Most AI programs are reviewed in status meetings that optimize for reporting, not results, so decisions drift and capital stays stuck. Treat the portfolio like a factory instead: every agenda item is a decision ticket with a named owner, two or three costed options, and an evidence pack, moved through four standard gates. Instrument the flow with throughput and quality metrics such as time-to-decision, exception rate, and rework, then measure decisions closed per hour. That is how a review stops being theater and starts producing funded, sequenced execution.
Why status meetings quietly stall AI programs
Walk into most AI portfolio reviews and you will watch an hour dissolve into slide narration. A team reports that a pilot is "on track," another shares a demo, a third flags a vague concern, and the meeting ends with no decision recorded, no capital moved, and no owner on the hook. The program feels busy and produces nothing. The root cause is structural: the review is built to report progress, not to decide the next move, so the scarcest resources in the building, executive attention and unallocated budget, sit idle while work drifts.
An AI factory inverts that. A factory is judged on throughput and quality: how many good units leave the line per shift, and how many defects escape. Applied to a portfolio, the unit of output is a decision, not a status. A decision-centric review exists to close decisions, advance use cases through gates, and reallocate capital toward evidence. Consider a mid-market lender running eight AI initiatives. In its old weekly review it made roughly two real decisions per session and left six items in limbo. Rebuilt around decision tickets and gates, the same 60-minute meeting closed nine decisions and freed a stalled document-processing pilot that had waited five weeks for a go or kill. The meeting length did not change. Its purpose did. The gain did not come from working harder; it came from designing the session to emit decisions the way a line is designed to emit units, with a clear standard for what a finished decision looks like and a count of how many left the room.
The five parts of a decision-centric review
The operating model has five components. Each one converts a soft ritual into a measured process, and each carries a metric so you can tell whether the review is actually running like a factory or has slid back into theater.
| Component | What it produces | Instrumented metric |
|---|---|---|
| Decision tickets | Each item names one decision owner and 2 to 3 costed options with trade-offs, no "for awareness" slots | Percent of agenda that is decidable in the timebox (target 90%+) |
| Evidence packs | A one-page packet of metrics, user feedback, and risk notes, linked not narrated | Time-to-decision from ticket opened to closed (target under 2 weeks) |
| Stage gates | Explore, Pilot, Scale, Operate, each with pre-agreed pass criteria and reallocation rights | Gate pass rate and average days a use case waits at a gate |
| Defect review | Weekly log of quality, latency, and policy defects with owners and due dates | Exception rate and rework rate on shipped work (target trending down) |
| Decision log | A versioned record of decision, rationale, owner, and next check date | Decisions closed per review hour (target 6 to 10) |
Read the table as one connected line. Tickets make items decidable, packs give the evidence to decide, gates route the decision, defect review protects quality, and the log makes every choice traceable. When all five run, the review produces a countable output. When one drops, the metric attached to it exposes the gap before the program silently stalls. In practice, the fastest diagnostic is the last row: if a 60-minute review is closing fewer than six decisions, one of the upstream components has quietly broken, and the metric on that row tells you which. A low decidable-agenda percentage means tickets are still status in disguise; a long time-to-decision means evidence packs are missing or bloated; a rising exception rate means quality is leaking faster than the gates can catch it.
Rebuild the review in four weeks
- Replace the status agenda with decision tickets: a one-page template of owner, 2 to 3 options with cost and risk, evidence link, current gate, and the specific ask. Cap the agenda to what the timebox can actually decide.
- Publish objective gate criteria for Explore, Pilot, Scale, and Operate, and tie each gate to a funding tranche so passing a gate releases the next block of capital automatically.
- Instrument the work now: track answerability percent, exception percent, time-to-decision, and rework rate, and post the numbers before every session so evidence, not opinion, drives the call.
- Assign four standing roles: a chair who keeps pace, a scribe who updates the decision log live, guardians for risk and compliance, and the product owners who bring the tickets.
- Set a layered cadence: a weekly decision review, a monthly portfolio re-rank, and a quarterly capital reset, so short-cycle decisions and long-cycle reallocation each have a home.
Where the factory slips back into theater
- Status creep, where narrated slides return and decisions vanish. Fix: make options and an evidence link mandatory fields, and let the chair reject any ticket that lacks them before it reaches the agenda.
- Vague gates, where "looks good" replaces a threshold. Fix: write pass or fail criteria as numbers, for example answerability above 85 percent and exception rate below 5 percent, and refuse to advance on sentiment.
- Unowned defects that never close. Fix: every defect gets a named owner and a due date in the log, and open defects past due are read aloud at the top of the next review.
- An evidence pack that becomes a deck. Fix: hard-cap it at one page of links and metrics; if it needs narration to be understood, the ticket is not ready to decide.
- A dead decision log nobody reads. Fix: link it from every calendar invite, open each session by reviewing decisions due for a check, and make the scribe update it in the room, not afterward.
What to ship before your next review
- Draft a one-page decision ticket template with owner, options, evidence, gate, and ask, and require it for every agenda item.
- Stand up a shared, versioned decision log and link it directly from the meeting invite.
- Instrument time-to-decision and decisions closed per hour, and publish both numbers weekly.
- Write pass or fail criteria for all four gates and attach each gate to a funding tranche.
- Name the chair, scribe, guardians, and product owners in writing so the roles are filled before the meeting starts.