Agentic AI is moving from concept to day-to-day reality, but enterprise adoption is not keeping pace with innovation.
In a GLG virtual event, AI leadership expert Andreas Welsch explained why the gap between what is possible and what organizations can absorb is widening—creating new demands for governance, enablement, and measurable business outcomes.
His core message: scaling agentic AI requires treating agents as software with human accountability, building guardrails that do not stifle innovation, and focusing on workforce enablement to avoid low-quality “AI workslop.”
The discussion also highlighted executive-level decision points: tool standardization versus capability, business pull versus IT resistance, board-level success measures, and why KPIs—not token usage—should define value.
Executive Summary
- Innovation is accelerating; enterprise absorption is slower, widening the adoption gap.
- Shadow AI grows when governance blocks innovation instead of guiding it.
- Tool choice matters less than workforce enablement and responsible use.
- Scale requires business-aligned KPIs, not “token maxing” usage metrics.
- Credibility risks rise when leaders outsource judgment and tone to AI.
Key Takeaways
- Welsch observed that hands-on use is now essential because AI cycles are shortening.
- He warned that shadow AI increases when employees experiment outside governance boundaries.
- He emphasized “dual lens” thinking: agents are software, and humans remain responsible.
- He recommended borrowing familiar lifecycle concepts (role, access, oversight) to govern agents.
- He argued most vendors “cook with water,” making enablement more critical than constant tool switching.
- He cautioned that low-quality AI output can damage leadership credibility and trust.
- He advised measuring success through business and process KPIs—not license counts or token usage.
What is Agentic AI?
Agentic AI refers to AI systems designed to operate with greater autonomy than a typical chatbot—handling multi-step tasks, using tools, accessing information, and producing outputs (or taking actions) with less constant prompting. In the conversation, Welsch described agents as “always on” software that can be responsive and available “at your fingertips,” while still requiring human oversight. He stressed that organizations should treat agents as software components—useful for scaling work—while maintaining human responsibility for decisions, quality, and accountability.
Why the Pace Shift Matters for Agentic AI Adoption
Welsch noted that leaders could previously stay broadly aware of AI trends, but the pace has “accelerated significantly” since the beginning of the year.
He pointed to the emergence of more structured agent tooling as an inflection point—moving beyond one-off prompting toward systems that stay available and do “more meaningful work.” In his view, this makes hands-on experimentation a requirement, not a nice-to-have.
Key Insight: Welsch argued that keeping up now requires practical usage. Only hands-on work reveals where agentic AI performs well, where judgment must intervene, and what leaders still need to understand to stay in control as cycles shorten.
The Widening Gap: Exponential Innovation vs. Logarithmic Absorption
According to Welsch, the industry is experiencing an exponential increase in AI innovation, announcements, and model updates—while organizations remain constrained by slower operating rhythms.
That mismatch creates a growing gap between what is available and what can be responsibly implemented. Businesses must evaluate applicability, run tests, estimate cost, manage change, and address privacy and cybersecurity—each of which adds necessary friction.
This dynamic can intensify executive anxiety: competitive pressure rises as vendors set the pace, while internal constraints prevent lockstep execution. Welsch framed the challenge as sense-making, not merely technology procurement.
Key Insight: Welsch described the core enterprise problem as “sense-making.” Even when technology exists, organizations struggle to extract, clean, contextualize, and apply data to business decisions and workflows at scale.
AI Governance Without Killing Innovation: Managing Shadow AI
Welsch described a familiar pattern: executives declare an “AI-first” push, and employees respond by experimenting—often outside official tooling and governance. He compared this “shadow AI” to shadow IT.
The risk is not experimentation itself, but the leakage of confidential data, unclear security practices, and inconsistent quality standards. In his view, governance must set boundaries while keeping entry barriers low enough to encourage learning and iteration.
He cautioned against heavy approval processes that take months only to deny access. That approach, he suggested, fails from the start by pushing employees to work around the system.
Key Insight: Welsch emphasized “guardrails within which you are allowed to use it” while staying open enough to foster innovation. Overly rigid gates can accelerate shadow AI and increase risk instead of reducing it.
Change Management: From Rollout to Real Adoption
When asked about AI change management, Welsch pointed to the need for coordinated leadership across functions—HR, legal, finance, operations, and IT—rather than isolated experiments.
He described a common friction point: organizations roll out an AI tool (for example, an enterprise assistant) without sufficient training, leaving employees unsure how to use it well and leaders disappointed with outcomes.
