The Critical Path Is Probably Wrong

A critical path reset starts by questioning the path, not by accelerating every task along it. We have become very good at speeding up work along a route we rarely question. Generative AI makes that habit more dangerous, not less. If the critical path was designed around the limits of human attention, production capacity and coordination, what happens when those limits change?

The critical path is probably wrong. Not because project managers have forgotten how dependencies work, but because some of the dependencies we treat as permanent were merely the most practical arrangement under yesterday’s constraints.

For years, agility challenged organizations to leave the comfort of sequential planning. The difficult part was often cultural: moving from certainty on paper toward learning in practice. AI introduces another challenge. It asks whether we should continue accepting the inherited sequence at all. An assistant that dutifully completes every familiar stage may be demonstrating excellent compliance with a process that no longer deserves to exist.

When the process becomes the objective

Consider a familiar initiative. A team identifies an opportunity, prepares a business case, obtains approval, gathers requirements, commissions a design, builds a prototype and finally discovers whether anyone wants the result. Each stage is defensible. Together they may produce a compelling audit trail. But the sequence contains a troubling assumption: that the work needed to justify a decision must come before the work that could inform it.

Sometimes a business case is a genuine decision instrument. It forces the organization to articulate economics, trade-offs and exposure. Sometimes it is simply a ticket to the next meeting. If an organization already has a strategy, policy, mission and agreed risk appetite, why should every exploratory initiative require the same ceremonial journey before it can produce evidence?

A critical path reset does not abolish governance. It is to stop mistaking one historical expression of governance for governance itself. The meaningful question is whether the organization has satisfactory evidence to proceed, change direction or stop. The form that evidence takes should be proportionate to the decision, not predetermined by a template.

Consider what that means in practice. A team might spend weeks obtaining permission to build a demonstration when a tightly bounded demonstration could have clarified the very uncertainty the approval meeting was meant to address. The order can be reversed without ignoring risk: authorize a small, reversible investigation within existing policy, learn from what happens, then ask for broader commitment. The scope of permission changes with the exposure, rather than every investigation inheriting the full approval machinery.

This distinction matters particularly in professional services. We have built sophisticated mechanisms for presenting confidence: decks, stage gates, steering committees, workplans and approval packs. Some are indispensable. Others may reflect the difficulty of producing a credible alternative quickly enough. When the economics of production change, the justification for the sequence must be revisited.

The outcome is the anchor. The route is provisional.

I am increasingly interested in a demanding proposition: make the outcome as stable as possible and the way toward it as volatile as necessary. That does not mean pretending outcomes are immutable in the face of new evidence. It means refusing to protect a delivery path simply because it was approved, staffed or elegantly documented.

Imagine an executive asking for a decision about a new customer-facing service. The conventional response may be a discovery program, a market study, a business case and then a prototype. A different response begins with the decision itself: what must be true for us to commit, and what could we credibly learn before committing?

Perhaps the answer is a working simulation, a narrow prototype, a small customer interaction or an operational test. Perhaps the answer is still a formal analysis because the downside of acting is substantial. The important shift is that the route follows the evidence requirement. It is not assumed in advance.

We should apply the same challenge to AI assistants. Ask an assistant to plan a project and it may reproduce a familiar human workflow: research, summarize, plan, seek approval, implement, review. Those steps may be appropriate. But they should not be accepted because the assistant recognizes them. Ask instead: what is the outcome, which limitations have you silently inherited from human work, and which stages would you eliminate if producing an alternative were cheap?

That is a different use of cognitive leverage. It is not simply asking AI to help us work through a plan. It is using AI to question why the plan has its particular shape. The distinction connects directly to the broader Human Premium Enterprise proposition: when cognition becomes more abundant, the scarce organizational capability shifts toward defining valuable outcomes and establishing credible grounds for action.

Prototyping changes which ideas deserve a chance

Prototyping offers a particularly clear example. Rapid prototyping existed long before generative AI. Yet it was still constrained by available people, specialist skills, production time and the cost of coordinating those resources. Those constraints shaped which ideas received attention. Teams tended to choose the concepts most likely to justify the effort of testing them.

When AI can help produce a plausible interface, a service simulation, a draft workflow or a testable interaction much more quickly, the selection logic can change. An idea that previously seemed too speculative to deserve a prototype may now be cheap enough to investigate. The organization can afford to explore possibilities that would otherwise have been rejected before they became visible.

But a prototype is not proof. A generated screen is not evidence of customer demand. A convincing simulated conversation is not evidence of operational viability. An AI-generated financial model is not an observed economic outcome. Faster artifacts can accelerate learning only when the test connects them to reality.

This is where the excitement requires discipline. The question is not how many prototypes an organization can create. It is how many consequential uncertainties it can resolve for a given investment of money, time and attention. A hundred impressive mockups can be less useful than one modest test that changes an investment decision.

Nor does lower creation cost mean zero total cost. Someone still needs to define what counts as evidence, expose the artifact to relevant conditions, interpret the results and deal with consequences. In its 2025 research, DORA describes AI as an amplifier of organizational strengths and weaknesses. A separate DORA analysis notes that effort saved in initial creation can reappear as auditing and verification work. That is a warning against confusing artifact throughput with system-level improvement.

Evidence can become a ritual too

There is a trap in replacing process worship with hypothesis worship. Hypothesis-driven work can be valuable, but the language of experimentation does not immunize an organization against bias. A team can select convenient hypotheses, design tests that are unlikely to falsify them and reinterpret disappointing results as support for its original preference. AI does not remove this problem. It may make it easier to generate polished rationalizations.

That is why I would not measure progress by the number of hypotheses generated. I would ask what evidence could actually change the decision. What result would cause us to stop? What observation would justify investment? Which assumptions remain untested, and who is entitled to interpret the outcome?

