AI Strategy for Generative AI

Generative AI adoption is accelerating—but an effective AI strategy must start with a leadership reality: the models rely on human-created and human-curated data, and that creates risk, cost, and complexity.

In a conversation on the Innovation Rockstars podcast, Andreas Welsch—an AI leadership expert focused on AI governance, strategy, adoption, and workforce transformation—explains why “generative AI has a growing problem, and that is humans,” and what business leaders should do about it.

The discussion spans practical enterprise starting points, how to evaluate use cases beyond hype, why AI centers of excellence can succeed (or fail), and how agentic AI could reshape workflows—from travel booking to procurement negotiation.

Executive Summary

• Generative AI inherits human data quality, bias, and curation trade-offs.
• Businesses should target measurable problems, not “AI for AI’s sake.”
• Feasibility still depends on data availability and domain context.
• Scalability must be designed in early to avoid permanent “interim” solutions.
• AI centers of excellence can align stakeholders and accelerate adoption.

Key Takeaways

• Human-produced internet data is both the fuel and the failure mode for generative AI.
• Content filtering and review are necessary—and can be mentally taxing for reviewers.
• Using generative AI to “review” AI training data can reintroduce the same errors.
• Bias shows up twice: in raw data and in human curation decisions.
• Enterprises should start with business problems and align stakeholders early.
• Desirability, feasibility, value, and scalability must be considered together.
• Agentic AI points to a future of goal-based delegation and multi-agent negotiation.

What is AI strategy?

In this context, AI strategy is an enterprise approach to applying AI—especially generative AI—to real business problems with clear ownership, measurable outcomes, and responsible controls. Andreas Welsch emphasizes aligning stakeholders on what matters, assessing whether data and resources exist to deliver a solution, and designing for rollout beyond a lab or pilot. The goal is not experimentation alone, but safe deployment that improves tangible KPIs while preparing the workforce for new AI-augmented ways of working.

Why generative AI’s “growing problem” is humans

Welsch’s core point is structural: large language models depend on data created by humans and curated by humans. That data is uneven—drawn from “the pretty and rosy places of the internet,” but also from corners where opinions and values may be extreme or unrepresentative.

Because training data is messy, organizations rely on humans to filter and review content. That work is labor-intensive and can be psychologically difficult, especially when reviewers must assess disturbing text or images and decide what is permissible.

Key Insight: Generative AI is not only a model problem—it is a human system problem. The same humans who generate imperfect data must also curate it, and that curation introduces additional bias and operational strain. A leadership-grade AI strategy must plan for this hidden work.

When human review loops create a vicious cycle

Welsch highlights an additional failure mode: task workers who label or review data may use generative AI to speed up their work. In one example, reviewers draw bounding boxes around car damage images to train or support insurance models.

The paradox is clear: humans are involved because models are not perfect. If humans use generative AI to do the review, the same errors and faults can be reintroduced into the pipeline—creating a “vicious cycle.”

Key Insight: Human-in-the-loop processes can silently become AI-in-the-loop processes. When reviewers rely on generative tools for speed, data quality and safety controls may weaken—exactly where enterprises assume human judgment is protecting them.

Bias, diversity, and the governance trade-off

The conversation underscores a leadership dilemma: adding human checks can reduce harmful outputs, but it can also introduce bias through the lens of reviewers and policy decisions. Welsch notes this “multi-layer problem” of bias—at the raw data level and at the curation level.

He also points to a practical example from image generation. When prompted to “create a portrait of a business leader,” results skewed toward “male caucasian and of a certain age.” Some vendors compensate by forcing more diverse outputs—but that can become problematic in historically specific contexts where diversity would be factually inaccurate.

Key Insight: AI governance is not only about preventing harm; it is also about choosing trade-offs. Efforts to counterbalance biased training data can collide with expectations for factual accuracy, creating reputational risk either way. Leaders need explicit policies for these edge cases.

Is generative AI overhyped—or finally accessible?

Welsch argues that the recent shift is less about novelty and more about accessibility. Before ChatGPT’s release in late 2022, analysts discussed another “AI winter.” Afterward, generative AI made outcomes available to non-specialists—without requiring a PhD or deep training in statistics and optimization.

That accessibility changes enterprise dynamics. Leaders across functions have tried the tools or are asked for an AI strategy, raising the urgency to move beyond prolonged proof-of-concept cycles.