Welsch linked adoption quality to credibility and trust. He highlighted a surge of AI-generated content—meeting transcripts, action items, and follow-up emails—that can be fast but shallow, incomplete, or inauthentic if not reviewed and improved by humans.
In his assessment, leaders must empower teams to use AI while also maintaining accountability for high-quality outputs.
Tool Sprawl vs. Capability: Why Enablement Wins
The conversation addressed a frequent executive concern: too many tools, too many updates, too much noise.
Welsch’s view leaned toward enablement over constant tooling churn. He argued that many vendors are drawing from similar underlying model capabilities, making the differentiator less about novelty and more about integration, enterprise readiness, and training.
He contrasted integrated enterprise ecosystems (such as productivity suites already embedded in daily workflows) with cutting-edge standalone tools that may innovate faster but require more integration and operational change.
His caution: organizations cannot chase every “tool is dead” headline because switching costs are real. Pick a platform, invest in learning, and migrate thoughtfully when sustained market signals justify the move.
Key Insight: Welsch warned that “AI workslop” creates hidden costs: managers spend time reviewing low-quality AI drafts. In his view, foundational training—what AI is, risks, and how to use it well—reduces rework and protects output quality.
Governance Model for Agents: Treat Them Like Digital Workers
Welsch offered a pragmatic way to frame agent governance: apply familiar people-management concepts to AI agents.
He described a “dual lens principle.” First, agents are still software and humans remain accountable for results. Second, organizations can borrow from the employee lifecycle to define and manage agents: role definition, access to data, behavioral guidelines (tone, truthfulness), interaction rules with other agents, and ongoing oversight.
He also highlighted an operational risk: if an employee builds high-value agents that are tied to the employee’s account, productivity can drop if that employee leaves and the agent is switched off. That makes ownership, continuity, and lifecycle management essential.
In the current environment, agents may be built by individuals, by business functions with IT, or delivered through major software vendors. Welsch emphasized that this fragmentation makes centralized visibility harder, increasing the need for orchestration and control layers.
Business Pull vs. IT Pushback: How Leaders Close the Gap
When asked how to respond to IT resistance, Welsch framed the tension as a classic business-versus-risk tradeoff: business functions push for efficiency and competitive speed, while IT emphasizes cost, security, and data privacy.
If alignment cannot be reached at the peer level, he advised elevating the issue to the next leadership level to establish shared priorities. The goal is not forcing adoption, but ensuring innovation happens in ways that meet organizational standards without becoming a reflexive “red card” against change.
Welsch also emphasized that successful AI programs are cross-functional by design. Central leadership—such as a chief AI officer role that can coordinate across HR, legal, finance, and IT—helps reduce fragmentation and conflicting mandates.
Moving Beyond “Pilot Purgatory” with Measurable Outcomes
Welsch acknowledged the value of experimentation—especially in early stages—but argued that scale happens only when AI is tied to meaningful business problems.
He recommended aligning AI initiatives with KPIs and process performance indicators that business leaders already own. Examples raised in the discussion included invoice cycle time and cash collection timing, as well as customer support automation for common questions with escalation paths to live agents.
He also urged gradual scaling rather than “big bang” rollouts—expanding from one region or subsidiary to the next while learning what breaks and what generalizes.
On measurement, Welsch cautioned against “token maxing” leaderboards and other usage-only metrics. Token consumption reflects spend, not value. AI success should be assessed through business impact indicators—revenue lift, productivity improvements tied to process metrics, and measurable customer outcomes.
Workforce Enablement: A Cascaded Model for Adoption
Welsch outlined a pyramid-style enablement approach, starting with senior leadership and cascading through the organization.
Senior executives need enough hands-on familiarity to sponsor change and support middle managers. Middle management then drives day-to-day adoption with teams. Champions or “multipliers” inside functions translate AI capabilities into functional workflows. Builders (for example, technical teams) require deeper tool-specific guidance. Finally, everyday users need practical training on responsible use and quality standards.
Welsch also described a credibility risk: employees can often tell when leaders rely too heavily on generic AI-generated communication. In environments already marked by job anxiety, inauthentic AI-driven messages can erode trust at the exact moment leadership needs it most.
Key Insight: Welsch argued that leaders should not outsource judgment. AI can help tighten writing and clarity, but feedback, high-stakes decisions, and trust-building communication must carry authentic leadership intent—or credibility can decline.
Leadership Implications
- Set “lightweight guardrails” for Agentic AI: define boundaries for data use, privacy, and acceptable tools without blocking experimentation.
