Rehearsing the Budget With Agents: The Behaviour Is Documented, the Simulator Is Not

The problem

In the last week of the United States federal fiscal year, agencies spend 4.9 times the weekly average of the rest of the year, and information technology projects bought in that week receive substantially lower quality ratings (Jeffrey Liebman and Neale Mahoney, 2017; measured from administrative procurement data, without a control group)¹. A summary by the American Economic Association puts it another way: 16.5 percent of all spending falls in the last month of the year and 8.7 percent in the last week, and 25.79 percent of projects started in the last week receive one of the two lowest quality scores, against 3.04 percent of projects started during the rest of the year².

This is a budget failing for behavioural rather than arithmetic reasons. The money was counted correctly; what went wrong was how people responded to a rule that unspent money disappears at year end.

The same pattern appears inside companies. Using daily transaction data on about $608.5 billion of spending by 539 public and private firms between 2010 and 2019, Paul Décaire and Denis Sosyura report that managers with a budget surplus raise spending in the final budgetary month by 44 percent, and that the average manager spends 88 percent of the surplus still available four weeks before year end (measured without a control group; working paper)³. They also find that managers who finish the year with money left get smaller budgets the following year, which is the incentive behind the spree, and that the effect is weaker where firms have strong owners such as blockholders and private equity investors³. Their data cover advertising budgets, so the size of the effect in other kinds of spending is not established.

The second behavioural failure is padding, known in management accounting as budgetary slack: asking for more than is needed, or committing to less than is achievable. Writing in 2001, Michael Jensen of Harvard Business School argued that corporate budgeting "encourages managers to lie and cheat, lowballing targets and inflating results"⁴. That is an argument from experience, not a measurement, and the research did not find a recent, openly readable study that measures how much slack firms carry.

“The money was counted correctly; what went wrong was how people responded to a rule that unspent money disappears at year end.”

Outside the United States, the European Bank for Reconstruction and Development reported in 2013 that governments in its countries of operation, where data were available, spent on average 76 percent more per month in December than between January and November (measured from government spending data, without a control group; figure more than three years old)⁵. Finland and Poland were among the exceptions, and countries rated lower on government effectiveness tended to show larger year-end increases⁵. Agenda Austria, a Vienna think tank, estimated in 2017 that the Austrian government pays out around 17 percent more than the annual monthly average in December (figure more than three years old)⁶.

Some governments have tried to remove the incentive rather than police the behaviour. In 2011 HM Treasury described its Budget Exchange scheme, which lets departments surrender an expected underspend before the financial year ends in return for a matching increase the following year, up to a limit; the scheme it replaced had led to accumulated stocks of around £20 billion across the UK government⁷. The American Economic Association summary of the federal study likewise points to the Department of Justice, which can roll over some of its revenue and keeps project quality steady at year end².

The common thread is that budgets are tested against arithmetic, not against the people who will live inside them. Departments respond to the rules, competitors respond to the plan, and customers respond to prices; none of that sits in the spreadsheet.

What already works

Human war games

Companies already rehearse reactions using people. In a 2011 McKinsey Quarterly article, John Horn described a two-day exercise in which "teams representing competitors and stakeholders role-played against a 'company' team"⁸. A human war game costs senior time, covers a handful of moves and cannot be replayed hundreds of times.

System dynamics

System dynamics models represent an organisation as stocks, flows and feedback loops, with fixed decision rules rather than adaptive actors. Its best-known teaching device, the beer distribution game, shows how far real decision-makers drift from what the arithmetic allows. In John Sterman's account, average team costs are about $2,000 against an optimum of about $200, "ten times greater than optimal" (measured in classroom experiments; figures more than three years old)⁹. His 1987 analysis of 44 players traced the gap to a specific habit: players "fail to account for control actions which have been initiated but not yet had their effect"¹⁰.

The lesson for budgeting is that the gap between plan and outcome came from predictable rules of thumb, which a model can represent once they are known.

Agent-based models calibrated on real data

An agent-based model simulates many individual actors, each following its own rules, and lets the overall outcome emerge from their interactions. Sebastian Poledna and colleagues, writing in the European Economic Review in 2023, built such a model of the Austrian economy with "millions of heterogeneous agents", populated from national accounts, input–output tables, government statistics and business demography data, and report that it "can compete with benchmark VAR and DSGE models in out-of-sample forecasting of macro variables" (VAR and DSGE are the standard statistical and theory-based forecasting models)¹¹. Out-of-sample means the model was tested on periods it had not been fitted to, which is the discipline a budget simulator would need.

Language-model agents that imitate people

A newer line of research uses large language models, the systems that generate text, as stand-ins for people. Joon Sung Park and colleagues built agents from two-hour interviews with 1,052 Americans and asked them General Social Survey questions the agents had not seen. Measured against each participant's own consistency when re-surveyed two weeks later, interview-based agents reached 83 percent of that benchmark and agents combining interviews and surveys reached 86 percent, compared with 74 percent for agents given only demographics (preprint, November 2024)¹².

In plain terms, an agent built from a long interview predicted a person's survey answers almost as well as that person's own answers two weeks later did; but survey questions are not budget decisions.

John Horton, Apostolos Filippas and Benjamin Manning put language-model agents through classic behavioural economics experiments and report results "qualitatively similar" to the original human studies (working paper, revised February 2026)¹³. CompeteAI, a research simulation, placed restaurant agents and customer agents in a virtual town and found that competition pushed restaurants to change strategy in ways the authors say "align well with existing market and sociological theories" (preprint, 2024)¹⁴.

