Logiciel Solutions Contact Us
Success Stories Tech News Contact Us
whitepaper

Twelve Annual AI Budgets Are Twelve Bets With No Exit Right.

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.

In depth

Twelve Initiatives, One Approval, Twelve Months.

01

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.

02

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.

The detail

What Separates A Portfolio From A Wish List.

Zone · 01

Options Bought

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.

Zone · 02

Kill Criteria Signed In Advance

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.

Zone · 03

Run Cost Modelled From Day One

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.

By the numbers

The figures that make it a board-level conversation.

25%
of AI initiatives have delivered the return expected of them on executives' own assessment, with around 16% scaled enterprise-wide
30%
of generative AI projects are abandoned after the proof of concept, on data quality, escalating cost and unclear business value
16%
of organisations produce regular reports to the CFO on the value their generative AI work creates. The other 84% are funding on narrative
40%
of agentic AI projects are expected to be cancelled by 2027
Inside the report

What you'll take away.

01

gate zero, cost the problem

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.

02

gate one, prove the technique offline

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.

03

gate two, measure a real delta

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.

04

gate three, fund the 3am owner

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.

Questions

Frequently asked.

Does tranching not just add four approval committees to every project?

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.

How do we set a kill rate without the portfolio looking like a failure?

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.

Should AI spend be capitalised or expensed?

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.

Is a first tranche of 5% too small for a sponsor to prove anything with?
Why report cost per outcome when the board asks for total AI spend?

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.

Get the whitepaper

Have it emailed to you.

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.

Download whitepaper
Next step

Buy the option first.

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