- Anchor scale to business KPIs: select use cases tied to measurable process outcomes (cycle time, cash collection, customer resolution rates).
- Invest in workforce enablement: train leaders, managers, champions, builders, and users with role-specific guidance on quality and responsible use.
- Clarify ownership and lifecycle management for agents: document who built which agents, what data access exists, and continuity plans if employees leave.
- Protect leadership credibility: require review and human judgment for sensitive communications and decision-making, avoiding low-effort AI-generated outputs.
Why this conversation matters
This discussion took place in a GLG virtual event focused on making agentic AI work at scale, aimed at business and technology leaders navigating rapid change.
The conversation matters because it connects technical momentum to leadership reality: governance, cross-functional alignment, workforce transformation, and measurement discipline. Welsch framed these as the practical constraints that determine whether agentic AI becomes a competitive advantage or a source of risk and rework.
It also connects to Welsch’s broader emphasis on AI leadership and adoption quality—particularly the need to enable employees to use AI well, not merely to use it often.
Conclusion
Scaling Agentic AI is less about chasing every model update and more about disciplined execution: governance that guides rather than blocks, enablement that lifts quality, and measurement that ties adoption to business outcomes.
Welsch’s perspective emphasizes a practical path forward—treat agents as software, hold humans accountable, and focus on workforce transformation so that agentic AI becomes a compounding advantage rather than a source of “AI workslop.”
FAQ
How can an enterprise scale Agentic AI without increasing risk?
Scaling Agentic AI requires lightweight governance guardrails that allow experimentation while protecting data and security. Welsch emphasized boundaries for acceptable use, clear accountability, and oversight so innovation can proceed without pushing employees into shadow AI behavior.
Practical steps include defining permitted tools, data-sharing rules, and ownership of agents across their lifecycle.
What should boards use to measure AI success versus failure?
Boards should measure AI success through business outcomes tied to strategy, not vague “do AI” mandates. Welsch argued that measurable objectives—such as revenue growth, market outcomes, or process KPI improvements—make it easier to assess progress and accountability.
This approach links AI programs to enterprise priorities rather than technology activity.
Is tool selection or upskilling more important for AI adoption?
Upskilling and enablement are often more important than constantly switching tools. Welsch suggested many vendors share similar underlying capabilities, making training, integration fit, and responsible usage practices the true determinants of adoption quality and value creation.
Tool choices still matter, but workforce readiness determines whether value compounds.
How should leaders prevent “AI workslop” from hurting productivity?
Leaders can reduce AI workslop by training teams to use AI well and setting quality expectations for outputs. Welsch described inbox overload from cheap content generation and warned that low-quality drafts shift time costs to reviewers and erode trust.
Quality gates, review norms, and prompting education help keep AI outputs useful and credible.
How should organizations handle business demand and IT resistance to AI?
Organizations should align business urgency with IT’s security and privacy responsibilities through shared leadership priorities. Welsch recommended elevating unresolved tensions to the next leadership level to ensure innovation continues while still conforming to enterprise standards.
The goal is joint ownership, not IT acting as a permanent brake.
What is the best way to move beyond AI pilot purgatory?
Moving beyond pilot purgatory requires selecting problems worth solving and tying them to KPIs that business owners care about. Welsch emphasized business alignment, representative data, and gradual scaling across regions or subsidiaries instead of a risky “big bang” rollout.
When stakeholders see measurable return, adoption tends to spread faster.
How should AI ROI be measured at the C-level?
C-level leaders should measure AI ROI using business and process performance indicators, not usage metrics like tokens. Welsch criticized “token maxing” as a poor proxy for productivity and advised linking AI to measurable cycle time, cash, or customer outcomes.
Licenses and logins show adoption activity, but not business impact.
Why does human accountability still matter with Agentic AI?
Human accountability matters because agentic systems are still software and outputs can be inconsistent or wrong. Welsch noted that these are probabilistic systems, so organizations must build trust through oversight, review, and clear decision ownership—especially under pressure.
Human-in-the-loop controls may be essential until performance and reliability are demonstrated over time.
How can leaders maintain credibility while being “AI-first”?
Leaders maintain credibility by ensuring important communication reflects authentic judgment, not generic AI output. Welsch warned employees can often detect AI-generated messages, and that perceived low-effort feedback can erode trust during periods of job anxiety and change.
AI can assist drafting, but leadership intent and review should remain visibly human in high-stakes contexts.