07 · AI Tools for Project Management¶

AI tools are genuinely useful in project management for a narrow, specific set of tasks — and actively dangerous when applied past that boundary without a human checking the output. The skill this module teaches isn't "how to use an AI tool" (that's product-specific and changes constantly); it's how to judge which PM tasks are safe to delegate to a model, and which ones only look automatable.

A framework: delegate, assist, or reserve¶

Task category	Delegate (AI does it, human spot-checks)	Assist (AI drafts, human decides)	Reserve for humans entirely
Status report drafting from raw data	Yes — summarising structured data into prose is low-risk and easy to verify
Risk register brainstorming		Yes — AI proposes candidate risks from similar past projects; PM judges relevance and rates probability/impact
EVM calculation from PV/EV/AC inputs	Yes — arithmetic is exactly where AI is reliable, still worth a spot-check
Stakeholder resistance diagnosis			Reserve — reading a specific person's political motivations requires context an AI doesn't have and shouldn't guess at
Meeting notes summarisation	Yes
Contract negotiation strategy		Yes — AI can draft a term sheet or model an FPIF scenario, but the actual negotiation call is a human's
Go/no-go recommendation on a gate review			Reserve — a gate decision (Level 3, module 05) carries accountability that shouldn't be laundered through a tool's output
Schedule dependency detection from a large plan		Yes — AI is good at flagging candidate dependencies humans miss in a large plan, bad at judging which ones actually matter

The dividing line running through this table isn't complexity — EVM arithmetic and gate go/no-go decisions can both be described in a few sentences — it's accountability and context. Arithmetic has one correct answer that's easy to verify. A gate decision is a judgment call that someone specific is accountable for, and using an AI recommendation as cover ("the tool said go") doesn't remove that accountability, it just obscures it until something goes wrong.

Worked example: using AI safely for a status report¶

Raw inputs a PM has: BAC $500,000, PV $220,000, EV $195,000, AC $210,000, three open risks, one blocked dependency.

What's safe to delegate: asking a tool to draft the EVM section of the status report from these numbers, and to summarise the three risks and one dependency into readable prose from the PM's own risk register entries.

python3 -c "
BAC, PV, EV, AC = 500000, 220000, 195000, 210000
CPI = EV/AC
SPI = EV/PV
print(round(CPI,4), round(SPI,4))
"

→ CPI = 0.9286, SPI = 0.8864. A PM should independently verify any AI-generated arithmetic against a direct calculation exactly like this one — not because AI is unreliable at arithmetic specifically, but because the cost of an unverified wrong number reaching an executive report (module 05) is much higher than the ten seconds it takes to check.

What's not safe to delegate: the report's "Ask" section — the recommendation for what the steering committee should decide given CPI 0.9286 and SPI 0.8864. That recommendation depends on context an AI tool doesn't have (how the sponsor has reacted to past Amber reports, what political capital exists, whether this project's Amber status is tolerable given other portfolio pressures) — exactly the kind of judgment this level's other modules are about.

Common failure modes¶

Failure	What it looks like	Fix
Automation bias	Accepting an AI-drafted risk register as complete because it looks thorough	Treat AI-generated risks as a starting brainstorm, not a final register — the PM's own project knowledge still has to add and validate entries
Laundering accountability	"The AI recommended it" used to justify a decision that went wrong	Every AI-assisted recommendation still needs a named human decision-maker who owns the outcome
Context blindness	AI drafts a stakeholder communication using a tone or fact pattern that doesn't fit the specific relationship history	Human review specifically for context, not just correctness
Stale data	AI summarising a document that's since been updated, producing a confidently wrong report	Point tools at the current source of truth, verify the timestamp

A practical adoption checklist for a PMO¶

Question	Why it matters
Is the task's output independently verifiable (like arithmetic) or a judgment call?	Determines delegate vs. assist vs. reserve
Does using the tool here require putting confidential project data into a system outside our data governance?	A tool that's fast but leaks vendor pricing data is not actually a net gain
Who is accountable if the AI-assisted output is wrong?	Must be a named human, every time
Does relying on this tool erode a skill the PM organisation needs to keep (module 06's capability framework)?	An associate PM who never learns to draft a risk register manually because a tool always does it first arrives at Senior PM with a skill gap

How It Actually Works¶

The delegate/assist/reserve framework maps onto a genuine error-cost asymmetry, not a comfort preference: tasks are safe to fully delegate to AI specifically when an error is cheap to detect and cheap to correct (a first-draft status report summary that a human reviews before sending — the review step catches errors at near-zero incremental cost), while tasks belong in "reserve" specifically when an error is expensive, hard to detect, or both (a stakeholder risk assessment where a subtly wrong probability estimate flows silently into an EMV calculation and compounds through every later decision built on that reserve number, per Level 3 Module 3's calibration-drift mechanic). The common failure modes in this module — confident wrong numbers, fabricated precedent — are specifically dangerous in PM contexts because they enter a workflow (EVM, risk registers, contract terms) that already assumes numeric inputs are ground truth and propagates them without a validation step, which is exactly the same "garbage in, silently compounds" mechanism as a miscalibrated risk register or an uncorrected estimating bias from Level 3 Module 8.

Exercise¶

Your PMO is evaluating whether to let an AI tool auto-generate the weekly portfolio rollup dashboard (Level 3, module 05) directly from each project's raw EVM data, with no PM review before it's sent to executives.

Using the delegate/assist/reserve framework, classify this specific proposal (auto-generate and auto-send, no review) and justify your classification.
Identify which row of the adoption checklist this proposal fails most clearly, and explain why.
Propose a modified version of the same idea that would move it from a failing proposal to something you'd approve — be specific about what step you'd insert and where.