What should a PMO automate first, and what should it leave alone?

The useful exercise is not "what can AI do". It is "what happens if it is wrong".

We built this catalogue out properly for a PMO platform and ended up with 42 catalogued decisions, each assigned to one tier. The number matters less than the exercise. Most PMOs that do this find the Tier 1 list is longer than they expected and the Tier 3 list is shorter than they claimed.

Reading is the best first automation in a PMO for four reasons.

• The input already exists in writing. No new data capture, no process change, no change management fight.
• The volume is real. Nobody reads every closure report. That is why the pattern in them never gets found.
• Checking is fast. A PM can verify a summary against the source in a minute. Compare that to checking a forecast.
• Being wrong is cheap. A bad summary gets corrected. A bad approval gets litigated.

What that looked like in practice

An insurance PMO was reading project closure reports by hand to find repeat problems. A senior PM sampled them once a year, because reading all of them was not realistic. After automating the reading, review dropped from 30 to 50 hours per cycle to 4 to 8 hours, and they moved from one cycle a year to four.

Anything that should be arithmetic

This is the most common and most expensive mistake. Critical path, float, capacity, earned value and budget rollups are deterministic calculations. They have one right answer and you want the same one every time you ask.

When we built a clinic scheduling tool, the scheduling core was deliberately not AI. Capacity recalculation, the overbooking guard and the audit log all run as plain code. A language model sitting where a formula belongs gives you a system that is slower, more expensive, and occasionally confidently wrong about a number.

Anything that commits the organisation

Approvals, escalations, vendor selection, contract decisions, anything touching a person's role or performance. These are Tier 3 and they stay there, not because the technology cannot produce a recommendation, but because the accountability cannot move.

Anything you cannot check

If nobody on the team can look at the output and tell within a minute whether it is right, do not automate it yet. Build the check first. An automation nobody verifies is not saving time, it is moving risk somewhere you cannot see it.

1. Summarise what you already collect. Closure reports, status updates, meeting notes. Tier 1.
2. Extract structure from unstructured text. Action items, decisions, risks out of transcripts and email. Tier 1.
3. Draft the recurring documents. Weekly status, portfolio snapshot, risk register. Tier 2, human sends.
4. Flag the exceptions. Schedule slips, stale risks, missing updates. Tier 2, human decides.
5. Stop. Run those for a quarter before adding anything. Most PMOs that fail at this add step six in month two.

In a readiness review for a national construction firm, 25 candidate initiatives scored 21 out of 30 on organisational readiness across a user base of more than 400 people. The blocker was almost never the technology. It was that nobody had decided who owns the output when the agent produces it. Decide that before you build. It is a one-hour conversation that saves a quarter.