Chapter 1 - The Administrative Tax on Project Leadership

AI PROJECT DELIVERY PATH SELECT > CONTEXT > DECOMPOSE > CLASSIFY > DELEGATE > CONTROL > EVALUATE > OPERATE > MEASURE
YOU ARE HERE: BASELINE / SELECT - identify friction, value, and the current state before choosing AI.

Project management has survived many predictions of its own obsolescence. Enterprise resource planning systems were supposed to integrate work. Collaboration platforms were supposed to eliminate email. Agile methods were supposed to remove heavyweight project administration. Low-code platforms were supposed to enable everyone to build the tools they needed.

Each changed the profession. None removed the need to coordinate human work under uncertainty.

Artificial intelligence will follow the same pattern, but with a larger effect on the information-processing layer of the job.

What a project manager actually produces

A useful way to understand the impact is to separate project management outputs from project management responsibilities.

Project managers produce schedules, reports, meeting notes, action logs, risk analyses, stakeholder maps, presentations, change assessments, decision packages, forecasts, issue summaries, and countless messages.

But those artifacts are not the purpose of the role. They are mechanisms through which the role performs its actual responsibilities: creating shared understanding, maintaining control, enabling decisions, anticipating problems, resolving ambiguity, managing commitments, and aligning people around outcomes.

AI is increasingly capable of helping create the artifacts. That creates an opportunity to redirect human attention toward the responsibilities.

Consider five common activities.

Activity 1: Turning discussion into commitments

A meeting may contain thirty minutes of conversation but only six items that materially affect delivery: two decisions, three actions, and one emerging risk. A human can identify them. A capable AI system can often identify them as well, format them, compare them to existing logs, and flag ambiguous ownership.

The project manager's higher-value work is confirming that the interpretation is correct, negotiating ownership where it is not, and deciding whether the risk requires escalation.

Activity 2: Reconciling inconsistent status

Workstream A says it is green. The schedule shows a missed predecessor. The vendor email says a delivery will arrive five days late. The project manager sees the inconsistency because the manager understands the system as a whole.

An AI system can help by comparing these sources systematically and surfacing contradictions. It should not decide which source is true without evidence.

Activity 3: Translating for different stakeholders

A technical lead needs defect-level detail. A sponsor needs impact, choices, and requested decisions. A customer may need reassurance and a recovery plan. The underlying facts are related, but the communication products are different.

Generative AI is unusually good at transformation: converting one form of information into another while preserving a defined set of facts. That can save substantial time, provided factual boundaries are controlled.

Activity 4: Detecting what is missing

Projects fail not only because known information is wrong but because critical information is absent. A statement of work may omit acceptance criteria. A plan may include development but no security review. A kickoff deck may name stakeholders but not decision rights.

AI can be used as a structured adversarial reviewer: "What would need to be true for this plan to succeed, and which of those conditions are not evidenced in the materials?" This is often more valuable than asking AI to create another plan from scratch.

Activity 5: Maintaining organizational memory

People forget. Teams change. Decisions lose context. A project can remember only what its operating system preserves.

AI systems with persistent workspaces, governed data connections, retrieval, and memory can make project history easier to interrogate. But memory without authority labels can create a new problem: a system that remembers old information but cannot distinguish superseded information from current truth.

Project context has to be designed, not merely uploaded.

The productivity trap

AI adoption can create the illusion of productivity while degrading delivery.

A team generates more documents because documents are easy to generate. Meeting summaries multiply. Stakeholders receive polished but redundant updates. Analysts produce five scenarios when one would have been enough. The project manager spends more time reviewing machine output than the original task required.

This is the AI productivity trap: lower production cost causes output volume to rise faster than decision value.

The solution is to measure workflow outcomes, not generation volume.

Useful metrics include:

• elapsed time from raw input to approved output;

• human review time;

• number of defects found after release of an artifact;

• percentage of material claims with traceable evidence;

• time from risk signal to escalation;

• time from meeting decision to system-of-record update;

• stakeholder response time;

• rework caused by misunderstanding;

• administrative hours returned to project leadership.

If an AI workflow creates more content but does not improve these measures, the workflow needs redesign.

What the evidence says: productivity is real, uneven, and task-bounded

The strongest evidence for generative AI at work does not support either extreme claim that AI is universally transformative or broadly useless. It supports a more useful conclusion: performance gains depend on the task, the worker, the workflow, and whether the task sits inside the system's current capability boundary.

In a preregistered experiment with 453 college-educated professionals performing writing tasks, Noy and Zhang found that access to ChatGPT reduced average completion time by 40 percent while increasing output quality by 18 percent. That result matters, but so does its domain: these were bounded professional writing tasks, not every form of managerial judgment. (Noy & Zhang, 2023).

A large field study of customer-support agents found a different but complementary pattern. AI assistance increased productivity on average, with much larger gains among less experienced and lower-skilled workers than among the most experienced workers. That suggests AI can transmit patterns of expert practice, but value is heterogeneous rather than uniform. (Brynjolfsson, Li, & Raymond, 2025).

