By Vanguard Enterprise Intelligence Unit
What our work with Meridian Group reveals about governance, data quality, and change fitness in AI transformation.

By early 2026, the executive team at Meridian Group had what many companies claimed to want: a broad portfolio of artificial intelligence pilots, enthusiastic business-unit leaders, active vendor relationships, and a board that believed AI would become central to the firm’s next phase of growth.

Our work with Meridian began at a moment when the company appeared, at least from the outside, to be ahead of the curve. It had not ignored AI. It had not waited for perfect certainty. It had encouraged experimentation across the enterprise and given business units enough freedom to discover use cases in real operating environments. For many executives, that alone would have looked like progress.

Meridian operated in a complex environment. Its businesses included logistics, industrial services, customer support operations, and regulated financial workflows. The company had valuable data, thousands of employees performing repetitive knowledge work, and multiple functions under pressure to reduce cost while improving speed. On paper, AI looked like an obvious answer.

The first year seemed promising. Sales teams used generative AI to prepare account briefs. Customer-service managers tested tools that summarized cases and drafted responses. Operations leaders experimented with predictive models for routing and delay management. Finance teams used AI to detect anomalies in reconciliations. IT launched early agentic workflows that could retrieve information, update systems, and draft internal reports.

The activity was impressive. The results were not.

When we reviewed Meridian’s AI portfolio, the gap between motion and value was clear. The company had more than forty AI pilots running across the enterprise, but only six had moved into sustained production. Several produced local productivity gains, but none had materially changed the economics of a major business process. Some tools were used heavily for a few weeks and then quietly abandoned. Others produced impressive demos but could not be integrated into legacy systems. A few created uncomfortable questions from compliance, legal, and cybersecurity teams. The board began asking a sharper question: why was a company spending heavily on AI still struggling to show enterprise-level value?

Meridian’s problem was not unusual. It had confused experimentation with scaling. It had treated AI as a portfolio of tools rather than as an enterprise capability. It had encouraged adoption before building the management system required to absorb adoption. Most importantly, it had underestimated two forces that now separate winners from stalled organizations: governance and change fitness.

The Pilot Plateau

Meridian’s early AI program followed a familiar path. The company began with speed. Leaders wanted quick wins, proof points, and internal excitement. Each business unit was encouraged to identify use cases and test tools. The central technology team provided support, but the company deliberately avoided heavy governance at the beginning because executives did not want to slow experimentation.

In our assessment, that decision was understandable. Employees needed permission to learn. Managers needed exposure to the technology. Business leaders needed to see what AI could do inside real workflows rather than in conference presentations. The company’s initial openness helped create energy.

But by the second year, the same openness created fragmentation. Different teams selected different vendors. Data access rules varied by function. Some pilots used approved enterprise systems, while others relied on tools procured locally. Model outputs were reviewed inconsistently. Business cases were written in different formats. Success metrics ranged from “hours saved” to “user satisfaction” to “strategic learning.” No one had a reliable enterprise view of what was working, what was risky, or what deserved more funding.

We came to describe this condition as the pilot plateau. It occurs when a company is capable of experimentation but not yet capable of institutionalization. The organization can start AI projects, but it cannot scale them reliably. It can produce local wins, but it cannot translate them into operating-model change.

At Meridian, the symptoms were visible. A customer-service AI tool reduced average drafting time, but it did not improve overall resolution time because the escalation process remained unchanged. A sales AI tool generated better account summaries, but pipeline conversion did not improve because managers never redesigned follow-up cadence or coaching routines. A supply-chain model identified potential disruption risks, but planners did not trust the recommendations because the data sources were unclear. An agentic workflow could update internal systems, but risk leaders stopped deployment because no one had defined what the agent was authorized to change.

The lesson was uncomfortable. AI had made pieces of work faster. It had not made the company meaningfully better.

The Governance Reset

Meridian’s turning point came after a failed agent pilot in its service operations group. The agent was designed to read customer tickets, retrieve contract information, summarize prior interactions, draft a response, and recommend whether the case should be escalated. In a controlled demo, it performed well. In production testing, it exposed multiple weaknesses.

