John McIntyre
Founder, HotPMO
When I talk to PMOs about AI, I see three camps. The first uses Copilot for meeting minutes and calls it a strategy. The second is using AI to help build processes: intake workflows, automation, the plumbing. The third, mostly software-heavy organisations, has bots wired into the delivery process itself, and in places the bot has become the process.
Most organisations I meet are in camp one, and the surveys suggest that’s where the value isn’t. The Work AI Index 2026 found 87% of digital workers using AI while only 13% report “significantly better” organisational performance, to use the survey’s own measure. The gap closes when AI changes how work flows, and that redesign is PMO territory. Here are three experiments that do it, each startable this month, none needing a procurement exercise.
Experiment one: meeting transcripts are not the product
I know. Every AI conversation in every PMO starts with meeting minutes, and I nearly refused to include them. But almost everyone does them badly, and done properly this is a different animal.
Start with what a transcript is. Teams records the meeting and transcribes it, and at the end you have a wall of text. That is raw material – not the final product. Getting stuff done off the back of the meeting is the final product: an action sitting somewhere visible, with a named owner, a due date and something nudging that owner, is far more likely to get done than the same good idea left forty messages back in a Teams chat nobody reopens.
The step that decides the quality is the hand-off, and it happens after the meeting. The organiser downloads the transcript as a .vtt file, uploads it to a controlled SharePoint library, and completes a short set of context fields: what the meeting was for, what it needed to achieve, what is out of scope, and what your organisation counts as a real action. Five minutes of typing, before the AI has read a word. The evidence says context is the dividing line between organisations that get value out of AI and organisations that get exhaustion. The Work AI Index 2026 found workers in context-rich organisations were 64% less likely to feel worn out by AI and half as likely to ship unverified work, while 53% say the information their AI needs sits somewhere it cannot see. The hand-off is where you close that gap for one meeting.
With context. The AI writes its proposed actions to a review list and never to the action tracker itself. A named reviewer works down that list and edits, rejects or approves each row against four questions: did somebody actually commit to this, can another person tell when it is done, is there one accountable owner, and is the date real. Rejection comments go back into the next AI pass, so the second draft answers the objections to the first. Only approved rows are published to the tracker, and a scheduled flow reminds each owner the day before their action falls due.
Call this shape the AI sandwich: human supplies the context, machine drafts, human approves before anything moves downstream.Â
Building it in Microsoft 365 is a couple of weeks of work and you end up with a shift from transcriptions that never get looked at, to actions that are tracked and delivered on.
The meeting-to-action workflow. The two red steps are the human control points; nothing reaches the Action Tracker until a named reviewer approves it.
I have written the whole build up in Experiment One: turn Teams meeting transcripts into agreed, tracked actions (PDF): the SharePoint library and the two lists, the AI Builder prompt with its grounding rules, both Power Automate flows step by step, the ten test cases to run before you go live, and the troubleshooting table for when it misbehaves.
Experiment two: don’t let your lessons die
So many PMOs have a lessons-learned graveyard: documents that were accurate on the day they were written, filed at project close, never opened again.
A paper published this year in Campbell Systematic Reviews carries a title I’d frame: “Don’t Let the Evidence Die” (Tezok et al., 2026). The authors run evidence reviews for policy-makers, and their problem is ours: a review starts rotting the day it’s finished. Their answer is living evidence, a continuously updated system with automated surveillance for new material, defined triggers for updating, and version control, in place of the one-off document.
Transplant that into PMO practice, because deriving lessons has never been easier. Your PPM system holds years of raw material: status updates, RAID logs, stage reports, closure documents. No human has time to read it all. A bot has nothing but time. Set it loose on the history and ask it questions you’d never resource a person to answer. Which characteristics preceded our last five escalations? What did status reports say, and avoid saying, in the three months before every project that went red? What do our best projects have in common at the thirty-day mark?
The question I keep coming back to: can AI find your next leading indicators? By which I mean the signals you’d already be tracking if you’d known they were signals.
Run it with the paper’s discipline, because this is human-in-the-loop as an operating procedure rather than a slogan. In their system, humans hand-check around 10% of what the machine screens, both to train it and to catch its drift, and humans keep every final judgement. PMI’s new Standard for Artificial Intelligence in Portfolio, Program, and Project Management (2026) takes the same position, treating human oversight as a source of value rather than a compliance chore. Sampled human validation, machine scale, human accountability. Then keep it living: a project closes, the data lands, the patterns refresh. The corpus setup and the validation sample are written up in Experiment Two: don’t let your lessons die (PDF).
Experiment three: codify the reporting flow
It’s now trivially easy to hand AI your monthly data and ask for the portfolio report, and it will oblige, beautifully. The trouble is it will write a different report every month. Different emphasis, different structure, different judgement about what matters, and it will silently change again when the model updates. Your stakeholders lose the one thing reporting exists to give them: confidence that a change in the report means a change in the world.
So flip it. Use AI to help you build the reporting process. What gets pulled, from which systems, aggregated how, with what thresholds triggering an exception. Design it together, iterate over two or three cycles, then codify the logic into code where possible and fixed templates where not. From then on, AI executes the flow and drafts the narrative around the numbers, and the numbers come off the same rails every month. Comparable in March and September, auditable in between.
Codified flows bought their keep this June, when a US export-control order took Anthropic’s Claude Fable 5 offline worldwide for nineteen days (CNBC has the story). A process written down as logic survives a model outage and a vendor switch. A process living in someone’s prompt history does not.
One counterweight so the rule stays balanced: for one-off, exploratory, ad-hoc work, use AI conversationally. Save the engineering for the things that have to be consistent. The Power BI and Power Automate rails behind this one are in Experiment Three: codify the reporting flow (PDF).
Where to start
Pick one experiment and start it this month; the meeting-actions pipeline is the gentlest on-ramp and teaches the habits the other two depend on. All three build guides are free to download, with no form to fill in.
The templates and working materials behind all three live in our PMO Success Hub, and if you’d rather talk it through first, that’s what we’re here for.
Sources
- Work AI Index 2026, Glean Work AI Institute.
- Tezok et al. (2026), Campbell Systematic Reviews, DOI 10.1177/18911803261462152.
- PMI (2026), The Standard for Artificial Intelligence in Portfolio, Program, and Project Management.
- CNBC (30 June 2026) on the Claude Fable 5 export-control order.
