Business Integration of Generative AI

2.1 Four Levels of Integration

Companies do not adopt GenAI in one way; they adopt it at a level, and the level determines the cost, the complexity, and how defensible any advantage will be. Drawing on practice and on research (e.g., Scott Cook, Andrei Hagiu, and Julian Wright in Harvard Business Review, January-February 2024, "Manage Generative AI by Strategizing at Four Levels"), four levels emerge:

Level 1: Adopt Publicly Available Tools. The entry point is simply using off-the-shelf GenAI tools, public chatbots (e.g., GPT, Claude), image generators (e.g., Midjourney, Gemini nano banana), video editors (e.g., Gemini Veo), or coding assistants, to make internal work go faster: drafting emails, summarizing documents, generating a first creative pass. Complexity and cost are low and the upfront investment is minimal, but customization of the underlying models is essentially nil. Any competitive advantage tends to be temporary, since the same tools quickly become "table stakes" across the industry. The main considerations are privacy and security. Relying on third-party services raises real concerns when sensitive information is involved, and the SAS report (2024, p. 6, 9) finds that 76% of organizations worry about data privacy and 75% about security with GenAI.

Level 2: Customize Existing Tools or Models. At this level, organizations tailor available AI tools or foundation models to their needs, typically by calling APIs from model developers (OpenAI, Anthropic) or fine-tuning pre-trained models on proprietary data. That opens the door to personalized support bots, AI-powered product features, and adaptive interfaces. Complexity and cost rise to moderate levels, since API integration, fine-tuning, and data preparation all require technical expertise; customization becomes high at the application layer even if the underlying model is only partially adapted. Competitive advantage can be significant when the customization rests on unique data or solves a sharply defined customer problem. The key consideration is data governance for fine-tuning material. The Shelf & ViB report (2025, p. 11) notes that 57% of companies fine-tune on their own data to tackle unstructured-data issues, which tells you how mainstream this approach has become.

Level 3: Build Proprietary GenAI-Powered Applications, Automated Workflows, and Agents. Here organizations move past basic customization and build their own applications, integrated AI workflows, and agentic systems, using GenAI as the foundation for automating specific business processes or shipping unique product capabilities. Examples range from document-processing apps and specialized chatbots to expert assistants and end-to-end agent systems that reason, plan, and act autonomously, typically by combining GenAI APIs, prompt engineering, business logic, and workflow automation. Complexity and cost are high because of the development, integration, and maintenance involved, but so is the payoff in fit: solutions are shaped to specific workflows, vertical needs, and existing enterprise software and data. The competitive advantage can be substantial, since well-designed automation and agentic systems are hard for competitors to copy. But success depends on sound governance and security, real change management, careful workflow and prompt design, continuous monitoring, and cross-functional teams that actually bridge business and technology. I have watched more than one Level 3 project stall not on the model but on the org chart, and the pattern is consistent enough to spell out. The pilot works; the demo impresses; and then the project discovers that the process owner, the data owner, the budget owner, and the person accountable for errors are four different people, none of whom has the authority to change the workflow the system is supposed to automate. Six months later the "AI project" is still in the pilot phase, and the model was never the problem. The fix is organizational, not technical: one named owner with the authority to redesign the process, a written decision about which actions the system may take without a human, and an agreed measure of success before integration starts. Chapter 14 turns that fix into a method.

Level 4: Develop Proprietary Models. The most demanding level involves building generative AI models from the ground up, foundation models trained for the firm's own problems, or significantly modified open-source architectures, using internal data and expertise. This path is realistic only for companies with unique data advantages or needs that off-the-shelf and fine-tuned models cannot meet. Complexity and cost are extremely high, demanding serious AI research talent, massive compute, and extensive datasets, but customization reaches its maximum: the firm controls the architecture and training process end to end. If the model delivers a breakthrough capability, the advantage can be both large and durable. Very few companies have the resources to attempt this, success is far from guaranteed, and development cycles are long. Still, the McKinsey report (2025, p. 8, Singla commentary) advises organizations to "think big" and aim for "wholesale transformative change," which for some firms points naturally toward this level.

