Hiromi Nakatani, founder of Axis International Corp., advising leaders on project management, AI adoption, cybersecurity and governance.
getty
A few months ago, I joined a portfolio review with the leadership team of a company, there to help them weigh the initiatives under consideration. The project management office had done its job: 14 AI initiatives on one dashboard, 12 of them green. Then the CFO asked which three should receive next year’s funding. The room went quiet. Nobody could answer—not because the PMO had failed to report, but because it had reported the wrong thing. The dashboard answered “Are we on schedule?” when the question that mattered was “What have we learned, and what deserves more money?”
In earlier articles, I argued that AI initiatives should pass a readiness test before funding and be run on evidence gates rather than milestones. Both ideas work at the level of a single project. But most organizations are not running one AI project. They are running a dozen, and the body that sees them all is the PMO. If the PMO keeps collecting the same information it always has, leaders will keep getting dashboards that cannot answer the CFO’s question.
Why The Reporting PMO Goes Blind On AI
The classic PMO was built to aggregate variance against a plan: percent complete, milestone dates, budget burn, a red-amber-green rating. For a conventional project with known scope, those signals are meaningful. For an AI initiative they are not, because an AI project can hit every milestone and still prove nothing. A model that is 90% built has not shown it changes any business decision. The PMO is not failing at its job; the job was designed for a different kind of work.
The fix is to change what the PMO collects, what it compares, what it puts in front of leadership and what it pools across the portfolio. I think of these as the four functions of a steering PMO.
Function 1: Collect evidence, not status.
Replace percent complete with three facts per initiative: which evidence gate it is at (feasibility, value or adoption), what evidence it has produced and what evidence it still needs to pass the next gate. “Gate two, handling time down from eight minutes to six across one team, second team not yet live” tells a leader more than “72% complete, amber.” The PMO’s role is to insist on this format and to push back when a team submits activity in place of evidence.
Function 2: Make initiatives comparable.
Comparison is the PMO’s most underrated service. When every AI initiative reports on the same one-page sheet—hypothesis, current gate, evidence produced, evidence missing, business owner, stop criteria, next review date—leaders can finally place two initiatives side by side and decide where the next tranche goes. Without a common format, funding follows the most persuasive sponsor. With one, it follows the evidence.
Function 3: Put the stop decision on the agenda.
In my experience, AI pilots rarely end because nobody is responsible for proposing that they end. The steering PMO owns the calendar of gate reviews and brings a short list of stop candidates—initiatives that missed a gate twice or whose stop criteria have been met—to every portfolio meeting. The PMO does not make the decision; executives do. But a stop decision that is never proposed is never made, and the money stays trapped.
Function 4: Pool what AI projects share.
Most AI initiatives in a portfolio are solving the same problems alone: securing rights to use customer data, assessing model risk, negotiating terms with the same AI vendors, finding people who can evaluate a model. The PMO can solve each once and make the answer available to all. It should also maintain an inventory of every AI model in production, with its business owner, the data it uses, its last retraining date and its next review date. As an auditor, I have asked for that inventory in organizations with dozens of models in production and been handed a spreadsheet assembled over a weekend. An organization that cannot list its models cannot govern them.
Who Runs This
None of this requires replacing the PMO’s staff. It requires changing its inputs, its outputs and the measure of its success. Add at least one person who can read a model evaluation and ask a data scientist the right follow-up question. Stop measuring the PMO by reports produced and start measuring it by decisions enabled: funding reallocated, initiatives stopped and shared problems solved once instead of 12 times.
What To Do Next
If your organization runs more than a handful of AI initiatives, here is where I would start:
1. Replace the portfolio status dashboard with a one-page evidence sheet per initiative, in one common format, and read them side by side at your next portfolio review.
2. Give the PMO the gate-review calendar and standing authority to put stop candidates on the executive agenda, and expect at least one at every meeting.
3. Build the inventory of every AI model in production: owner, data, last retraining, next review. If it takes more than a week to assemble, that is your first finding.
4. Pool the shared constraints: data access rights, model risk assessment, vendor terms. Assign the PMO to solve each once.
5. Change how you measure the PMO. Ask it, each quarter, what decisions its work made possible.
Executives often ask me whether the PMO will survive the AI era. I think that is the wrong question. The PMO will survive if it becomes the leadership team’s eyes on a portfolio that no single executive can see whole—and it will fade if it keeps producing reports that answer questions nobody is asking. Steering is harder than reporting. It is also the only version of the job worth funding.
Forbes Business Council is the foremost growth and networking organization for business owners and leaders. Do I qualify?











