Most AI budgets are approved once, for a year, with no scheduled decision point between the cheque and the post-mortem. This framework sets out four gates and four tranches at 5%, 15%, 30% and 50% of lifetime budget, and the kill criteria a sponsor signs before build money moves.
The trap most investment committees walk into: approve twelve initiatives once, for twelve months, then meet the real decision point at next year's planning cycle, by which time fifteen of fifteen are still nominally alive, five to eight should have been stopped, and the eleventh dormant pilot is delaying the two that worked.
What the better investment committees do instead: release capital in tranches of roughly 5%, 15%, 30% and 50% of lifetime budget, each gate converting exactly one assumption into evidence a controller accepts, with the kill criterion (metric, threshold, date, named decision maker) signed by the sponsor before any build money moves.
Not Projects
Venture investors are not better at picking winners than investment committees are. They are better structured: capital goes in as small commitments, each priced by what the last tranche proved, with the right to stop at every round. Annual funding gives that right away on the day it is approved.
A metric, a threshold, a date and a named decision maker. Extraction accuracy on the held-out invoice set below 92% by 14 March, decision the Group Financial Controller. Signed before Gate 1 money moves, a stop executes the sponsor's own prior decision rather than defeating them in a meeting.
Traditional software gets cheaper per user. Inference tracks usage, so a successful rollout is a cost event. Four lines go missing: evaluation and monitoring at 15% to 25% of build cost a year, scheduled version migration, funded human review, and compliance overhead that grows with models in production.
Roughly 5% of lifetime budget proves a costed problem, an instrumented process and a named profit and loss owner who signs for the benefit. Release on baseline volume, cycle time, error rate and loaded cost, pulled from the operational system and signed by finance.
At 15%, an offline evaluation on a held-out set the team did not build on, plus cost per transaction at forecast volume rather than at pilot volume. A quality bar cleared at a unit cost that still works when the volume arrives.
At 30%, real users doing real work against something that is not the new system. A holdout or staggered rollout across one business cycle, with the measured delta and the run cost reported in the same paper rather than in separate ones.
The final 50% needs unit economics that hold at volume, a twelve-month run-cost forecast, a support model, and evaluation coverage that survives a provider version change without a rebuild. Plus a named owner for the night it breaks.
It replaces one uncapped approval with four dated decisions that already have their evidence defined. Gate 0 and Gate 1 are paper reviews of a signed baseline and an offline evaluation. The committee time goes down, because
nobody is arguing about what would count as proof.
Report kills as capital returned, in the same table as capital deployed. Five to eight stops out of fifteen initiatives is the base rate for this asset class, so a portfolio with no stops is the one that needs explaining, not the one with
seven.
It sits across the boundary, so agree treatment with your auditor before the first tranche. Initial development can often be argued into capex once authorised, continuous retraining looks like maintenance, and per-token inference is operating cost. This is not accounting advice and frameworks differ.
Total spend rises for both a working programme and a failing one, because inference cost tracks usage. Falling cost per processed invoice while volume rises is a working programme. The two cases are indistinguishable in a total.
Drop your details and we'll send Twelve Annual AI Budgets Are Twelve Bets With No Exit Right straight to your inbox - no spam, unsubscribe anytime.
Bring your current AI budget to a portfolio review with our engineering leads and we will map each initiative to a gate, a tranche and a kill rule.
Book a portfolio review