As McKinsey's 2025 report observes, organizations are still in the "early days" but are actively "redesigning workflows, elevating governance, and mitigating more risks" as they move through these levels (McKinsey, 2025, p. 2).

2.2 Business Models and Value Creation

What does GenAI actually do to a business model? The evidence is starting to accumulate: in workplace trends, in productivity studies, and in how organizations restructure themselves. The McKinsey (2025, p. 2) report title says it plainly: "The state of AI: How organizations are rewiring to capture value."

Perceived Potential and Adoption Trends: Adoption keeps climbing. A note on vintages before the numbers: McKinsey's "State of AI" survey runs in waves, and this section cites the wave fielded in July 2024 (published 2025); Section 2.5 cites the late-2025 wave, in which the same adoption measure had risen further to 88%. Read them as two points on one curve. In the July 2024 wave, 78% of survey respondents reported their organizations use AI in at least one business function, up from 72% in early 2024 and 55% a year prior. For GenAI specifically, 71% of respondents said their organizations regularly use it in at least one business function, a jump from 65% in early 2024 (McKinsey, 2025, p. 17). The SAS report (2024, p. 27) tells a similar story: over half (54%) of businesses have begun to implement GenAI, and 86% invested in it in 2024 or planned to in 2025.

Enterprises are committing significant resources to GenAI across various functions. The Shelf & ViB (2025, p. 3, 6) survey found the highest levels of GenAI commitment (planning, PoC, deployed, or scaling) in software development (87%), data management/BI/analytics (86%), and operations/process automation (83%).

Productivity and Output Quality Gains: The productivity evidence is unusually concrete for such a young technology, though it needs honest reporting. In a controlled experiment, developers completed a coding task 55.8% faster with AI assistance (Peng et al., 2023, "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot"). Consultants completed tasks 25% faster with about 40% higher-rated quality on tasks within the technology's competence (Dell'Acqua et al., 2023, "Navigating the Jagged Technological Frontier"), and professional writing tasks took roughly 40% less time (Noy and Zhang, 2023, "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence"). The gains show up wherever knowledge work involves drafting, synthesis, or iteration, but they are not universal: a 2025 METR randomized trial found that experienced open-source developers working in codebases they knew deeply were actually slower with AI assistance, while believing they had been faster. The lesson for managers is to expect large gains on unfamiliar, drafting-heavy work, smaller or even negative ones where deep individual expertise already dominates, and to measure rather than trust self-reports.

The benefits are not only about speed. The SAS report (2024, p. 4) found that among organizations embracing GenAI, 89% reported improved employee experience and satisfaction, 82% noted savings on operational costs, and 82% stated higher customer retention. Still, the outcome companies target first, according to Shelf & ViB (2025, p. 3, 7), is improving operational efficiency (61% of respondents).

Revenue and Cost Impacts: Organizations are beginning to see tangible financial benefits. McKinsey's latest survey (2025, p. 22-23) indicates an increasing share of respondents reporting value creation, with larger shares than in early 2024 stating their GenAI use cases have increased revenue and led to cost reductions within deploying business units. For instance, in the second half of 2024, 70% of those using GenAI in strategy and corporate finance reported revenue increases, and 61% using it in supply chain and inventory management reported cost decreases.

Accelerated Employee Development: Research by Brynjolfsson, Li, and Raymond (2023, NBER WP 31161, "Generative AI at Work") found that AI assistance raised customer support agents' resolutions per hour by about 14% on average, but the headline hides the interesting part: gains reached roughly 34% for novice agents while barely registering for the most experienced ones. The tool was, in effect, distributing the experts' playbook to everyone else. This suggests GenAI can democratize expertise and accelerate onboarding, and it foreshadows a management question you will meet again in Chapter 13: if the system encodes your best people's know-how, how do the next generation of experts develop?

