Jesse Jacoby
Picture a scenario that’s becoming increasingly common. The head of transformation at a Fortune 500 manufacturer pulls a trusted advisor aside after a steering committee meeting and says something that’s surfacing with growing frequency across the field: “Our governance model is fighting the work.”
Her organization is eighteen months into a multi-year operating model transformation. Halfway through, they integrated AI into three of the program’s core workstreams: a machine learning model optimizing their supply chain network design, a generative AI tool accelerating process documentation across 14 sites, and an AI-assisted change readiness assessment that was surfacing adoption risks faster than any survey-based approach could. The technology was delivering. The problem was that every governance structure around the program had been designed for a world where these capabilities didn’t exist.
Milestone gates were structured around sequential deliverables, but the AI workstreams operated in rapid iteration cycles that didn’t map to quarterly stage-gate reviews. The steering committee couldn’t evaluate AI-influenced recommendations because most members lacked the literacy to ask the right questions about model assumptions, training data, or confidence intervals. The resource model allocated people by function and phase, but AI-augmented work was collapsing traditional phase boundaries and shifting where human effort was actually needed. The PMO was tracking progress against a plan that no longer described how the work was actually happening.
This is the governance gap that almost nobody is talking about. Not the question of how to govern AI itself, which gets plenty of attention, but the question of how to govern transformation programs that are increasingly AI-augmented. The PMO structures, steering committee models, and program governance frameworks that most organizations rely on were designed in a pre-AI world. They assume a pace, predictability, and linearity that AI-augmented programs no longer follow.
The Architecture of a Pre-AI PMO
To understand what needs to change, it helps to name what most program governance structures assume.
Traditional PMO architecture was built around sequential phases with defined handoffs. Requirements lead to design, design leads to build, build leads to test, test leads to deploy. Each phase has entry and exit criteria. Steering committees meet on a cadence to review progress against plan, adjudicate scope changes, and approve the next phase. Success is measured by adherence: are we on time, on budget, and on scope relative to the baseline?
This model works well when the work is predictable and the primary risk is execution discipline. For ERP implementations, facility consolidations, and organizational restructurings with defined end states, sequential governance provides the control and accountability that complex programs require.
But AI-augmented programs break several of the assumptions this model depends on. The work is iterative, not sequential. Outcomes are probabilistic, not deterministic. The boundary between “build” and “deploy” blurs when AI models are continuously learning from production data. And the skills required to evaluate progress shift from operational expertise to a hybrid of domain knowledge and technical literacy that most steering committee members don’t possess.
The result is a governance structure that generates the appearance of control without providing the substance of it. Reports get filed. Gates get passed. But the people making the decisions don’t fully understand what they’re approving, and the milestones they’re tracking don’t reflect where the program’s real risks actually sit.
Three Governance Failures in AI-Augmented Programs
The breakdown shows up in three specific patterns that recur across industries and program types.
The milestone mismatch. Traditional stage gates assume that work proceeds in a roughly linear fashion toward a defined deliverable. AI workstreams don’t operate this way. A machine learning model goes through cycles of training, testing, validating, retraining, and revalidating before it’s production-ready. A generative AI workflow may be usable within weeks but require months of fine-tuning to reach the accuracy thresholds the business requires. Forcing these workstreams into quarterly milestone reviews either creates artificial checkpoints that don’t correspond to actual decision points, or it creates pressure to declare “completion” at a stage where the AI component is still maturing. Both outcomes undermine the governance system’s ability to surface real risk.
The literacy gap on the steering committee. In most organizations, steering committees are composed of senior leaders with deep operational and financial expertise. They can evaluate whether a process redesign is sound, whether a technology implementation is tracking to plan, and whether the business case assumptions still hold. What they typically cannot evaluate is whether an AI model’s recommendations are trustworthy, whether the data it was trained on is representative, or whether the confidence levels it reports are meaningful in the context of the decisions being made. This isn’t a criticism of those leaders. It’s a structural problem. The governance model puts people in a position to approve things they don’t have the literacy to assess. And when approvers can’t evaluate what they’re approving, governance becomes ceremony.