The most directly relevant evidence for this book comes from the field experiment that introduced the idea of a jagged technological frontier. Dell'Acqua and colleagues found that consultants using AI performed better and faster on tasks inside the tested capability frontier, yet were less likely to reach the correct answer on a difficult task outside it. The managerial lesson is not simply "use AI." It is "learn where the frontier sits for this workflow, and build controls for the places where confidence can exceed competence." (Dell'Acqua et al., 2026).

That empirical pattern is why this book decomposes project workflows instead of assigning AI to entire job roles. A status process can contain retrieval steps that are easy to automate, deterministic calculations that should not be generated by a language model, synthesis steps where AI is powerful, and escalation decisions that remain human. Productivity comes from matching the right capability to the right unit of work.

Workflow Lift Principle: Treat productivity claims as workflow hypotheses. Prove them against your own work before scaling them.

The augmented role

The AI-augmented project manager becomes more, not less, responsible for judgment.

As AI makes analysis, drafting, planning, and monitoring cheaper, the scarce work shifts. The project manager has to decide which analysis matters, which message is appropriate, whether assumptions are realistic, and when intervention is necessary.

AI shifts the center of gravity from production toward evaluation and decision.

That change demands new competencies:

1. Work decomposition - seeing where information enters, transforms, and exits a process.

2. Context engineering - structuring the knowledge and rules the AI needs.

3. Delegation design - defining what the AI may propose, decide, or do.

4. Verification design - determining how outputs will be checked.

5. Tool selection - matching systems to workflows.

6. Agent supervision - overseeing multi-step machine work.

7. AI governance literacy - understanding privacy, security, retention, contractual, and policy boundaries.

8. Economic judgment - measuring whether the automation is worth maintaining.

These are management capabilities, not data-science capabilities.

Workflow Lift Principle: Automate around judgment, not through it

The highest-value AI workflows remove preparation, reconciliation, formatting, searching, and routine transformation around a human judgment point.

For example:

Poor design: AI reviews all project evidence and automatically changes the project health rating.

Better design: AI reviews all project evidence, identifies signals that may justify a health-rating change, shows the supporting evidence, and routes the proposed change to the accountable project manager.

The second design preserves the machine's speed while keeping material accountability where it belongs.

Failure Mode: The virtual intern fallacy

A common metaphor describes AI as an intern. The metaphor is useful only up to a point.

An intern is a legal person with continuity, social awareness, organizational membership, accountability structures, and the ability to understand when a situation feels wrong. An AI system has none of those characteristics by default.

Treating AI as a junior employee can cause managers to overestimate implied judgment. The better metaphor is a highly capable, nonhuman work engine with uneven competence and configurable access.

It can be extraordinary in one task and unreliable in the next. It can work at machine speed. It can also scale an error at machine speed.

Case Evolution - Project Alderline

Situation. Alderline’s project manager begins with a one-week friction log rather than an AI shopping list. The log shows that the largest recurring burden is not writing itself: it is collecting workstream updates, reconciling schedule and risk information, converting meeting discussion into controlled actions, and rebuilding the same facts for different audiences.

Baseline. The weekly executive-status cycle consumes roughly four hours of project-manager effort, with most of that time spent gathering and reconciling evidence. Meeting closeout often takes another hour because actions, decisions, and scope signals are manually transferred into separate logs.

PM decision. The manager does not declare “status reporting” an automation target yet. Instead, Alderline records the current time cost, error points, handoffs, and judgment points. The first artifact in the case is therefore a friction log, not a prompt library.

What evolves. Alderline now has a measured current state. Every later AI claim will have to improve on that baseline rather than rely on enthusiasm or anecdote.

Field Application: The one-week friction log

For one working week, do not begin by asking, "What can AI do?"

Instead, track moments when project work experiences friction. Record:

• what triggered the work;

• what information you had to collect;

• what transformation you performed;

• what judgment you applied;

• what output you created;

• what system you moved it into;

• how long the task took;

• what made it tedious or error-prone.

At the end of the week, classify the work into four categories:

Search and retrieval. Finding information.

Transformation. Summarizing, reformatting, translating, extracting, or drafting from known information.

Analysis. Comparing, identifying patterns, testing assumptions, or generating options.

Judgment and accountability. Making a material decision, negotiating, approving, escalating, or accepting risk.

The first three categories contain many AI opportunities. The fourth requires the greatest caution.

This simple log is often a better AI strategy document than a list of popular tools.

Before You Move On

Answer these without looking back. If one is unclear, revisit the chapter before completing the implementation action.

1. Why does the chapter begin with a friction log instead of a list of AI tools?

2. Which categories of project work are common AI opportunities, and which category requires the greatest caution?

3. What workflow measure would demonstrate that AI reduced administrative burden rather than simply increased output volume?

Implementation Close

Action: Track administrative friction for five working days.

Artifact: A friction log ranked by time, delay, and consequence.

Measure: Hours spent on collection, reconciliation, reformatting, and follow-up.