In an evidence-first organization, the evidence standard precedes the preferred answer. The cheapest test is not automatically the best test. It must be credible enough for the decision it is meant to inform. A lightweight simulation may justify another experiment. It may not justify a production rollout affecting customers, regulated processes or material financial exposure.

The executive challenge is to define the required confidence without forcing every decision through the same expensive path. That is the operating logic behind the AI Value Assessment: establish whether a value proposition merits further commitment before treating transformation as an inevitable destination. The aim is not to make every decision fast. It is to avoid making decisions expensive for the wrong reasons.

Does a human really need to approve every decision?

Here is an uncomfortable question for organizations that claim to be AI-native: why should we assume a human must intervene at every decision point?

It is tempting to respond with the familiar formula that a human must always remain in the loop. But that can become another inherited critical path. In some bounded situations, an AI system may hold more relevant operational context than the person asked to approve its next action. Requiring that person to click a button may create delay without adding meaningful scrutiny.

The stronger design question is not whether to include a human by default, but where human intervention adds value and where it merely transfers accountability onto someone who cannot realistically inspect the decision. Delegation can be broad in low-impact, reversible, observable situations. It should become narrower as uncertainty, irreversibility, rights impacts or exposure increase.

That is not a claim that autonomous systems are universally more reliable. It is a proposal to make decision rights explicit and evidence-based. The AI decision-rights question is therefore inseparable from critical-path redesign: who or what may decide, within which boundaries, using which information, with which monitoring and escalation mechanism?

For consequential applications, the NIST AI Risk Management Framework offers a useful reminder that trustworthiness is a system property to be governed and evaluated, not a personality trait we attribute to a model. Oversight need not mean manual approval of every routine step. It does mean retaining meaningful accountability, traceability, testing and the ability to intervene when the system departs from its authorized scope.

The human premium is not the number of times a person signs off. It is the quality of the judgment applied where judgment matters.

The incentive problem survives automation

Even when technology permits a new route, organizations may resist it for reasons that have little to do with technology. We say we reward outcomes. Yet many professional environments continue to recognize visible effort, complex coordination, exhaustive documentation and the performance of being busy.

There is a cultural contradiction here. A team that produces the required evidence in two days may appear less industrious than one that follows the approved process for six weeks. A consultant who questions whether a workstream is needed can threaten the commercial logic that rewards delivering the workstream. A manager who delegates a bounded decision to an AI system may be judged more harshly for a visible mistake than a colleague whose traditional process quietly consumes months.

These are plausible operating-model tensions, not measured universal laws. Their importance is that AI cannot resolve them by producing more material. If incentives still reward the inherited journey, the organization may use AI to generate its artifacts faster while leaving the underlying journey intact.

That is an expensive way to achieve a modest result. It preserves the approval theater, the handoffs and the cognitive burden, then celebrates the speed of preparing the slides. The real opportunity is to make the organization comfortable with a credible outcome reached by an unfamiliar route.

Reinvest the Capacity Dividend in the future

I have long been inclined to spend resources sharpening the axe before swinging it. AI makes that instinct more consequential. If the effort required to build a reusable capability, investigate an alternative or improve a workflow falls, the excuse for postponing future-oriented work becomes weaker.

This is why the Capacity Dividend interests me more than a narrow productivity calculation. The point is not merely to recover hours and take them out of the system. Reinvestment can produce more output, better outcomes and new possibilities without a proportional increase in human cognitive effort. But the dividend must be earned and measured; it cannot be assumed to exist at a fixed percentage.

Suppose a team genuinely frees a quarter of the effort required for a recurring activity. That is a thought experiment, not a forecast. It need not choose permanently between quality, experimentation and delivery. Some capacity can strengthen verification; some can improve the process itself; some can explore opportunities that were previously beyond reach. The right allocation depends on the marginal value of each use and the constraints that remain.

For an individual orchestrating several AI-supported workstreams, this can feel like an expansion of the practical limits of multitasking. Yet orchestration has costs: context management, prioritization, exceptions, integration and the temptation to start more than can be responsibly finished. The achievement is not keeping more tasks in motion. It is reaching more valuable outcomes while protecting the attention needed to recognize when the system is going wrong.

In other words, capacity is not the destination. It is an option on better decisions and more ambitious work.

How to run a critical path reset

For an initiative about to begin, I would ask leaders to try a bounded reset before authorizing the familiar plan.

The test of a critical path reset is whether it changes the decisions made, not merely their speed. Did the team stop an unpromising investment sooner? Did an inexpensive prototype expose a customer constraint the original plan overlooked? Did delegated decisions reduce handoffs without increasing harmful exceptions? If the organization cannot answer questions like these, it may have changed the vocabulary of the process without changing its economics.

The reset itself is subject to the same rule as the project. Do not redesign the path for the satisfaction of proving a predecessor wrong. If the established process is already the most credible and efficient route, keep it. An outcome-oriented mindset must apply to the act of redesigning work as much as to the work being redesigned.

The real constraint may be what we refuse to question

I cannot claim to have solved this problem. I am still discovering how often my own assumptions about the correct route shape the instructions I give to AI. That is part of the experiment. The interesting possibility is not that AI has eliminated human limitations, but that it exposes how many of those limitations we quietly encoded into our operating models.

We should not confuse an accelerated journey with a better destination. We should not mistake a convincing prototype for evidence, or a formal approval for judgment. And we should not assume the future of professional services is to execute yesterday’s workplan with fewer people.

If the route were designed from the outcome backward, would we choose the same critical path?

The question worth asking is simpler and harder: If we designed the path from the outcome backward, with today’s capabilities and today’s evidence requirements, would we choose the same critical path?