procedures. With the right architecture and access controls, employees can receive answers grounded in company information rather than depending only on knowledge learned during a general model’s training.

Executives should, however, separate company-specific AI from the assumption that every company needs to train its own model. These are different decisions. Enterprises can often add proprietary context to a capable general-purpose model through retrieval systems, controlled data connections, system instructions, and other forms of customization. Training or fine-tuning a dedicated model becomes appropriate when the use case, economics, security requirements, or performance requirements justify it.

The same nuance applies to small language models, or SLMs. Smaller models can offer lower inference costs, reduced latency, easier deployment in controlled environments, and sufficient performance for narrow tasks. They are not automatically more accurate than large models. Their value depends on the task, training data, evaluation criteria, and deployment design.

Holweg and Davenport’s criticism of public LLMs should also be read as their assessment rather than a universal empirical conclusion. Public and general-purpose models can provide substantial value for coding, drafting, translation, research support, summarization, and other activities. The enterprise question is whether a generic model has enough verified context to support the specific business decision being made.

Security and governance also influence the choice. Sending proprietary information into an uncontrolled public service can create confidentiality, privacy, intellectual-property, or regulatory risks. Enterprise versions of general-purpose models may provide stronger contractual protections, access controls, data-retention policies, and administrative oversight. Leaders should assess the specific deployment rather than treating every public or externally developed model as having the same risk profile.

The stronger strategy is therefore not simply “small models instead of large models.” It is to match model architecture to business value. Generic models can handle general tasks. Company-grounded systems should support activities where proprietary knowledge matters. Specialized models are justified when they produce better economics, control, or performance for a defined workload.

This shifts competitive advantage away from basic access to an LLM. As access becomes widespread, the more defensible capabilities are high-quality proprietary data, effective workflow integration, rigorous model evaluation, and institutional knowledge that competitors cannot easily reproduce.

For the C-suite, this creates a practical investment test. Do not fund proprietary AI because ownership itself sounds strategically valuable. Fund it when company-specific knowledge measurably improves accuracy, cost, speed, security, or business outcomes. The architecture should follow the use case, not the other way around.

Organizations must preserve data provenance and access to verified ground truth

AI-generated information is only as trustworthy as the evidence behind it. When companies cannot trace an AI output to its source, verification becomes slower and accountability becomes weaker. For important business decisions, provenance should therefore be part of the AI architecture from the start.

Matthias Holweg, professor at the University of Oxford’s Saïd Business School, and Thomas H. Davenport, analyst, argue that enterprises should track the history of both structured and unstructured data. This means retaining information about where data originated, how it was changed, and which AI systems or people transformed it.

The distinction between source material and generated material is critical. A customer interview, for example, contains direct statements, facts, emotions, and contextual information. An AI summary can make that interview easier to use, but it is a derived output. It may omit qualifications, compress conflicting views, or emphasize details differently from the original conversation.

Organizations should therefore preserve the original material and connect AI-generated outputs back to it. If an executive questions a conclusion in a customer-research report, employees should be able to inspect the underlying interviews rather than relying on another generated summary. The same principle applies to contracts, financial records, research, operating data, employee information, and other consequential sources.

This requirement becomes more important as AI-generated information passes between systems. Without recorded provenance, a downstream employee or model may have no reliable way to distinguish an original fact from a generated interpretation. Once that distinction disappears, unsupported content can start functioning as organizational knowledge.

Provenance also supports accountability. A useful enterprise system should record the source, relevant version, transformation history, and, where necessary, human approval associated with important outputs. These controls make it easier to investigate mistakes, correct information, reproduce decisions, and demonstrate how an output was produced.

The appropriate level of traceability should depend on consequence. Maintaining extensive lineage for every low-risk AI draft would create unnecessary overhead. Financial, legal, regulatory, strategic, customer-facing, and other high-impact outputs warrant stronger controls because errors can produce material costs.