What does not exist

The research found no published case of a company or government running its own budget through a simulation in which departments, competitors, customers and suppliers act as adaptive agents, and no study calibrating such a simulation against an organisation's past budget years. The building blocks sit in separate fields: documented budget behaviour in accounting and public economics, calibrated agent-based models in macroeconomics, and language-model agents in social science. Joining them for a single organisation's budget would have to be invented.

The possibility

Before a budget is approved, the finance team would run it through a simulated year. Each department would be an agent with its own goals: a head of marketing who protects next year's allocation, a head of operations who pads to avoid being caught short, a sales team that reaches its target by discounting in the last quarter. Outside the company, competitor agents would respond to price changes, customer agents to price and service, and supplier agents to volume commitments. The simulation would run the year many times, and the output would not be a single forecast but a map: where the plan breaks, which rules invite gaming, and which assumptions about rivals carry the most risk.

One way this could work is in three layers. The first is ordinary data plumbing: past budgets, actual spending by month and cost centre, prices, volumes and market data, drawn from finance and sales systems. The second is a rules engine, closer to system dynamics than to AI, that encodes known behaviours, such as year-end spending when a surplus remains and cuts for underspenders, with parameters estimated from the organisation's own history. The third is the AI layer: language-model agents that play the parts where fixed rules are too rigid, for example how a competitor might respond to an unexpected price cut, or how a department head might react to a new carry-forward rule.

Where AI fits. Machine learning could estimate behavioural parameters from past spending; generative agents could improvise reactions to scenarios with no history. The ledger, data feeds, scenario design and governance of results are ordinary software and finance work.

The role of AI would therefore be narrower than the idea first suggests. If the rules engine, calibrated on past years, explained most of the gap between budget and outturn, the language-model agents would add little and could be left out.

The most useful early version might not forecast at all. It could test budget rules rather than budget numbers: what happens to year-end spending if unspent funds can be partly carried forward, as in the UK scheme and the Department of Justice case⁷², or if underspenders stop being cut³. In these cases the direction of the effect is documented, and the simulation would only need to size it for one organisation.

What would have to be true

1. The behaviour would have to be visible in the data. Year-end spending can be measured from transaction dates, as the federal and corporate studies show¹³. Padding is harder: it is the gap between what was requested and what was needed, and the second number is rarely recorded.
2. The simulator would have to reproduce past years before anyone trusted a future one. A model calibrated on several past budget years would need to reproduce a held-back year, including its year-end pattern, in the way the Austrian agent-based model was tested out of sample¹¹. Without that test, the output is a story, not evidence.
3. Language-model agents would have to behave like the actual people, not like plausible people. Interview-based agents came close on survey answers¹², but no study was found showing agents reproducing a real manager's budgeting behaviour.
4. Competitor and customer agents would need ground truth. A firm sees its rivals' prices and its customers' purchases, not its rivals' internal decisions, so competitor agents could be checked only against observed market outcomes.
5. Leaders would have to accept results that implicate them. A simulation showing that one division games its targets is useful only if that division cannot veto the model.
6. The cost would have to fall below that of a human war game. A two-day exercise with senior teams⁸ sets the benchmark the simulation would have to beat, in money and in management time.

What could stop it

What the evidence shows

Language-model agents are unreliable at the level of individuals. Nikolay Petrov, Gregory Serapio-García and Jason Rentfrow found that when models were given the demographic profiles of real people, their answers to personality questionnaires showed "poor psychometrics properties", and concluded that the work "casts doubt on LLMs' ability to simulate individual-level human behaviour" (preprint, May 2024)¹⁵. CEO-Bench, a 2026 benchmark that tested five frontier models on 13 corporate resource-allocation scenarios, found that all produced structurally valid plans but diverged sharply on strategic calibration, with failure modes including "single-advisor capture, conservative default under ambiguity, and historical amnesia" (preprint, June 2026)¹⁶.

Human war games, which a simulation would partly replace, have documented limits of their own: they are most valuable "when a company has very few but discrete choices to test", work best at moderate uncertainty, and some stakeholders, such as legislators, are hard to portray⁸.

The supporting evidence for language-model simulation comes mostly from preprints and working papers^12131416, and none of these studies tested agents on budgeting decisions.

Reasoning

Would simpler remedies do the job? The documented budget failures already point to non-AI fixes: carry-forward limits, multi-year funding, and not cutting underspenders²³⁷. An organisation could adopt those directly, without a simulator. The simulation would add value only where several effects interact or are hard to size in advance.

Would it change the behaviour it models? If managers knew a simulator was predicting their padding, they could shape their requests to look normal to the model. Any behavioural model used to set budgets would invite the same gaming it was built to expose.

Who bears the risk? A plausible but uncalibrated simulation could give a finance team false confidence, and the cost of a wrong call would fall on the departments whose budgets were cut on its say-so. The more fluent the agents, the harder it would be to see that they are guessing.

Whose incentives work against it? Department heads benefit from slack and would have little reason to supply the honest data a simulator needs. Sellers of simulation tools would have reason to promote agent simulations before they had been checked against past budgets.

Is the AI needed at all? The beer game shows that much dysfunction follows from a few known rules of thumb⁹¹⁰. A rules-based model calibrated on an organisation's own history might capture most of the effect. If so, the case for generative agents would rest on the narrow area of reactions with no history, which is also the area hardest to validate.

Signs to watch

• A published backtest of a budget simulation, calibrated on past budget years, against a held-out year's actual spending, including its year-end pattern.
• Publication of the private-firm results of the Décaire and Sosyura study³.
• Peer-reviewed publication of the interview-based agent study¹² or of CEO-Bench¹⁶, and any change to their reported results.
• A finance ministry or treasury piloting agent-based or behavioural simulation in budget execution, or using one to evaluate a change to carry-forward rules.
• A study testing language-model agents against real managers' budget requests rather than survey answers or stylised games.