From Welsch’s perspective, last year was heavy on experimentation and piloting. This year, enterprises are increasingly implementing and rolling out use cases—especially those that are easy to start, such as summarization, translation, and text generation.

Why businesses should care: productivity, documents, and better drafts

Welsch connects business value to practical pain points. Many employees dislike writing proposals or responding to emails. Generative AI can accelerate those tasks, producing a “good enough draft” in seconds that humans can refine.

He also emphasizes how text- and document-heavy businesses are. He cites an estimate (attributed in the discussion to Everest Group) that 80–90% of business data is stored in documents—requiring humans to open, read, and extract information. Generative AI can reduce that friction with rapid summarization and faster access to “the gist.”

Examples referenced in the conversation include HR workflows (resume creation and candidate identification) and commerce scenarios like generating product descriptions using internal context.

Where to start an AI strategy: begin with the business problem

Welsch advises organizations to resist starting with solution workshops alone. The best starting point is identifying the business problem worth solving and then determining where generative AI fits.

Otherwise, organizations risk pursuing vague outcomes such as “happier employees,” which may be positive but can be difficult to tie to business KPIs or monetizable value.

He also points to the practical requirement of baseline literacy: teams must understand what generative AI can and cannot do. However, he notes that in many audiences, a large majority have already used tools like ChatGPT, lowering the barrier for internal enablement.

Why business and technical leaders must co-design use cases

Welsch emphasizes that strong outcomes come from combining business expertise with technical capability. Business experts understand recurring pain points; technologists bring the toolbox of what is feasible today.

This joint design approach helps teams move faster from idea to deployment—while keeping safety in mind.

Desirability, feasibility, value, scalability: treated as equally important

When asked how to get to ideas that are desirable, feasible, viable (valuable), and scalable, Welsch describes why each dimension matters—and why neglecting any one can derail delivery.

Desirability is stakeholder alignment: unless business, IT, and data leaders agree the problem is real and worth solving, the initiative risks losing buy-in, budget, or leadership sponsorship.

Feasibility still comes back to data. Welsch notes that the data question “is not going away” because generic models often need augmentation with company context (policies, domain knowledge, and internal information).

Value addresses whether the organization can solve the problem within constraints—and whether it remains worth solving even with more resources.

Scalability is where many projects fail: a solution may work in a lab or pilot, but breaks when rolled out globally across regions, languages, or operating models. Welsch warns against “interim solutions” that remain in place long after their creators have moved on.

AI centers of excellence: what they are and why they help AI adoption

Welsch describes AI centers of excellence (CoEs) as an organizational nucleus that helps a business become more capable with AI. Their role includes setting guidelines, standards, processes, and technologies; enabling early pilots and proofs of concept; and supporting safe deployment and scaling.

CoEs are especially valuable when there is uncertainty, distributed ownership, and a need for coordination across multiple business functions.

How CoEs should source and validate use cases

In Welsch’s view, one key success factor is building relationships with business stakeholders. He recommends creating a community of “multipliers”—subject matter experts across the organization who are trained on AI basics and act as listening posts for problems and opportunities.

The mutual benefit is clear: CoEs gain insight and adoption pathways; multipliers gain credibility as innovators within their business unit.

Welsch also acknowledges a common enterprise risk shared with innovation labs: CoEs can fail if they operate in silos, chase technology for technology’s sake, or build solutions disconnected from today’s business metrics. Tangible problems, tangible technology, and near-term value help CoEs “earn the right to sit at the table.”

Workforce transformation: the most important skills go beyond prompt engineering

Welsch identifies three skills that are universal for human collaboration—and become even more critical with generative AI.

• Asking better questions: moving beyond yes/no prompts toward targeted, open-ended inquiries.
• Giving better instructions: clear, tangible direction that reduces rework—whether delegating to people or AI assistants.
• Adopting a Socratic mindset: using follow-up questions and dialogue to explore perspectives and improve outputs.

He also notes a mindset shift for early-career professionals: rather than seeing AI as replacement, focus on how it removes low-value work (such as emails and proposals) and enables faster accomplishment.

Agentic AI: what changes when AI moves from answers to actions

Looking ahead, Welsch highlights the emerging direction of “agentic AI”: systems that can be given a task or goal and, within defined boundaries, decide next steps and execute across tools.

