A project manager has to keep people working from the same understanding: what was agreed, what changed, and what happens next. Project phases help because humans cannot hold entire projects in working memory. A PM tracks maybe 50 requirements before context loss; handoffs lose information, and keeping people synchronized takes meetings. Gantt charts, sprint ceremonies, and status updates provide that cognitive support, compensating for the limits of sequential human processing.
Much of what fills a project manager’s day is coordination overhead that exists only because the execution layer is human.
AI has none of these constraints. It holds a full project context, processes parallel inputs, synthesizes across domains simultaneously. The phases that existed purely as information-processing scaffolding — breaking work into digestible chunks because human brains need digestible chunks — no longer have a structural reason to exist.
This is happening now, in industries where cognitive work dominates. Research, analysis, synthesis, scheduling optimization — the activities that filled project timelines and justified headcounts — are compressing into parallel AI-driven workflows.
The limits of that compression depend on what each project phase is doing.
The Three Walls
Three constraints remain beyond information processing.
The Physics Wall. ALICE Technologies, a construction AI platform, recovered 42 days on a major project by optimizing scheduling across millions of permutations . The recovery came entirely from smarter cognitive work — better sequencing, fewer conflicts, optimized resource allocation. But concrete still cures in 28 days. Large power transformers have 128-week lead times. Safety inspections require physical presence. The knowledge layer compressed. The physical layer did not.
The Politics Wall. Research on AI-driven decision-making found that AI use creates new tensions between stakeholders rather than resolving existing ones . Power dynamics, conflicting interests, and consensus-building arise from the structure of human social systems, not from information-processing limitations. A scheduling AI cannot negotiate between a VP who needs resources and a director who needs them more. It cannot read the room in a steering committee.
The Accountability Wall. In just three months — May to July 2025 — trust in agentic AI dropped 89% . Only 6% of companies fully trust AI agents to autonomously run core business processes. Even accounting for early-stage volatility, the pattern reflects a structural problem: someone human must own the outcome. Every AI-augmented project now requires a governance layer that did not previously exist — defining what AI can and cannot decide, who reviews outputs, and who signs their name to the result.
This leaves project managers with a different allocation of time: cognitive phases compress dramatically, while physical, political, and accountability phases persist — or grow.
Velocity Drift
One objection is that these limits are temporary. Give AI five years. Capabilities compound.
Perhaps. But the evidence from organizations that acted on this assumption is not encouraging.
Ninety-five percent of AI pilot programs fail . Not because the technology fails — it works. Because organizations automate flawed processes. Technology does not fix misalignment. It amplifies it. Automating a flawed process helps you do the wrong thing faster.
This is the pathology researchers call Velocity Drift : execution compression forces judgment expansion. As development cycles collapse, the bottleneck migrates to requirements clarity. Organizations ship faster in the wrong direction.
In a randomized controlled trial of experienced open-source developers, those using AI tools took 19% longer on tasks . They believed they were 20% faster. A perception-reality gap of nearly 40%. PR review time increased 91% — the bottleneck shifted from creation to verification, but nobody adjusted the process to account for it.
The walls are not falling. Organizations that pretend they have fallen are shipping faster toward outcomes nobody validated.
The Judgment Amplifier
As AI speeds up implementation, a good or bad judgment reaches further into the project before someone corrects it. The need for judgment therefore becomes more consequential.
The IMF’s framework makes the distinction cleanest: prediction assigns probabilities to outcomes; judgment assigns values to consequences . AI handles prediction. It amplifies judgment gaps. In a Harvard study, high-performing Kenyan entrepreneurs gained 10 to 15% profit with AI while low-performing entrepreneurs experienced an 8% decline. Same tool. Opposite outcomes. The difference was the ability to evaluate which AI suggestions were good and which would backfire. This is why experience is more valuable now than it has ever been — the judgment that comes from decades of pattern recognition is exactly what AI cannot replicate.
An eight-month study of 200 employees at a U.S. tech company found that AI adoption led to three forms of work intensification : task expansion, blurred work-life boundaries, and increased multitasking. Work became ambient — something that could always be advanced further. Without the discipline of phases — deliberate pauses, sequencing, review gates — the result was cognitive fatigue and weakened decision-making.
A limitation of this framework is that the cognitive-political-physical distinction is too clean. In practice, every phase blends all three. A construction planning meeting involves scheduling optimization (cognitive), contractor negotiations (political), and site logistics (physical) simultaneously. The categories describe tendencies rather than separate kinds of meeting. Phases exist on a spectrum from mostly-cognitive to mostly-political, and the compression gradient follows that spectrum.
When people disagree about the goal
Some projects contain problems that resist AI by definition, not by degree.
Wicked problems — as Rittel and Webber formalized in 1973 — have contested definitions, no stopping rules, irreversible consequences, and value-laden choices. Both ChatGPT and Claude acknowledged they cannot solve wicked problems when asked directly . The issue is not capability but category: these problems require negotiation among competing values, not optimization of a function.