Transforming Knowledge Management Systems: One dimension of GenAI's value gets overlooked: what it does to institutional knowledge. Every organization sits on a pile of documents, wikis, slide decks, and old email threads that nobody can find when they need them. GenAI can turn those static repositories into systems people actually use in the flow of work. Embed AI into routine activities, team meetings, employee orientation, client interactions, and knowledge stops being something you go look for; it arrives when needed. In practice this looks like a few specific things. GenAI can pull together information from scattered sources to reconstruct institutional memory, and conversational interfaces cut through data silos that formal search never managed to. New employees onboard faster because the system delivers knowledge tailored to what they are doing that week. A support agent on a live call gets the relevant policy in seconds instead of putting the customer on hold. Here is the catch: none of this comes free with the technology. It requires investment in metadata infrastructure, real integration work across systems, and governance woven into daily workflows so people keep trusting what the system tells them. And it is ultimately a cultural change, which means visible executive sponsorship and change agents who help employees turn the new capability into better daily work.

Strategic Rewiring for Value Capture: Getting full value out of GenAI means changing the organization, not just installing the software. McKinsey (2025, p. 2) reports that companies are "redesigning workflows, elevating governance, and mitigating more risks," and 21% of respondents using GenAI said their organizations have fundamentally redesigned at least some workflows (McKinsey, 2025, p. 4). That kind of change rarely happens from the middle. As Alexander Sukharevsky puts it: "Effective AI implementation starts with a fully committed C-suite and, ideally, an engaged board" (McKinsey, 2025, p. 4).

2.3 Challenges and Considerations in GenAI Integration

The potential is real, but so are the failure modes. In my consulting work the same five problems come up again and again: data quality, governance, strategic alignment, technical integration, and talent.

1. Data Quality and Unstructured Data Management. "Garbage in, garbage out" applies with full force: the quality of the data used to train and prompt these models directly determines the quality and reliability of their outputs. The Shelf & ViB (2025) survey shows the scale of the problem. Fully 85% of organizations manage over a million documents and files (51% handle more than ten million), and 92% report that unstructured data issues have impacted their GenAI initiatives, with 30% describing the impact as large or significant. The rot runs deep within each estate, too: 68% say more than half of their files have at least one issue, and 42% report that over 70% of their documents have an issue capable of hindering GenAI success. The most common problems are duplicate files and multiple versions (66%), out-of-date information (53%), and conflicting versions (47%). SharePoint is the primary source of unstructured data feeding these initiatives (67%), followed by email (46%) and Microsoft OneDrive (45%), ordinary enterprise systems that were never designed with GenAI in mind. While 74% plan to use unstructured data despite the issues, 55% intend to remediate over the next 12 to 24 months, typically by fine-tuning on existing data (57%) and by adding new data management and quality solutions (48%). As the SAS report (2024) notes, "data management and analytics tools can detect outliers and sources of bias in the raw data used to feed LLMs."

2. Governance, Risk, and Compliance (GRC). The risks are well known by now: bias, hallucinated misinformation, IP infringement, security vulnerabilities. Preparedness is another matter. SAS (2024 fieldwork, before the EU AI Act's obligations began to bite) found that only one in ten organizations had done the work needed to comply with GenAI regulations, and a striking 95% lacked a comprehensive governance framework; Chapter 13 covers what has changed since. Data privacy (76%) and security (75%) top the list of concerns, and McKinsey's survey (2025, p. 7) shows mitigation efforts intensifying for inaccuracy, IP infringement, and privacy. Monitoring is the weakest link: only 5% of organizations have a reliable system for measuring bias and privacy risk in LLMs, 71% cannot continuously monitor their GenAI systems, and McKinsey (2025) found that only 27% of organizations are mitigating accuracy risks for all relevant GenAI use cases. Leadership matters here. CEO oversight of AI governance correlates with higher bottom-line impact, though only 28% of organizations using AI say the CEO is responsible. Structurally, risk, compliance, and data governance for AI tend to be centralized, while tech talent and adoption of AI solutions are typically managed in a hybrid model (McKinsey, 2025, p. 5).