Data quality remains a separate requirement. Provenance can show where information came from, but it cannot make a poor source accurate. Enterprises still need rules covering authoritative sources, data ownership, access permissions, retention, and quality. AI governance works best when it builds on strong data governance rather than operating as a separate program.

For C-suite leaders, the architectural priority is straightforward: generated content must not become a substitute for authoritative information. AI can transform and summarize enterprise knowledge at scale. The underlying evidence should remain identifiable, retrievable, and verifiable.

Holweg and Davenport’s position is that companies should understand their “ground truth” information and preserve links to authentic content when AI modifies or summarizes it. This gives enterprises the controls needed to use generative AI extensively without losing confidence in the information on which decisions depend.

Poorly governed AI could repeat the productivity problem seen with earlier corporate computing

Technology adoption does not guarantee productivity growth. Holweg and Davenport warn that generative AI could repeat the “productivity paradox” associated with the expansion of corporate computing roughly half a century ago. Companies may invest heavily in technology while failing to capture corresponding improvements in business performance.

The authors state: “If we fail to address the uncontrolled proliferation of generative AI in our business processes, we are likely to see a rerun of the ‘productivity paradox’ observed with the growth of corporate computing half a century ago.”

The key constraint is process design. Installing AI into an existing workflow can make individual activities faster without removing the steps, controls, incentives, or organizational structures that limit overall productivity. Additional verification and correction work can further reduce the net benefit.

This is particularly important with generative AI because content creation is cheap. Employees can produce more emails, reports, presentations, code, applications, and analyses in less time. But higher output volume is not the same as higher economic output. If colleagues must read, verify, correct, or respond to substantially more generated material, AI can increase organizational workload despite reducing the cost of individual production.

Executives therefore need to measure AI productivity beyond adoption and activity. The number of employees using an AI assistant, prompts submitted, documents generated, or hours reportedly saved can help track usage, but they do not establish business value.

The stronger measures sit closer to outcomes: end-to-end process cost, cycle time, error and rework rates, revenue impact, customer outcomes, employee capacity, and the quality of important decisions. The relevant metric depends on the workflow, but it should establish whether AI improves the final result after verification and downstream work are included.

Historical research gives additional context to the productivity-paradox concept, although these studies were not cited in the source article. Economist Robert Solow famously observed in 1987 that the computer age could be seen “everywhere but in the productivity statistics.” Later research offered a more developed explanation. Erik Brynjolfsson and Lorin Hitt argued that information technology often delivers larger benefits when companies make complementary organizational and process changes rather than treating technology investment as sufficient on its own.

That distinction is highly relevant to generative AI. Companies may need to redesign roles, approvals, information flows, performance measures, and controls before AI’s full productivity benefits become visible. Simply adding an assistant to every employee’s existing work does not address inefficient underlying processes.

There is also reason for a constructive outlook. AI provides capabilities that can directly reduce knowledge-work costs, and organizations are gaining a clearer understanding of where those capabilities perform well. The management task is to convert technical performance into operational improvement.

For the C-suite, this means shifting the objective from AI adoption to business redesign. Identify expensive or slow processes. Determine the true constraint. Decide which work AI can remove, which decisions still require human judgment, and how outputs will be verified. Then measure the complete process after deployment.

Holweg and Davenport’s warning is ultimately about execution rather than technological potential. AI can improve productivity. But companies will capture that value only when they redesign work around the technology and control the new costs that AI introduces.

Sustainable AI advantage depends on combining human capital with organization-owned AI capabilities

Microsoft Chairman and CEO Satya Nadella frames enterprise AI around two assets: “human capital” and “token capital.” Human capital includes employees’ knowledge, judgment, relationships, ingenuity, and ability to recognize patterns. Token capital refers to AI capabilities that a company builds and owns. His argument is that the strongest economic value comes from combining the two.

The distinction matters because access to AI models is becoming widely available. Simply giving employees an AI assistant is therefore unlikely to create a durable competitive advantage. The more defensible value comes from connecting AI with knowledge, workflows, feedback, and operating practices that are specific to the company.