The agent occasionally retrieved outdated policy documents. It recommended different treatments for similar cases because customer records were incomplete. It drafted language that sounded authoritative but did not always reflect contractual nuance. It also revealed a deeper issue: no single executive owned the end-to-end service workflow. Customer operations owned the agents. Legal owned policy language. IT owned the system integrations. Data governance owned access standards. Risk reviewed the controls. But no one owned the combined outcome.

We recommended that the company pause deployment and conduct a structured review. At first, the business unit viewed the pause as bureaucratic interference. But the review produced a better diagnosis. The issue was not that the agent was useless. The issue was that Meridian had not built the governance architecture required to let agents operate safely.

The CEO established an AI Governance Council, but with one important condition: it could not become a committee that merely said no. Its mandate was to make responsible scaling faster. The council included leaders from technology, legal, cybersecurity, data, risk, HR, operations, and the major business units. It was not designed to review every prompt or approve every small experiment. It was designed to define the rules of the road.

The first step was inventory. Working with Meridian’s leadership team, we helped map the company’s AI activity: approved tools, informal tools, vendor systems with embedded AI, internal models, agent pilots, data sources, business owners, risk levels, and production status. This exercise was humbling. The company discovered more AI usage than executives expected, much of it outside formal oversight.

The second step was classification. Meridian created risk tiers for AI use cases. Low-risk productivity tools received clear but lightweight rules. Medium-risk systems required documentation, testing, business ownership, data review, and monitoring. High-risk systems required legal review, bias testing where relevant, cybersecurity approval, human oversight, escalation paths, audit trails, and executive accountability. Certain uses were prohibited until further review.

The third step was ownership. Every AI system moving beyond pilot stage needed three owners: a business owner responsible for value, a technical owner responsible for performance and integration, and a risk owner responsible for controls. This ended the common pattern in which everyone supported an AI initiative but no one could be held accountable for its outcome.

The fourth step was monitoring. Meridian required production AI systems to be observed over time. Accuracy, usage, cost, exception rates, user overrides, customer impact, incident reports, and business outcomes had to be measured. AI governance became less like a launch approval and more like an operating discipline.

This reset changed the mood inside the company. Some leaders expected governance to slow AI down. Instead, it clarified which projects were worth scaling and which were distractions. It created a common language for risk. It gave business units confidence that if a use case met the standard, it could move faster. Governance became not the enemy of innovation, but the condition for serious innovation.

Data Quality as the Hidden Gatekeeper

The second lesson Meridian learned was that AI strategy is often limited less by model capability than by data quality. Several of the company’s stalled pilots failed for the same reason: the model could not compensate for fragmented, outdated, inaccessible, or poorly governed data.

This was most obvious in logistics. Meridian wanted to use AI to predict delays, recommend routing changes, and notify customers proactively. The models performed well in narrow tests, but production results were inconsistent. The underlying data came from multiple systems, some updated in real time and others manually refreshed. Customer commitments were stored in different formats. Exception codes varied by region. Some planners relied on local spreadsheets that never entered the central system.

The AI did not create this disorder. It revealed it.

This is one of the defining features of AI transformation. It exposes the quality of the enterprise beneath the surface. Companies with strong data foundations can move quickly from insight to action. Companies with fragmented data find that AI produces either generic output or unreliable recommendations. The model becomes a mirror.

Meridian responded by treating data quality as a prerequisite for scaling, not as a parallel IT project. The company identified priority data domains tied to its most valuable AI use cases: customer records, contract terms, shipment status, service history, pricing logic, vendor performance, and operational exceptions. It assigned data owners, cleaned definitions, improved metadata, and created access rules for AI systems.

This work was not glamorous. It did not produce immediate headlines. But it changed the trajectory of the AI program. Once the data foundation improved, several pilots became viable. The service agent could retrieve current policy documents. The logistics model could distinguish between routine delay signals and meaningful disruption risks. The sales assistant could generate account briefs based on reliable customer history rather than partial records.

Meridian’s executive team came to understand that data quality is not a technical hygiene issue. It is strategic infrastructure. AI cannot scale reliably when the organization does not know what it knows.

Change Fitness

Governance and data quality solved only part of the problem. Meridian also had to confront a human issue. AI changed work faster than the organization had prepared people to absorb.

Employees were not uniformly resistant. Many were curious. Some were enthusiastic. Others were anxious. The strongest performers quickly learned to use AI to extend their capability. Weaker performers sometimes used AI to produce polished but shallow work. Managers struggled to evaluate output when they did not know how much of it came from employees and how much came from machines. Some teams accelerated. Others became overwhelmed by the volume of AI-generated drafts, summaries, recommendations, and alerts.