3. Strategic Alignment and Organizational Understanding. Effective deployment needs a clear strategy and a shared understanding of the technology. Neither is widespread. SAS (2024) found that 93% of senior tech decision-makers admit they do not fully understand GenAI or its potential impact on business processes. Sit with that number for a moment: the people making the investment decisions mostly do not understand what they are buying. Two-thirds of companies (66%) lack a standard process for prioritizing GenAI use cases (Shelf & ViB, 2025), and fewer than one in three organizations follow most of the 12 key adoption and scaling best practices identified by McKinsey (2025). Even basic guardrails are missing in many places: 39% of organizations have no GenAI usage policy for their staff (SAS, 2024).

4. Technological Integration and Tools. Wiring GenAI into existing systems and workflows is hard, unglamorous work. Nearly half (47%) of decision-makers say they lack the right tools to implement GenAI, and 41% report compatibility issues when combining GenAI with current systems (SAS, 2024). More than half (52%) hit obstacles using public and proprietary datasets effectively, and over a third (34%) name technological limitations as the biggest challenge to monitoring GenAI in production.

5. Talent and Skills. Demand for GenAI-proficient professionals still outstrips supply, though the picture is shifting. Half of organizations (51%) worry that they lack the in-house skills to use GenAI effectively, and 39% identify insufficient internal expertise as an active obstacle to implementation (SAS, 2024). McKinsey (2025) reports that companies are hiring for new risk-related roles such as AI compliance and AI ethics specialists; while hiring difficulty has eased for many AI roles, AI data scientists remain in particularly high demand. To close the gap from within, organizations are leaning on reskilling, with many expecting to do more AI-related reskilling in the next three years than they did in the past one.

2.4 The Future of Work with Generative AI

Generative AI is changing how work gets done, and with it how people think about their careers. Cheap access to powerful tools lowers the barriers to starting and scaling a business, fueling a rise in solopreneurship and small, nimble firms. Inside larger organizations, the working pattern is shifting toward human-AI collaboration: "copilot" scenarios where AI assists with drafting, research, coding, and analysis to augment rather than replace people. McKinsey (2025, p. 19-20) data shows C-level executives leading the charge in personal GenAI use, potentially modeling this collaborative approach for the rest of the workforce. Closely related is the rise of prompt engineering as a transferable skill. The ability to communicate with GenAI through well-crafted prompts is becoming a baseline competency across many roles.

Beyond the individual workstation, GenAI is enabling deeper workflow automation at both the company and individual levels, moving past simple task automation into complex multi-step processes. The next step is autonomous AI agents that understand goals, plan, and execute independently: agents that can manage projects, conduct research, or run parts of a business with minimal human oversight. The workforce impact will be uneven. McKinsey (2025) finds that while most respondents expect little immediate change in headcount, declines are anticipated in service operations and supply chain/inventory management, with growth in IT and product or service development. Notably, when headcount reductions do occur as a result of GenAI, they are among the organizational attributes most strongly associated with bottom-line value realized. Skill demands will keep shifting in tandem. As Lareina Yee observes, "the difficulty of finding AI talent, while still considerable, is beginning to ease... the long-term workforce effects are still only beginning to take shape" (McKinsey, 2025). Continuous learning stops being optional.

As Michael Chui concludes in the McKinsey report (2025), "AI only makes an impact in the real world when enterprises adapt to the new capabilities that these technologies enable." The adaptation, not the technology, is the hard part.