The academic research on AI value alignment confirms this . All three leading approaches — crowdsourcing, RLHF, and constitutional AI — fail to accommodate reasonable moral disagreement. Whether to lay off 200 people to save a company is not an optimization problem. It is a moral question that no algorithm can adjudicate with democratic legitimacy.
This matters because most significant projects contain wicked sub-problems. Building a highway involves engineering (tame, AI-addressable) but also community displacement, environmental impact, and equity considerations (wicked, AI-resistant). The project manager sits at the intersection — and AI only helps with one side.
Previous automation waves changed the tools within phases. Gantt charts digitized scheduling. Agile shortened iterations. DevOps automated deployment. None eliminated the cognitive work within the phase itself. AI does — which is why the non-cognitive constraints become the binding constraint for the first time.
What Replaces Phases
Continuous work still needs explicit checkpoints.
PMI projects 25 million new project-oriented roles needed by 2030 . The demand for project management is growing, not shrinking. But the work inside those roles is transforming. Cognitive coordination — tracking, synthesizing, status-reporting — compresses. Political navigation, specification quality, and accountability governance become primary.
The new rhythm is not Plan then Execute then Review. It is continuous AI-driven flow punctuated by human checkpoints at three gates. The Physics Gate: can this actually be built or delivered in the physical world? The Politics Gate: are the stakeholders aligned and committed? The Accountability Gate: who owns this outcome and what happens when it goes wrong?
If the first article in this series argued that your experience is more valuable than you think, and the second asked who comes next when the expertise pipeline breaks, this one asks a structural question: which parts of your project process compensate for cognition — and which serve physics, politics, or accountability?
To apply this to a project, ask what each phase is there to accomplish. Time spent moving information is a candidate for compression. Time spent agreeing on a difficult trade-off, reviewing a consequential decision, or waiting for physical work serves a different purpose. Use the three walls to decide where to shorten the process and where people still need time and support.
References
-
“How Gen AI Could Change the Value of Expertise.” Harvard Business Review, March 2025. https://hbr.org/2025/03/how-gen-ai-could-change-the-value-of-expertise
-
“ALICE Technologies Case Studies.” ALICE Technologies. https://www.alice.construction/case-studies
-
“GenAI’s Human Infrastructure Challenge.” Center for Strategic and International Studies. https://www.csis.org/analysis/genais-human-infrastructure-challenge-can-united-states-meet-skilled-trade-labor-demand
-
“Navigating Power Dynamics in the Public Sector Through AI-Driven Algorithmic Decision-Making.” Government Information Quarterly, 2025. https://www.sciencedirect.com/science/article/pii/S0740624X25000474
-
“Workers Don’t Trust AI. Here’s How Companies Can Change That.” Harvard Business Review, November 2025. https://hbr.org/2025/11/workers-dont-trust-ai-heres-how-companies-can-change-that
-
“Harvard Business Review Survey: Only 6 Percent of Companies Trust AI Agents.” Fortune, December 2025. https://fortune.com/2025/12/09/harvard-business-review-survey-only-6-percent-companies-trust-ai-agents/
-
“Why Most AI Projects Fail.” Project Management Institute, 2025. https://www.pmi.org/blog/why-most-ai-projects-fail
-
“Cognitive Compression in the AI Era: Powerful Shift.” Strategic AI Leader, 2025. https://www.strategicaileader.com/cognitive-compression-in-the-ai-era-powerful-shift/
-
“Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.” METR, July 2025. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
-
“Machine Intelligence and Human Judgment.” Agrawal, Gans, and Goldfarb. IMF Finance & Development, June 2025. https://www.imf.org/en/publications/fandd/issues/2025/06/machine-intelligence-and-human-judgement-ajay-agrawal
-
“Artificial Intelligence: Human Judgment Drives Innovation.” Harvard Business School. https://www.hbs.edu/bigs/artificial-intelligence-human-jugment-drives-innovation
-
“AI Doesn’t Reduce Work — It Intensifies It.” Harvard Business Review, February 2026. https://hbr.org/2026/02/ai-doesnt-reduce-work-it-intensifies-it
-
“Can AI Solve Wicked Problems?” Peter Levine, 2024. https://peterlevine.ws/?p=35246
-
“AI Value Alignment and Moral Disagreement.” PMC, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12628449/
-
“Rising Demand for Project Managers: Why the World Needs More Project Leaders.” Columbia University School of Professional Studies. https://sps.columbia.edu/news/rising-demand-project-managers-why-world-needs-more-project-leaders
-
“Your PM Tool Was Designed for Humans.” Paddo.dev, 2025. https://paddo.dev/blog/your-pm-tool-was-designed-for-humans/
-
“Why Human Judgment Is Essential in the Age of AI.” Sir Andrew Likierman. London Business School, 2025. https://www.london.edu/think/human-judgement-essential-in-ai