During our work with Meridian’s senior team, the COO began using a phrase that captured the missing capability: “change fitness.” It meant the organizational capacity to adapt repeatedly without losing coherence, trust, or performance. Change fitness was different from traditional change management. It was not a communication plan attached to a single transformation. It was a standing capability for working in an environment where tools, workflows, and roles would continue to evolve.

Meridian defined change fitness across four dimensions.

The first was managerial fluency. Managers needed to understand what AI could do, where it failed, how to review AI-assisted work, and how to redesign workflows. The company trained managers not only to use tools, but to ask better questions: What decision is this system influencing? What data is it using? What happens when it is wrong? Which human judgment must remain in the process? What metric will tell us whether the workflow improved?

The second was role redesign. Meridian stopped treating AI adoption as an individual productivity initiative and began decomposing work into tasks. Some tasks could be automated. Some could be augmented. Some required human judgment. Some became more important because AI was involved. This led to revised job expectations in sales operations, customer service, finance, and logistics planning.

The third was trust-building. Employees needed clarity about whether AI was being introduced to help them perform, monitor them more aggressively, or replace them. Meridian’s leaders avoided vague reassurance. They acknowledged that roles would change, then described the new skills and pathways employees would need. They invested in training, internal mobility, and new apprenticeship models for junior employees whose routine work was being altered.

The fourth was learning velocity. Meridian created feedback loops so teams could report where AI was useful, where it failed, and where workflows needed redesign. The company treated frontline experience as strategic intelligence. Employees were not merely users of AI systems. They became sensors for how those systems performed in real work.

Change fitness became the missing link between technology and value. Without it, even well-governed AI systems struggled to gain adoption. With it, employees began to understand how AI fit into the work rather than feeling that AI was being imposed on top of the work.

The Agentic Autonomy Question

The most difficult decisions involved agentic AI. Meridian’s leaders saw the potential clearly. Agents could reduce handoffs, monitor workflows, retrieve information, trigger actions, and support decisions across functions. But they also introduced a new risk category. A chatbot could be wrong. An agent could be wrong and act.

We helped Meridian develop a risk model for autonomy. It asked four questions for every agentic workflow. First, how consequential is the action? Second, how reversible is the action? Third, how reliable is the data environment? Fourth, how clear are the rules governing the workflow?

Low-consequence, reversible actions could be automated more easily. Examples included creating internal summaries, drafting routine updates, scheduling follow-ups, and preparing recommended actions for review. Medium-risk actions required supervised execution. The agent could complete parts of the workflow, but humans approved exceptions, external communications, pricing changes, or customer-impacting decisions. High-risk actions required tight controls or remained human-led. These included regulated decisions, legal commitments, major financial approvals, sensitive personnel decisions, and actions affecting customer rights or obligations.

This framework prevented two common errors. It stopped enthusiasts from pushing every workflow toward autonomy simply because the technology could perform tasks. It also stopped skeptics from blocking all agentic systems because some use cases were risky. Meridian learned to separate autonomy by risk, reversibility, and context.

The company also gave agents governed identities. Each agent had defined permissions, system access, logging, escalation rules, and performance monitoring. Agents were not treated as invisible background tools. They were treated as controlled actors inside the enterprise architecture.

This discipline allowed Meridian to scale agents where they made sense. In service operations, agents handled retrieval, summarization, drafting, and internal routing, while humans approved customer-impacting decisions. In logistics, agents monitored exceptions and prepared mitigation options, but planners authorized significant changes. In finance, agents flagged anomalies and prepared supporting documentation, but controllers retained approval authority.

The result was not full automation. It was governed delegation.

From Local Wins to Enterprise Value

Within eighteen months of the reset, Meridian’s AI program looked different. The company had fewer active pilots, but more production systems. AI investment was no longer spread thinly across every idea with executive sponsorship. Funding moved toward use cases with clear process ownership, reliable data, measurable value, and governance readiness.

