With AI, judgement is the constraint that doesn't scale

AI scales execution, not judgement

AI will not make strategic execution simple by making project tasks faster. It will make execution more demanding because it scales production faster than it scales judgement.

That is the central constraint. AI can draft, summarise, analyse, forecast, classify, schedule and generate options at a pace that makes traditional coordination look slow. But it cannot absorb accountability. It cannot decide what the organisation is trying to become. It cannot know, on its own, which trade-off is acceptable, which stakeholder concern is material, which assumption deserves distrust, or which risk should be carried rather than avoided.

The leadership problem is therefore not how to inject AI into project management or other operations. It is how to design an execution architecture in which human and machine work combine safely, productively and repeatedly, with consistency and coherence. That architecture is not a technology blueprint, instead it is the management system through which choices become work, work produces learning, and learning reshapes choices in an ongoing value delivery loop.

We see some of these system capabilities emerge in product-centric operating models, or those adopted by software developers, but outside of those contexts, they are more uncommon.

The ideal execution architecture has four pillars:

Projects as temporary vehicles. Projects still matter because strategic work needs bounded commitments, budgets, deadlines and accountable teams. But projects are no longer the natural container for every ambition. They are vehicles for changing, building, testing, scaling or retiring something that should outlast the project itself.

Capabilities as the enduring assets. The durable asset is the organisation's repeatable ability to create value: data, platforms, workflows, decision rights, skills, governance routines, operating know-how and trust. AI makes this clearer because a local model or agent has little strategic value unless it strengthens the organisation's capacity to act again, better and faster.

Portfolios as steering systems. Portfolios should not be calendars for approving work and receiving status packs. They should be live systems for choosing, sequencing, funding, stopping and redirecting strategic work as evidence changes.

Governance as the trust infrastructure. Governance is not the compliance layer added after innovation. It is the infrastructure that makes human-machine execution accountable, traceable and safe enough to delegate into. So much of what drives true innovation requires effective governance, so it becomes intrinsic from the outset.

These pillars also dissolve an old (and false) bifurcation. Strategy is not a fixed plan set above the organisation and then handed down for execution. Execution is the way strategy becomes real and is continuously shaped. Every resource allocation, every stop decision, every exception, every workflow design, every governance threshold and every capability investment expresses the strategy. Recall Roger L. Martin's famous essay Strategy Is What You Do, Not What You Say.

So, the execution architecture is strategy in operational form.

The old question for leaders to ask was: Are the projects on track? The better AI-era question is: Is the organisation building the judgement, capabilities, decision rights and trust conditions needed to adapt while it acts?

That question cannot be delegated to the PMO, the CIO, the chief data officer or a responsible AI committee. It is a general management question to be asked at all levels because AI changes the system by which judgement is applied to work.

Faster production exposes weaker decision systems

The first visible wave of AI in knowledge work looks deceptively administrative. Meeting notes, first drafts, document comparisons, project updates, risk summaries, compliance reports and reporting packs get faster. That matters. Organisations spend a large share of managerial time producing artefacts that exist mainly to coordinate other work and provide stakeholders with a sense of what is happening.

But the deeper shift is not that administration disappears. It is that the pace of production becomes less often the main constraint. When a plan can be regenerated, a market scan refreshed, a project report drafted or a risk register updated in minutes, the scarce resource moves elsewhere.

Someone still has to decide what question matters. Someone still has to judge whether the answer is good enough. Someone still has to know what evidence is missing, what assumption is fragile, what stakeholder will resist, what regulatory exposure has been introduced and what trade-off is being hidden by a confident sentence.

This is why the common phrase "human in the loop" is too weak. A loop can be performative, as it can mean a tired reviewer clicking approve after the machine has already set the frame and context for the decision. AI-enabled execution needs humans in the decision architecture, not merely in the loop. Humans must frame the work, set thresholds, contest recommendations, resolve ambiguity, accept accountability and redesign the system when the pattern of errors changes.