2.5 The Adoption Paradox: What the Field Data Shows

Before we move on to the practical chapters, it is worth pausing on an uncomfortable set of numbers. By late 2025, 88 percent of organizations surveyed by McKinsey reported regular AI use in at least one business function, up from 78 percent in the mid-2024 wave cited earlier in this chapter. Yet only 39 percent could point to any effect on earnings, and most of those attributed less than 5 percent of EBIT to AI (McKinsey, State of AI, late-2025 wave). McKinsey's consultants gave this gap a name: the gen AI paradox. Nearly everyone is using the technology. Almost no one can find it in the profit and loss statement.

Why? Part of the answer lies in what companies chose to deploy first. Horizontal tools such as enterprise copilots and chat assistants spread quickly precisely because they demand so little change; roughly 70 percent of Fortune 500 companies had adopted Microsoft 365 Copilot by mid 2025, by Microsoft's own account. Their benefits, however, arrive as small time savings scattered across thousands of employees, which makes them nearly invisible in financial statements. The function-specific applications that could move earnings behave in the opposite way: McKinsey estimates that around 90 percent of these vertical use cases never made it out of the pilot phase (McKinsey, 2025).

Researchers writing in Harvard Business Review have described the same phenomenon from the inside of firms. One team calls it the micro-productivity trap: organizations accumulate isolated, individual-level improvements that never aggregate into measurable business outcomes (Dutt et al., 2026). Another, drawing on experiments at Siemens and Procter & Gamble, documents a productivity J-curve, an initial dip in output after adoption while people and processes adjust, followed by gains only for the organizations that persist through the awkward phase with disciplined experimentation (Berndt et al., 2026).

The pattern shows up at the level of individual workers too. BCG's annual AI at Work survey found in 2025 that regular use among leaders and managers had passed 75 percent while frontline usage was stuck near 51 percent, a gap the firm called the silicon ceiling. A year later that ceiling had cracked: frontline daily use jumped to 74 percent, and 42 percent of regular frontline users reported saving eight or more hours per week, a full working day (BCG, 2026). Here is the uncomfortable part. Two-thirds of those employees said they received little or no guidance on what to do with the time they saved, and more than half admitted they did not reinvest it in higher-value work. The productivity was real. The value evaporated.

What separates the companies that capture value from those that merely adopt tools? The evidence converges on a simple, hard answer: work redesign. McKinsey found that high performers were nearly three times more likely to fundamentally redesign workflows rather than bolt AI onto existing processes (McKinsey, 2025). BCG reached the same conclusion from a different direction with its 10-20-70 principle: in successful AI programs, roughly 10 percent of the effort goes to algorithms, 20 percent to data and technology, and 70 percent to people, processes, and cultural change (BCG, 2025). BCG's data also shows that winners concentrate: leading companies pursued an average of 3.5 prioritized use cases against 6.1 for everyone else, and anticipated more than twice the return.

A case from MIT Sloan Management Review makes the human side concrete. Novo Nordisk scaled Microsoft Copilot from a few hundred users in January 2024 to 20,000 employees by February 2025, measuring average savings of 2.17 hours per employee per week (Wade et al., 2025). The finding that stayed with me is not the hours. Employee satisfaction correlated about three times more strongly with perceived improvements in the quality of their work than with time saved. People did not love the tool because it made them faster. They loved it because it made their output better.

One more finding deserves attention because it says something about where the energy for adoption actually sits. McKinsey's Superagency research, based on surveys of over 3,600 employees and 238 senior executives, found that C-suite leaders underestimated their own employees' AI use by a factor of three, and that 47 percent of executives believed their companies were developing AI too slowly. Employees ranked training as the single most important factor for adoption, and nearly half reported receiving little or none (McKinsey, 2025). The report's conclusion was blunt: the workforce is ready, and the biggest barrier is leadership. Keep that in mind as you read the chapters that follow. The bottleneck in most organizations is not the technology, and it is usually not the people using it.