Can Companies and Institutions Scale AI without Losing Control?

“We’re building AI into critical workflows with fewer disciplines than we’d ever tolerate in other infrastructure. You wouldn’t run payroll on an untested system with no rollback plan, but organizations are doing the equivalent with AI every day.”

Candice Quadros, Bay Area Enterprise AI & Transformation Leader

The pressure on enterprise leaders is no longer simply to adopt artificial intelligence. It is to scale it, govern it, and keep it from outrunning the institutions built to manage it.

Across industries, including education, companies that once celebrated successful pilots now confront a different and more difficult challenge: the organizational gap between what AI can do and what an enterprise is actually prepared to absorb. AI transformation is not a technology problem. It is an organizational evolution problem, and companies learning that lesson late pay for it in stalled programs, eroded credibility, and investments that never return their value.

The gap shows up differently depending on the sector, but the underlying failure is consistent. A pilot succeeds in a controlled environment, generates executive enthusiasm, and stalls the moment it touches the broader organization. Legacy systems resist integration. Data ownership fragments across teams that do not communicate. Governance frameworks adequate at the pilot stage prove inadequate at scale. The workforce never consulted during the design phase finds reasons, some technical and some cultural, not to use the tool at all. By the time leadership recognizes the problem, the investment has already been made, and the credibility of the initiative has already begun to erode.

What makes this moment particularly consequential is the speed at which AI moves from productivity experiment to operating infrastructure. The same tools that once lived at the edge of enterprise workflows are now being built into critical systems: financial processing, compliance decisions, customer-facing services, and clinical operations. That shift changes the stakes of getting governance wrong. A failed pilot wastes budget, and then a failed AI system embedded in core operations affects customers, regulators, and institutional credibility simultaneously.

The organizations scaling AI responsibly are the ones that recognized this transition early and built their operating models accordingly, before the pressure to move fast made careful thinking harder to justify. And we spoke with Bay Area Enterprise AI & Transformation Leader, Candice Quadros, to better understand what separates organizations that succeed in scaling through these new operational and workforce goals.

Candice Quadros is an enterprise AI and transformation leader based in the SF Bay Area, with over two decades of experience driving large-scale program management, AI governance, and digital transformation across the technology industry.

Most recently, Quadros served as director of program management and productivity at Roku, where she built and led the enterprise PMO, scaled AI adoption to 1,000+ employees, and established responsible AI governance frameworks across more than 40 cross-functional teams. Prior to Roku, she held leadership roles at Google and Microsoft.

Quadros was named one of the Top 100 Women of Influence by the Silicon Valley Business Journal (2021), a winner of the Leadership Excellence in Technology Award (2021), and recipient of the Authentic Leadership Award by the National Diversity Council (2020). She is an active speaker and mentor, with a particular focus on advancing women and program managers in technology careers.

She holds a master’s in computer science from the State University of New York and a bachelor’s in computer engineering from the University of Mumbai, India.

Most enterprise AI programs fail in the same place and for the same reasons. The pilot may run cleanly, the model performs, and somebody gives a presentation with good numbers. What nobody tends to map carefully enough, however, is the distance between the controlled environment where the pilot lived and the messier, more contested terrain of the real organization.

Production data is not the curated dataset the pilot ran on. It sits scattered across legacy systems, owned by teams with different definitions of the same fields and no particular incentive to coordinate. The workflow that seemed clear during scoping turns out to be four different workflows depending on which business unit gets asked. The executive who sponsored the project attends the kickoff and disappears after it. The person now accountable for making the system work in production is not the person who designed it, and that handoff was never made explicit.

“The first thing that surfaces isn’t a technical problem. It’s a trust problem,” says Candice Quadros, an AI and transformation leader with over twenty years of experience in large-scale program management, AI governance, and digital transformation across the technology industry. “The pilot ran in a controlled environment with a motivated team. Deployment means everyone else: skeptics, edge cases, undocumented workflows, and users who weren’t in the room when the demo got applause.”

Data readiness, process clarity, and change ownership are the three things that collapse first, and each one is routinely underestimated until the project is already in trouble. Data that appeared usable is scattered and inconsistently governed. The workflow the AI was designed to support becomes contested the moment it has to become the workflow rather than sit beside it. And the champion who drove the pilot forward has no clear successor accountable for what comes next.

Pilot success and deployment readiness are different disciplines. Organizations treating them as continuous rather than distinct make a category error that compounds over time. Pilots are optimized to show what is possible. Deployment requires demonstrating what is ready, which is a harder and more honest standard to meet.

“The companies that get this right treat deployment readiness as a separate workstream from pilot success,” Quadros says. “Not a continuation of it.”

Scaling AI without losing control requires the institutional infrastructure to manage decisions, resolve disputes, and maintain accountability when things do not go as planned. Program management offices, technical program managers, and formal governance functions exist precisely for this purpose.

In AI initiatives where speed and agility get treated as substitutes for structure, they are consistently underused. The result is AI programs that generate activity without generating accountability, and that accumulate risk without building the mechanisms to surface or respond to it.

“AI strategy documents are easy. Execution is where it falls apart, and that’s exactly where PMO and TPM functions earn their value,” Quadros says.

Effective governance translates strategy into named decision-makers rather than committees, because diffuse ownership is functionally equivalent to no ownership. It builds escalation paths before they are needed, not after the first failure surfaces in a customer-facing output. It treats fallback plans as standard artifacts rather than emergency responses. And it coordinates across the functions that AI programs inevitably touch: data, legal, IT, finance, and the business units that live with the outcomes.

So the focus becomes organizational rather than procedural. “The PMO shouldn’t be the AI police. But it should be the connective tissue that keeps accountability from evaporating when things get complex.”

