Most AI budgets are set once a year, then run on inertia until half the portfolio is zombie pilots nobody will admit failed. The reason is simple: money follows momentum, not proof. A quarterly capital cadence fixes it by releasing funds in small tranches against exit criteria agreed before the work starts, then reallocating to whatever is proving out every ninety days. Because not advancing becomes the default outcome of a missed gate, killing a bet stops carrying stigma and capital finally moves toward evidence while it is still fresh. Here is how to build the loop.
Fund at the speed of proof, not the speed of enthusiasm
Most AI portfolios are funded once a year and then run on inertia. A pilot gets a number in the annual plan, the number becomes a floor no one revisits, and twelve months later the initiative is still burning cash with no decision ever made to continue or stop it. The money followed the original enthusiasm, not the accumulating evidence. By the time finance notices, half the portfolio is zombie pilots that are neither dead nor scaling, each defended by a sponsor who does not want to admit the thesis failed.
The fix is a quarterly capital cadence: release money in small tranches tied to milestones, judge each bet against exit criteria that were agreed before the work started, and reallocate freed capital to the bets that are proving out. A mid-size company running twelve AI initiatives at a combined 4 million dollars a year typically finds that three are working, five are ambiguous, and four should have been killed two quarters ago. A gated cadence surfaces that split every ninety days instead of once a year, so capital moves toward proof while the proof is still fresh.
Why annual budgeting fails AI specifically
Annual budgets assume you can forecast value at the point of funding. AI initiatives cannot be forecast that way, because the value depends on data readiness, adoption, and model behavior that are only knowable after you build. A cadence that re-decides quarterly matches the funding rhythm to the rate at which uncertainty actually resolves, which for most AI work is roughly one meaningful learning per quarter. It also changes the political economy of stopping. Under an annual budget, killing a bet means admitting a plan was wrong and surrendering the line item; under a quarterly cadence, not advancing is the default outcome of a missed gate, so stopping carries no stigma and the sponsor keeps standing to pitch the next idea. That single shift, from a kill being an exception to a kill being routine, is what actually frees capital.
An investment loop that repeats every quarter
Structure each bet as a stage with a named gate. The gate defines the metric threshold to advance, the evidence pack that proves it, the decision owner, and the size of the next tranche. The thresholds tighten as a bet moves from explore to pilot to scale, because the cost of a wrong bet rises with each stage. An investment committee should be able to read this table across a portfolio and decide continue, kill, or reallocate in a single ninety-minute review.
| Stage | Advance criteria | Evidence pack | Tranche size | Decision owner |
|---|---|---|---|---|
| Explore | Problem validated; data source exists and is accessible | Use-case brief; data availability check | 25k to 75k | Domain sponsor |
| Pilot | Model accuracy meets bar on held-out data; 10+ real users | Eval report; adoption log; risk note | 75k to 300k | AI product lead |
| Scale | Payback under 12 months; drift monitored; runbook exists | ROI model; monitoring dashboard; ops plan | 300k and up | Investment committee |
| Reallocate | A bet misses gate twice, or a peer bet exceeds its bar | Portfolio scorecard; freed-capital ledger | Up to 15% of pool | CFO co-sign |
| Kill | Thesis disproven, or no path to the accuracy or payback bar | Post-mortem; reusable-asset inventory | Recover residual | Investment committee |
A gate with no pre-agreed threshold is not a gate, it is a status meeting with better lighting.
A worked example
Take a document-classification bet funded at 60k to explore. At the Q1 gate it hit 84 percent accuracy against a 90 percent bar, so instead of killing it the committee released a reduced 120k pilot tranche conditioned on reaching 90 percent within the quarter. It hit 91 percent in Q2 with fourteen active users and an eight-month payback, cleared the scale gate, and drew a 350k tranche. The capital for that increase came from a chatbot bet that missed its adoption bar twice and was reallocated under the 15 percent rule. Same pool, better placement, decided in two reviews rather than argued for a year.
Standing up the cadence without freezing the portfolio
- Publish a one-page gate template per stage with quantitative thresholds, for example 90 percent accuracy to leave pilot and a payback under twelve months to reach scale, so decisions turn on numbers rather than narrative.
- Calendar the quarterly review in week ten or eleven, require evidence packs 72 hours ahead, and publish every continue, kill, or reallocate decision within 48 hours so the portfolio view is a single shared document.
- Release capital in tranches sized to the next milestone, never as a full-year allocation, so a stalled bet stops consuming cash the moment it stops producing proof.
- Pre-authorize a reallocation buffer of 10 to 15 percent that named decision owners can move between bets on gate outcomes, with CFO co-sign required only for larger shifts.
- Automate the evidence pack by pulling accuracy, adoption, and cost metrics from source systems, so the numbers in the review cannot be hand-massaged by the sponsor who wants the bet to survive.
How stage gates quietly stop working
- Vague thresholds that invite debate instead of decisions. Fix: express every advance criterion as a number on a shared instrument, not an adjective a sponsor can argue around.
- Funding tied to the calendar rather than the milestone. Fix: release the next tranche only when the prior milestone is evidenced, so elapsed time never substitutes for proof.
- Gates that only ever say continue. Fix: require the committee to kill or reallocate at least one bet per quarter, or the cadence is theater that never moves capital.
- One reallocation owner with no bound. Fix: cap the buffer at 10 to 15 percent and require CFO co-sign above it, so reallocation stays fast without becoming unaccountable.
- Uniform risk treatment across very different bets. Fix: tier the evidence pack by domain, so a regulated-decision bet carries a heavier risk note than an internal productivity tool.
Cadence moves for the next 30 to 90 days
- Publish gate criteria for explore, pilot, and scale as a single one-pager with the numeric thresholds filled in.
- Set the next quarterly review date and require evidence packs to arrive 72 hours before it.
- Enable a 10 to 15 percent reallocation buffer and name the decision owner who controls it.
- Convert this year's remaining AI allocations from annual lump sums into milestone-linked tranches.
- Run one live gate on your weakest current pilot and record a real kill or reallocate decision, not a deferral.