This changes project and execution roles. The manager valued mainly for task chasing, status collation and document production will be squeezed. Whereas the leader who can orchestrate human-machine workflows, validate evidence, expose trade-offs and hold the line on accountability becomes more valuable. The same is true in portfolio management. The value moves from producing the pack to improving the allocation decision. While there is promise that administrative automation will create more time for healthier decision debate, the right management system needs to be in place to bring about that debate.

There is also a difficult capability problem inside this shift. Judgement is not produced by outsourcing tasks to an AI. It is built by practice, exposure, feedback and responsibility. If AI absorbs too much entry-level analysis, drafting and coordination, organisations may weaken the apprenticeship through which future senior judgement is formed. A firm that automates the training ground and then creates a management system that demands better judgement from its people is designing a contradiction.

The sensible design is neither some luddite-inspired AI refusal nor automation maximalism. Routine production should be automated where it is low-risk and repetitive. But the work that builds judgement must be deliberately preserved, redesigned or replaced with new learning mechanisms. Otherwise AI will make organisations faster and shallower at the same time.

Projects vs. BAU isn't really a thing anymore

The inherited distinction between projects and business-as-usual work is losing it's place as a mental model from which to reason about strategic change. It was useful when change could be treated as episodic and operations as stable. AI weakens that assumption. Strategic work increasingly lives in a continuous flow where delivery, learning, automation, process redesign and capability building happen together.

Projects are not obsolete. They remain useful because concentrated effort still needs scope, funding, deadlines, governance and accountability. A project can create a temporary coalition for change. It can protect attention in a noisy organisation. It can force a decision about what will be done by when.

But the project is not the strategic unit. The strategic unit is the capability, value stream, decision process, platform service, data asset, customer journey or operating routine that becomes stronger because work was done. The test is not whether the work was labelled a project, program, product increment, operational improvement or BAU activity. The test is whether it delivered near-term value while increasing the organisation's enduring ability to create value again.

This is the crucial shift. A strong execution architecture does not separate delivery from capability building. Every significant piece of work should do both. A pricing automation should improve today's pricing decisions and strengthen the data, governance and workflow capability behind future pricing decisions. A customer-service agent should reduce today's handling effort and improve the escalation logic, knowledge base and accountability model for future service. A portfolio analytics upgrade should produce better current decisions and strengthen the organisation's decision intelligence capability.

The old model made it too easy to optimise the wrong thing. Projects were approved for local business cases. Operations were funded to keep the machine running. Capability foundations were treated as overhead unless attached to a visible initiative. AI exposes the cost of that split. Local automations can fragment data. Separate pilots can duplicate platforms. Fast prototypes can bypass controls. Business units can create impressive local productivity while making enterprise execution harder.

The better rule is simple: strategic work should be evaluated as part of one continuous flow. Approval, continuation, acceleration and stopping decisions should ask whether the work remains strategically valuable, whether it strengthens shared capabilities, whether it is safe to scale, and whether the next dollar or hour would be better used elsewhere. The label on the work is secondary.

This also changes stopping discipline. AI lowers the cost of generating proposals, prototypes, analyses and business cases. The organisation can drown in plausible work. Stopping weak work is not failure. It is the core act of strategy under scarcity. Continuation should be earned, not inherited from last quarter's commitment.

The PMO, transformation office or portfolio function must therefore stop defending the project inventory and start governing the flow of strategic work. Its purpose is to make sure temporary work strengthens enduring capability, not merely to report delivery. A green project that leaves no reusable asset behind is not healthy. An operational improvement that quietly builds a strategic capability may be more valuable than a formally branded transformation.

Portfolios become living choice systems

Portfolio management has often been a periodic discipline. Leaders approve an annual plan, review quarterly packs, debate red and amber items, and intervene when cost, schedule or dependencies become troublesome. AI makes that cadence increasingly inadequate.

When plans, analysis and reporting can be regenerated dynamically, the portfolio should become a steering system. Steering is different from reporting. Reporting asks whether committed work is on track. Steering asks whether the commitments should change.