He describes travel booking as an illustrative scenario. Rather than navigating multiple systems, an AI agent could use known preferences (e.g., seat class based on flight length, car size) and complete bookings through integrations.

He also points to a longer-standing concept: multi-agent systems. In procurement, one agent could represent the buyer’s company and policies, while another represents the supplier’s negotiation rules, potentially negotiating within constraints before a human reviews the outcome.

Leadership Implications

• Govern the data loop, not only the model. Define controls for training data review, labeling, and content filtering workflows.
• Design for scalability from day one. Anticipate rollout across regions, languages, and operating models—even if deployment starts small.
• Align stakeholders early. Validate desirability across business, IT, and data leaders to protect funding and sponsorship.
• Invest in workforce enablement. Build capability around questioning, instruction-giving, and iterative dialogue—not only prompting.
• Use a CoE to standardize and accelerate. Set policies, tooling guidance, and a repeatable path from pilot to production.

Why this conversation matters

This podcast discussion is aimed at business leaders navigating the gap between generative AI excitement and enterprise execution. It frames generative AI as a leadership and workforce transformation topic—not just a technology upgrade.

Welsch’s perspective connects AI leadership, AI governance, and AI adoption to operational realities: biased inputs, curation trade-offs, human-in-the-loop fragility, and the practical need to deliver measurable value. It also highlights how AI centers of excellence and skill development can help organizations move from experimentation to scalable impact.

Conclusion

Generative AI can deliver speed and productivity, but Welsch’s message is that the hard part is human: training data, bias, curation, review loops, and organizational adoption. An effective AI strategy anchors on real business problems, designs for scalability, and enables people with the skills and structures to apply AI responsibly.

AI centers of excellence, stakeholder alignment, and thoughtful preparation for agentic AI are not optional extras—they are practical mechanisms for turning experimentation into durable enterprise capability.

About the Author

FAQ

Why does Andreas Welsch say generative AI’s growing problem is humans?

Generative AI depends on human-created and human-curated data, so human bias, poor data quality, and inconsistent review standards propagate into model behavior. Welsch emphasizes that content filtering and labeling are labor-intensive and can still reintroduce errors into the system.

What should an enterprise AI strategy prioritize first?

An enterprise AI strategy should prioritize clearly defined business problems tied to measurable KPIs, not tools or hype. Welsch recommends identifying where the organization has a real need, then assessing how generative AI can help, safely and pragmatically.

Is generative AI overhyped for business leaders?

Generative AI is highly visible, but Welsch highlights a genuine shift: accessibility. Employees and executives can now use powerful AI capabilities without deep technical training, which accelerates adoption pressure and shortens the runway for long proof-of-concept cycles.

What are “low-hanging fruit” generative AI use cases in enterprises?

Low-hanging fruit typically includes text generation, summarization, and translation—tasks common across functions and easy to pilot. Welsch notes value in drafting emails and proposals, summarizing meetings, and extracting the gist from document-heavy business information.

Why do AI projects fail when moving from pilot to scale?

Many AI projects work in a lab but struggle in global rollout across countries, regions, and languages. Welsch stresses that scalability must be considered early; otherwise organizations accumulate “interim” solutions that persist, become fragile, and are difficult to retire.

What is an AI center of excellence (CoE) and what does it do?

An AI center of excellence is an organizational nucleus that defines standards, guidelines, tooling, and processes for AI adoption. Welsch describes CoEs as enablers of pilots, production deployment, and scaling—while helping the business understand AI’s potential and constraints.

How should AI centers of excellence gather enterprise use cases?

CoEs should build relationships with business functions and create a community of multipliers who surface problems and opportunities. Welsch recommends using these subject matter experts as innovation outposts, jointly vetting ideas for desirability, feasibility, value, and scalability.

What skills matter most for generative AI upskilling beyond prompt engineering?

Welsch points to three skills: asking better questions, giving better instructions, and adopting a Socratic back-and-forth mindset. These capabilities improve collaboration with humans and AI assistants, reduce rework, and help teams refine outputs through iterative dialogue.

What is agentic AI, and why does it matter for workflow design?

Agentic AI refers to systems that can be given a goal and decide next steps within defined boundaries. Welsch describes scenarios like travel booking and procurement, where agents may execute actions across tools and even negotiate, with humans reviewing outcomes as needed.