AI in Project Delivery: What a PMO Can Automate Today (and What It Shouldn’t)

We use AI daily in our own project management practice: this website itself is largely built and maintained with it, as regular readers of our AI experiments know. That makes us enthusiasts. Thirty years of delivery scars make us careful enthusiasts. Here is our honest map of where AI belongs in a PMO today, and where it does not.

What AI does genuinely well in delivery

Meeting minutes and actions. Feed a transcript to a capable model and you get minutes, decisions and an action list with owners in minutes, consistently formatted, every time. The PM’s job shifts from typing to verifying. This alone returns hours every week on a multi-workstream programme.

First-draft status reporting. Given the raw inputs (milestone data, risk log changes, workstream updates), AI drafts a coherent one-page status faster than any human. The colour of the report is not its call (more on that below), but the assembly, consistency and plain-language summary absolutely are. It also never gets tired on week 40 of a 60-week programme, which is precisely when human-written reports degrade.

Risk and issue log hygiene. Deduplicating overlapping risks, flagging stale entries with no update in three weeks, spotting that a risk and an issue describe the same problem, drafting mitigation language for review. Log quality is the classic PMO task that everyone agrees matters and nobody has time for: ideal automation territory.

Document synthesis at speed. “Summarise what changed between contract v3 and v4.” “Extract every commitment with a date from this 60-page proposal.” “Compare these two vendor SOWs for scope gaps.” Tasks that took an analyst a day now take an hour including verification, and the verification is the PM’s job, not optional.

Communication tailoring. The same delivery update needs different words for the steering committee, the technical team and the affected business unit. AI is excellent at re-voicing verified content for each audience, a real, unglamorous time-saver.

What AI should not decide

The RAG colour. As we argued in RAG status reporting that executives trust, the overall status is a judgement against defined criteria, owned by an accountable human. AI can assemble the evidence and even challenge the PM (“milestone variance suggests Amber; you reported Green; explain” is a genuinely useful prompt pattern), but accountability cannot be delegated to a model. Red means Red only if a person stands behind it.

Stakeholder and political reads. Whether the sponsor is losing confidence, whether a vendor’s reassurance is genuine, whether a team is quietly burning out: delivery leadership runs on signals that do not live in structured data. The most consequential PM decisions are trust decisions.

Estimates and commitments. AI can pressure-test an estimate, surface forgotten scope and check historical analogies. But a commitment given to a client is a human promise with a name attached. “The model said six weeks” is not a defence any steering committee will accept, nor should it.

Anything unverified going outward. Models produce fluent, confident text that is occasionally wrong. Internally, a wrong draft costs a correction. Externally (a status to a regulator, a milestone claim in a contract letter) it costs credibility that took years to build. The rule is boring and absolute: AI drafts, humans sign.

The quiet risk: skill atrophy

A PMO that lets AI write every report eventually forgets how to write one, and worse, forgets how to read one critically. Keep the thinking muscles working: rotate who verifies, review the AI’s misses openly, and treat the model as a junior analyst whose work is always checked, never rubber-stamped.

How to start, pragmatically

1. Pick one workflow. Minutes or report assembly are the safest first wins.
2. Define the verification step before switching anything on. Who checks, against what.
3. Keep confidential data governance boring and strict. Know what may leave your tenant and what may not.
4. Measure the hours returned and reinvest them in the judgement work: stakeholders, risks, decisions.

The pattern that works is not “AI replaces the PMO”. It is “AI does the assembly; people do the judgement”, and the projects run by teams who understand that split are simply better governed. Both of these capabilities (the delivery discipline and the AI implementation) are things we bring to clients through our IT project management services, tested first on our own operation.

Frequently Asked Questions

Editorial note — This article was researched and drafted with the assistance of Claude (Anthropic), and reviewed and approved by Amazing Projects before publication.