This is a sharper managerial burden, not a lighter one. AI can create more forecasts, scenarios, dependency maps and recommendations than leaders can absorb. The value comes from structuring choices, not from increasing information volume. Portfolio forums should spend less time hearing status and more time deciding sequencing, funding, capacity, risk appetite and strategic trade-offs.

Dynamic portfolio management does not mean constant churn. Churn is movement without learning. Steering is movement in response to evidence. The distinction depends on pre-agreed priorities, thresholds and decision rights. If a forecast changes, who can move funding? If capacity tightens, who decides which work slows? If a machine-generated scenario shows a better path, what evidence is required before reallocating? If a capability-building investment has no immediate business-unit sponsor but is essential to the platform, who protects it?

The portfolio office therefore becomes less like a compliance function and more like an intelligence and orchestration function. It maintains the living view of strategic options. It integrates financial, operational, workforce, technology, customer and risk data. It separates signal from noise. It records decisions so they can be revisited. It keeps the flow of work attached to strategic intent rather than to the political durability of existing commitments.

AI also changes portfolio metrics. Cost, schedule and risk remain necessary. They are no longer sufficient. Leaders need to see leading indicators, option value, capability accumulation, adoption, benefits realisation, trust conditions, workforce effects and the quality of the measurement system itself. The question is not whether a dashboard is richer. The question is whether it makes better strategic choices possible.

This returns to the point that execution and strategy are not separate acts. Portfolio steering is where strategy is continuously expressed. Every decision to accelerate, pause, stop, fund, sequence or protect work is a strategic decision. A portfolio that merely reports work cannot perform that role. A portfolio that steers choices can.

Governance moves inside the work

Governance is often treated as the chapter that comes after innovation. That sequencing is wrong for AI-enabled execution. Governance is not a brake added to the system. It is part of the system's steering, accountability and learning mechanism.

The reason is structural. AI can influence what work is proposed, how evidence is interpreted, how priorities are compared, how risks are described, how resources are allocated and how operational actions are triggered. If governance sits outside that flow, it will either be too slow to matter or too superficial to protect the organisation.

Governance must therefore move inside the work. It defines who may use AI for what purpose, what evidence is required before reliance, when human review is mandatory, who can override or contest a recommendation, how decisions are logged, how data and model provenance are maintained, how incidents are escalated, and how systems are monitored after deployment. These are not compliance niceties. They are the conditions under which delegation becomes safe.

This is why accountability cannot remain at the level of vague role descriptions. AI-assisted execution requires decision rights. A delivery leader may ask AI to generate a recovery plan. A portfolio forum may receive AI-generated funding recommendations. A business unit may deploy an agent to execute a bounded workflow. In each case the organisation must know who owns the decision, who owns the system, who owns the risk, who can challenge the output and what must be recorded.

Provenance also expands. It is no longer enough to know the final decision. Leaders need to know what data was used, what assumptions shaped the recommendation, what alternatives were considered, what the AI suggested, what the human changed, why the change was made and what action followed. Without that chain of custody, learning becomes anecdotal and accountability becomes theatrical.

The hardest part is that governance must be proportionate. Not every AI-generated note needs executive review. Not every prompt requires a formal approval trail. But higher autonomy, higher consequence, weaker explainability, sensitive data, customer impact and regulatory exposure should increase the governance burden. This is not bureaucracy. It is managerial discrimination.

Governance is therefore the trust infrastructure of the execution architecture. A system that cannot explain who decided, on what basis, with what authority and using what evidence is not an execution system. It is an activity system.

Strategy is realised in the execution architecture

The strategic risk is to mistake AI adoption for execution renewal. Adoption is easy to count. Execution renewal is harder to see because it lives in the connections between work, decision rights, governance, capabilities and value.

The organisations that gain from AI will not simply run more projects faster. They will design better execution architectures. They will use projects as temporary vehicles without making projects the strategy. They will treat capabilities as the assets that strategic work must strengthen. They will steer portfolios through live evidence rather than fixed approval cycles. They will make governance the infrastructure of trust rather than an after-the-fact review.