Governance operating as enforcement tends to slow programs down without improving their quality. Governance operating as coordination keeps accountability intact while preserving the speed organizations need to move from strategy to execution. That difference is a design choice, and it has to be made deliberately before the pressure of delivery makes it harder to make well.

One of the most consistent failure modes Quadros observes is shadow AI: public servants and enterprise employees using AI tools without any governance framework in place, not because they are trying to circumvent policy, but because no policy exists that addresses how they actually work. The risk is not that people use AI. It is that they use it without the organizational structures that allow failures to be caught, decisions to be traced, and accountability to be maintained. Without governance designed around how AI actually gets used, not how it was intended to be used, accountability evaporates precisely when it matters most.

Usage dashboards and adoption rates are easy to generate, easy to present, and almost entirely disconnected from whether the organization operates differently because of AI. They measure that the engine is running. They say nothing about where the car is going or whether it is headed in the right direction.

This is where enterprise AI measurement most consistently falls short, and where the gap between reported progress and actual value tends to widen quietly over time.

“Tool adoption rates and usage dashboards are input metrics. They tell you the engine is running. They don’t tell you where the car is going,” Quadros says.

The metrics that surface real operational value are different in kind. Cycle time reduction reveals whether AI is actually accelerating work that matters. Error and rework rates reveal whether decision quality is improving or whether AI is simply producing errors faster. Decision quality and speed, and what people do with the capacity that AI frees up, are the measures that connect AI output to organizational performance.

The clearest example Quadros offers comes from a financial close automation program that tracked not only hours saved but where those hours went. The headline figure was more than 300 hours saved per cycle.

“The real story was that the finance team was now spending that time on variance analysis they’d previously deprioritized, which actually changed decisions,” she explains.

The value of the AI was not what it automated. It was what the organization became capable of doing because of what it automated.

Building that measurement discipline requires deciding in advance what change looks like, not retrospectively claiming credit for whatever happened after deployment. It requires baseline data before implementation. And it requires the organizational willingness to report honestly when AI is not producing the change it was designed to produce. That is a governance and culture question as much as a metrics question, and it connects directly to whether leadership has built the conditions for honest reporting or only the conditions for good-looking dashboards.

As AI moves from productivity tool to operating infrastructure, the governance question shifts from adoption to resilience. The systems now most dependent on AI are often the same ones where failure carries the highest cost: financial processing, compliance monitoring, customer service operations, clinical decision support.

That concentration of risk in AI-dependent workflows creates a category of institutional exposure most enterprises have not yet built protocols to manage. The disciplines applied to conventional infrastructure, rollback plans, incident response, and lifecycle governance have not migrated to AI at the same pace as the technology itself.

“We’re building AI into critical workflows with fewer disciplines than we’d ever tolerate in other infrastructure. You wouldn’t run payroll on an untested system with no rollback plan, but organizations are doing the equivalent with AI every day,” Quadros says.

Resilience requires four things operating in parallel. Model lifecycle governance that does not default to vendor decisions, because outsourcing that judgment is a risk exposure. Deliberate human-in-the-loop design for each specific use case, applied as a genuine design constraint rather than a checkbox. Incident response plans for AI failures that specify what a break looks like, who detects it, and who has the authority to act. And cross-functional AI steering that is not siloed inside IT or legal, because AI failures rarely respect functional boundaries.

The discipline Quadros prioritizes above all others is visibility. “Organizations that treat errors as edge cases to hide will be blindsided. The ones that build cultures where failures surface quickly will learn faster and recover better.”

That is a cultural commitment as much as a structural one. It requires leadership that treats AI failure as information rather than liability, and that builds the organizational conditions for honest reporting before a crisis makes honesty feel dangerous.

The organizations that build resilience before something breaks share a common characteristic: they plan for failure as a certainty rather than a possibility. They do not assume the model will perform in production the way it performed in testing. They do not assume the vendor will catch problems before the organization does. And they do not assume that human oversight, once designed into a system, will remain meaningful without deliberate maintenance. Resilience is not a feature of the technology, but a feature of the institution built around it.

Clearly, technology stacks and business cases currently dominate enterprise AI conversations. Budget cycles, model selection, vendor evaluation, integration architecture: these fill the agenda. The human adoption layer, the layer that ultimately determines whether any of it works, receives far less attention and, in most programs, far less funding. That imbalance reflects a persistent assumption that adoption follows capability, that if the tool is good enough, people will use it. Enterprise deployments consistently argue otherwise.

“Most enterprise AI conversations are about the technology stack and the business case. Almost none of them are about whether the people who are supposed to use the AI actually trust it, understand it, or felt like they had any say in how it was introduced,” Quadros says.

Trust, comprehension, and agency are not soft factors. They are the variables that determine whether a capable AI system becomes an operational asset or an expensive and underutilized investment.

“Adoption failure is almost never a feature problem. It’s a change management problem that was never funded.”

So accountability shifts away from the technology and toward the organizational choices made before and during deployment. Whether frontline teams were consulted. Whether their concerns shaped the design. Whether the introduction of AI was treated as a change management effort or a rollout announcement. Her recommendation is concrete: before the next AI deployment, ask frontline teams what they are afraid of, not as a focus group, but as a design input.

“The answers will tell you more about your deployment risk than any pilot metric.”

The enterprises most likely to scale AI responsibly are not necessarily the ones with the most sophisticated models or the largest implementation budgets. They are the ones that have built the organizational conditions for AI to work: governance that maintains accountability, measurement that surfaces real outcomes, resilience protocols that treat failure as an operational reality, and enough respect for the humans in the loop to treat their concerns as something worth designing around.

Ultimately Quadros emphasizes enterprises pursue “less AI hype [and exercise] more respect for the humans in the loop.”