Humans remain essential to that process. Nadella argues that people should set goals, guide AI systems, and identify useful patterns so AI is not “running in circles.” Employees supply business context and determine what constitutes a good result. AI can then help execute, retrieve, synthesize, and improve work at greater speed.

This requires companies to define quality in measurable terms. Nadella points to internal evaluations that test AI performance against company-specific benchmarks. A customer-service system, for example, should not be judged only on the fluency of its answers. The company needs measures connected to its own requirements, such as resolution quality, policy compliance, cost, and customer outcomes.

Repeated evaluation can also turn operational experience into reusable institutional knowledge. When a company identifies a workflow that performs better, it can preserve the prompts, context, tools, data, evaluation criteria, and other operating knowledge that contributed to the result. Employees and AI systems can then reuse that knowledge instead of repeatedly reconstructing it.

Nadella describes the result as institutional memory that is “query-able.” This is an important shift for enterprise knowledge management. Valuable knowledge often exists across documents, systems, conversations, and individual employees. AI can make more of that information accessible, provided the organization has reliable source data, appropriate permissions, and controls that preserve provenance.

The feedback process can become increasingly valuable as the organization operates it. Nadella writes: “Every improved workflow generates a better training signal, which accelerates the accumulation of tacit knowledge unique to the firm.” In practical terms, successful interactions provide evidence about which approaches work under the company’s specific operating conditions.

The source article also notes a potential cost benefit. Better institutional memory can allow systems to retrieve relevant knowledge with fewer tokens rather than repeatedly processing large amounts of context. Because model usage is commonly priced partly according to token consumption, reducing unnecessary token use can lower operating costs. The article provides no numerical estimate for these savings.

Executives should treat this claim with appropriate precision. Better workflows do not automatically become model training data, and organizations do not need to retrain a model every time a process improves. The useful feedback can instead inform prompts, retrieval systems, evaluation sets, workflow logic, fine-tuning, or future model development. The right mechanism depends on the task and economics.

Ownership also needs careful definition. A company does not necessarily need to own the underlying foundation model to build proprietary AI capability. Competitive value can reside in company data, evaluation methods, workflow design, integrations, permissions, feedback signals, and accumulated knowledge even when the base model comes from an external provider.

This leads to a more useful C-suite investment model. Human expertise and AI infrastructure should not be managed as separate transformation programs. Companies need workflows in which employees improve AI-supported processes and those improved processes help employees perform better. That requires investment in data quality, evaluation, governance, workflow redesign, and employee capability alongside model access.

Satya Nadella, Chairman and CEO of Microsoft, presents this combination of “human capital” and “token capital” as a continuous learning process. His position supports the broader argument in the article: the strongest enterprise AI strategy does not remove human expertise. It captures that expertise, makes it easier to use, and applies AI where doing so improves measurable business performance.

Concluding thoughts

The core AI challenge is no longer access. It is control. Generative AI makes producing content cheap and fast, but it can also increase verification work, weaken provenance, and spread low-quality information across connected processes.

Executives should therefore stop treating AI adoption as the goal. Measure the complete workflow instead. Track whether AI reduces total cost, cycle time, errors, and rework while improving decision quality. A faster individual task has little value if another team must verify or correct its output.

The strongest AI strategy also protects the scarce asset AI cannot create on demand: company-specific knowledge grounded in real evidence. Preserve source data. Keep important outputs traceable. Define where human judgment remains accountable. Use AI to synthesize and execute rather than allowing generated content to become an unchecked source of truth.

Competitive advantage will not come from giving everyone access to the same models. It will come from combining capable models with proprietary knowledge, disciplined workflows, strong evaluation, and experienced people.

The executive mandate is clear. Automate where the economics work. Preserve human judgment where consequences matter. Measure outcomes rather than AI activity. Companies that get those choices right can capture AI’s productivity gains without allowing faster content production to weaken the knowledge their business depends on.