That is not a top-down cascade from strategy to execution. It is strategy being realised through execution choices. The organisation learns what its strategy means by funding some work and stopping other work; by deciding which capabilities matter; by setting thresholds for machine action; by choosing where human judgement must remain close; by adapting when evidence changes; and by refusing work that is attractive locally but destructive systemically.

AI makes this more important because it increases the speed and volume of plausible action. More options appear. More reports can be written. More scenarios can be generated. More workflows can be automated. The bottleneck becomes the quality of choice. Without stronger judgement, faster execution produces faster drift.

There is still uncertainty. Multi-agent orchestration is not yet the everyday reality of most organisations. Evidence of long-run enterprise value is still developing. Many claims about AI-enabled portfolio management remain practitioner-led rather than independently proven. Direction of travel should not be confused with maturity of deployment.

But uncertainty about pace is not uncertainty about the managerial problem. As AI scales execution, judgement does not scale by itself. It must be designed into the work through capabilities, portfolios, governance and learning systems.

The durable position is clear. AI-enabled strategic execution is not a project management upgrade. It is a judgement challenge requiring an execution architecture. Leaders who treat AI as tooling will get faster fragments. Leaders who treat execution architecture as strategy in action have a chance to build organisations that move faster without becoming less wise.

References

• Organisations need AI strategy, systems, people and governance capabilities to move from pilots to scaled value. — Enterprise AI Maturity Update (MIT CISR)
• Organisations should reduce project overload by stopping more initiatives and treating continuation as an active strategic choice. — Focus on Fewer Projects (HBR)
• AI-enabled organisations should move from isolated projects to companywide capability-building around shared data assets, central governance and business use cases. — Building AI Organisations (HDSR)
• AI copilots require redesigned workflows, decision rights and governance around human judgement rather than simple tool adoption. — From Tools to Teammates (Boston University)
• Senior leaders maintain trust in AI-assisted execution by embedding governance, documentation, testing and human oversight throughout the AI lifecycle. — AI Risk Management Framework (NIST)
• LLMs can support strategic decision-making by generating and evaluating strategies, but human strategists still need distinctive judgement and complementary assets. — AI and Strategic Decision-Making (Strategy Science)
• Portfolio management should become continuous, AI-supported steering with plans, analysis and resource allocation updated in near real time. — Organisational Transformation in the Age of AI (WEF)
• AI is reshaping expertise by shifting human value from answering questions to providing context, judgement and accountability. — Rethinking Expertise in the Age of AI (MIT Sloan)
• AI should be governed as a choice-architecture problem that shapes the decision environment, not merely as a model-performance problem. — Intelligent Choice Architectures (MIT Sloan)
• Agentic AI is forcing operating-model and governance change rather than merely improving existing management routines. — Emerging Agentic Enterprise (MIT Sloan/BCG)
• AI-enhanced KPIs should become descriptive, predictive and prescriptive, revealing relationships across organisational silos. — Future of Strategic Measurement (MIT Sloan/BCG)
• Governance must shift from after-the-fact control to embedded infrastructure, certification and cross-functional operating teams. — Scaling Enterprise AI Governance (IBM)
• AI shifts the bottleneck in knowledge work from execution to judgement, coordination and accountability. — Operating Model for the Age of AI (Bain)
• AI-enabled dynamic steering can make planning, forecasting and scenario analysis faster and more adaptable. — Dynamic Steering in Financial Planning (BCG)
• Organisations should build human-machine orchestration, digital trust, shared decision rights and continuous adaptability for AI-powered work. — Human Capital Trends (Deloitte)
• AI capabilities in project-based organisations support value definition, value creation and value capture, especially when organisational agility is present. — AI and Value in Project-Based Organisations (Emerald)
• Agentic AI makes audit trails, accountability and identity governance harder because autonomous systems can act without clear human approval. — Auditing Agentic AI (ISACA)
• Knowledge workers are shifting from direct task execution towards supervising, steering and validating AI-supported workstreams. — Future of Work Report 2025 (Microsoft)