If there's one cake and everyone asks for a slice, people ask for big slices because they know they'll get cut. Then everyone gets cut, and the asking gets bigger next time.
In most budget rounds, teams expect their request to be trimmed, so they ask for more. Finance expects inflated requests, so it trims. Both are acting sensibly, and the result is a lot of effort and numbers nobody trusts. Changing a few rules can make honest requests the sensible choice.
Budget allocation with expected haircuts is a game whose equilibrium is inflated bids and across-the-board cuts. Use-it-or-lose-it adds year-end spending. Mechanisms that help: allocate against decision-level evidence, keep a central reserve with transparent draw-down, reward forecast accuracy, and let underspend carry or pool.
The game
Each team lead submits a budget request. Finance has less money than the total asked for, so it trims requests. Everyone knows this happens. A team lead who asks for exactly what they need will be trimmed below it, so the sensible move is to ask for more. Finance knows requests are inflated, so it trims harder. Next year, team leads inflate more.
Nobody in this game is behaving badly. Each person is responding sensibly to what they expect the others to do. That's what makes it a game in the technical sense, and also why lecturing people about honest budgeting rarely works, much like estimate padding.
The year-end twist
Many organisations add a second rule: money not spent by year-end is lost, and a team that underspends may get less next year. That gives every team a reason to spend what's left in the final months, whether or not the spending is the best use. It also hides real need, because a team that spent its whole budget looks as if it needed all of it.
Why AI budgets make it worse
AI spending adds uncertainty. A team asking for Copilot Credits or agent capacity often can't say how much it will use, because nobody has used it before. The safe move is to ask for a lot. With usage-based billing, unused capacity may simply not be spent, but the request still blocks money other teams could have used.
Rules that change the game
These rules don't rely on anyone being more honest. They change what's sensible.
| Rule | Why it helps |
|---|---|
| Fund decisions, with evidence | Requests tied to specific decisions, with volumes from the decision inventory, are harder to inflate and easier to check |
| Keep a central reserve | A shared reserve that teams can draw on openly reduces the need for each team to carry its own buffer |
| Reward forecast accuracy | Track requested against actual spend. Teams with accurate forecasts get faster approval next time |
| Let underspend carry or pool | Remove the year-end spend reason. See Who gets the pooled allowance? A fair rule |
| Ask for ranges | A range with a confidence level is more honest and can be checked, as in Forecasting AI spend as a range |
A worked example
The numbers are illustrative. Four teams share a £100,000 AI budget. Each expects a 20% cut, so each adds 25% to its real need.
| Team | Real need | Request | After 20% cut |
|---|---|---|---|
| A | £30,000 | £37,500 | £30,000 |
| B | £25,000 | £31,250 | £25,000 |
| C | £20,000 | £25,000 | £20,000 |
| D | £30,000 | £37,500 | £30,000 |
| Total | £105,000 | £131,250 | £105,000 |
The requests total £131,250 against a budget of £100,000, so finance needs a cut of about 24%, more than the 20% everyone expected. Every team ends up below its real need, and each concludes it should inflate by more next year.
Now the same round with a central reserve. Suppose about £25,000 of the £105,000 real need is uncertain AI use that may or may not happen. Teams request only their committed need, £80,000 in total, and finance holds the remaining £20,000 as a reserve, released during the year when a team shows the usage. Nobody needs to inflate, because the reserve covers the uncertainty once instead of four times.
Where I got stuck
The central reserve works only if teams trust that they'll get money from it when they need it. If the first request is refused or delayed, the old behaviour comes back. I think the reserve needs a published rule for drawing on it, decided in advance, and I haven't yet written one I'm happy with.
Sources
This article applies standard game-theory reasoning about strategic responses to incentives. It uses no external facts or figures; the budget example is illustrative.