In customer service, AI-assisted workflows reduced drafting time, improved consistency, and helped managers identify recurring issues. More importantly, the company redesigned escalation logic, knowledge management, and quality review, which improved resolution speed. In logistics, better data foundations and supervised agents helped planners respond faster to disruption signals. In sales, account intelligence tools became more useful after they were integrated with pipeline management and coaching routines. In finance, anomaly detection and AI-generated explanations improved review efficiency without removing human accountability.

The most important change was not any single tool. It was the operating model. Meridian had learned that AI value comes from connecting technology, data, governance, workflow redesign, and human adoption. Once those pieces moved together, AI stopped being a series of experiments and became a management system.

The company’s leaders also became more sober. They no longer described AI as a magic productivity layer. They described it as a capability that required discipline. During one board review, the CEO put the shift plainly: “We are not trying to have the most AI. We are trying to have the most useful AI.”

The Framework: Governance Plus Change Fitness

Meridian’s experience suggests a broader framework for enterprises trying to scale AI in 2026. Successful scaling requires two forms of maturity that must develop together.

The first is governance maturity. This includes inventory, risk classification, ownership, data controls, monitoring, escalation paths, security review, vendor oversight, and auditability. Governance maturity determines whether AI can be trusted enough to scale. It answers the question: can the organization control what it is deploying?

The second is change fitness. This includes managerial fluency, role redesign, employee trust, training, workflow adaptation, and feedback loops. Change fitness determines whether AI can be absorbed into real work. It answers the question: can the organization change fast enough to capture value?

Many stalled companies have one without the other. Some have strong governance but weak change capacity. They create policies, committees, and controls, but adoption remains shallow because managers and employees do not redesign work. Others have high enthusiasm and rapid experimentation but weak governance. They move quickly until risk, data issues, or integration failures stop them. The winners build both.

This dual maturity is especially important as agentic AI expands. Agents require more than technical deployment. They require trust architecture, workflow clarity, data reliability, human oversight, and organizational readiness. Without governance, agents are dangerous. Without change fitness, agents are unused or misused.

Strategic Recommendations for Leaders

From Meridian’s experience, we draw several recommendations for leaders seeking to scale AI responsibly.

Leaders should begin by reducing the number of disconnected pilots. This may sound counterintuitive, but focus is often the first step toward value. A smaller portfolio of strategically important use cases is better than a large portfolio of weak experiments. The priority should be business processes where AI can affect revenue, cost, speed, risk, customer experience, or decision quality.

They should then establish a visible AI inventory. Leaders cannot govern what they cannot see. Every meaningful AI initiative should have a business owner, risk tier, data source map, vendor profile, integration plan, and success metric. This inventory should include not only official tools, but also embedded AI in vendor platforms and informal usage patterns.

Next, companies should create autonomy thresholds. Agentic AI should not be governed by enthusiasm alone. Leaders need clear standards for which actions can be automated, which require supervision, and which must remain human-led. These standards should reflect consequence, reversibility, data reliability, and regulatory exposure.

Data quality should become part of AI funding decisions. If a use case depends on unreliable data, leaders should fund the data foundation before scaling the AI layer. Otherwise, the company risks building elegant systems on unstable ground.

Finally, executives should invest in change fitness as deliberately as they invest in platforms. Managers need training. Roles need redesign. Employees need credible pathways. Teams need feedback mechanisms. AI transformation is not only a technology rollout; it is a new operating discipline.

The Real Lesson

The story of scaling AI in 2026 is not that some companies have better models and others do not. Access to powerful models is spreading. The deeper difference is institutional. Some organizations are learning how to govern intelligence, redesign work, and build trust at the same time. Others are still treating AI as a collection of tools.

Meridian’s case shows why governance and change fitness now define the frontier. Governance without change becomes bureaucracy. Change without governance becomes risk. Together, they create the conditions for scale.

This is the new executive reality. AI can accelerate the organization, but it can also expose every weakness the organization has avoided fixing: poor data, unclear ownership, fragmented workflows, weak management, shallow training, and inconsistent accountability. The companies that stall will often blame the technology. The companies that win will recognize that the technology revealed the work they had to do all along.

In the high-stakes environment of 2026, AI advantage will not belong to the loudest adopters. It will belong to the firms that build resilient platforms, disciplined governance, and adaptive human systems around the technology. The winners will not simply move from pilots to production. They will move from experimentation to institutional capability.

That is the real test of AI leadership. Not whether the company can deploy intelligent systems, but whether it can become intelligent enough to use them well.