The resource model disconnect. Traditional program resource planning allocates people by role, function, and phase. You need a certain number of business analysts during requirements, a certain number of developers during build, a certain number of change managers during deployment. AI-augmented programs redistribute effort in ways that this model can’t capture. When an AI tool handles 60% of the process documentation work, you need fewer technical writers but more data quality reviewers. When a predictive model accelerates risk identification, you need fewer survey administrators but more people who can interpret model outputs and translate them into action plans. The PMO’s resource model keeps tracking headcount against the original plan while the actual work has shifted underneath it.
Redesigning Governance for How the Work Actually Happens
Updating transformation governance for AI-augmented programs doesn’t require scrapping everything and starting over. It requires targeted adjustments in three areas where the traditional model is most likely to produce blind spots.
Replace rigid stage gates with adaptive checkpoints. AI workstreams need governance touchpoints that align with their actual decision cycles, not with the calendar cadence of the broader program. This means building a dual-track governance structure: the overarching program retains its phase-gate model for the workstreams where sequential governance makes sense, while AI-augmented workstreams operate on a checkpoint model tied to defined performance thresholds, validation milestones, and risk triggers. The steering committee still has visibility and authority. But what triggers a review is a meaningful event in the AI workstream’s lifecycle, not just the passage of another quarter.
Build AI evaluation capability into the governance body. This can take multiple forms depending on the organization’s maturity. At a minimum, it means adding a technical advisor to the steering committee who can translate AI performance data into business-relevant terms: what the model is doing, how confident we should be in its outputs, what the known limitations are, and what would cause us to pause or reverse course. In more mature organizations, it means developing a standard evaluation framework that all AI-augmented workstreams report against, covering data quality, model performance, human override rates, and downstream impact metrics. The goal is not to turn every executive into a data scientist. It’s to give the governance body the information it needs to make genuinely informed decisions rather than defaulting to trust because they can’t evaluate the alternative.
Shift resource governance from headcount tracking to capability tracking. Instead of measuring whether the program has the right number of people in each phase, measure whether the program has the right capabilities deployed against the work as it actually exists. This means the PMO maintains a capability map that reflects what the program needs at any given point, including the new capabilities that AI-augmented work demands: data quality assurance, model performance monitoring, AI-human workflow design, and responsible AI oversight. When the capability map diverges from the resource plan, that’s a governance conversation. It surfaces the real risk, which is not “are we staffed to plan?” but “are we staffed for the work we’re actually doing?”
The PMO’s Evolving Role
These adjustments have implications for the PMO itself. In a traditional program, the PMO’s primary value is tracking execution against plan and escalating deviations. In an AI-augmented program, that function still matters, but it’s no longer sufficient.
The PMO in an AI-augmented transformation needs to become what’s best described as a translation layer. It sits between the technical teams building and deploying AI capabilities and the business leaders making strategic decisions about the program’s direction. Its job is to ensure that AI workstream progress, risks, and dependencies are communicated in terms that the governance body can act on, and that governance decisions are translated back into actionable guidance for the technical teams.
This is a skill set most PMOs don’t currently have. It requires people who understand both the business context of the transformation and the operational characteristics of AI-augmented work. Organizations that recognize this gap and invest in building or acquiring this capability will govern their AI-augmented programs far more effectively than those that ask a traditional PMO to stretch into a role it wasn’t designed for.
The Cost of Waiting
The temptation for many organizations is to treat this as a future problem. Most transformation programs are still primarily human-driven, with AI playing a supporting role in a handful of workstreams. The governance structures in place are “good enough” for now.
That window is closing faster than most leaders realize. As AI capabilities mature and organizations embed them more deeply into transformation programs, the governance gaps described here don’t shrink. They compound. A steering committee that can’t evaluate one AI workstream will be even less equipped when three of the program’s five major workstreams are AI-augmented. A resource model that misallocates effort on one AI-driven workflow will produce cascading misallocations when AI is integrated across the entire program. And a milestone structure that doesn’t reflect how AI work actually progresses will generate increasingly misleading signals about program health.
The organizations that will navigate this transition best are the ones that start adapting their governance structures now, while the AI components of their programs are still manageable enough to learn from. The ones that wait until AI-augmented work is the majority of their transformation portfolio will find themselves trying to redesign the plane while it’s already at altitude.
Your PMO was built for the programs of the last decade. The programs of the next decade are already arriving. The question is whether your governance will evolve to meet them, or whether it will keep generating confident reports about a version of the work that no longer exists.
