Zero-based budgeting asks every cost center to justify its spend from zero each cycle rather than inflating last year’s number. AI does not decide what survives. It does the assembly: pulling actuals from the ledger, grouping them into cost packages, attaching the contracts and drivers behind each one, and flagging what looks off. Owners still argue for every line.
That split matters, because the reason ZBB dies inside most finance teams is not disagreement about priorities. It is fatigue. Rebuilding a budget from the ground up means someone reconciles thousands of transactions into coherent decision packages, tracks down which contract governs which recurring charge, and answers the same question for the fortieth cost center at eleven at night. The analysis is worth doing. The clerical work in front of it is what makes people quietly revert to a growth-rate-on-last-year spreadsheet the following year.
Where the clock actually goes
Walk through a real ZBB cycle and the time does not sit where the methodology books put it. Very little of it is spent on the interesting question of whether a cost is worth its return. Most of it is spent getting to the point where that question can even be asked.
- Actuals arrive from the ledger at the account level, but a decision package is a business unit of spend. Mapping one to the other is manual and inconsistent, and two analysts will bucket the same vendor differently.
- Recurring charges have a contract behind them somewhere. Finding it, confirming the renewal date, and confirming whether the charge is committed or discretionary is a scavenger hunt across procurement systems and shared drives.
- The same supplier shows up as three spellings across three entities, so spend that should aggregate into one package fragments into three.
- Half the “zero-based” review ends up being a reconciliation exercise, because the numbers pulled for the exercise do not tie to the reported P&L and nobody trusts a package total that disagrees with the ledger.
This is entity resolution, reconciliation, and lineage work. It is exactly the kind of assembly a model does well and a person does slowly. The judgment about whether a line deserves funding is the part you want your people rested and focused for, and it is the part they never reach with energy left.
What the model assembles, and how you keep it honest
The unit of automation is the decision package: a cost, its history, the driver that explains its size, the contract that commits it, and the owner accountable for it. The model builds that package and hands it over. It does not rank packages and it does not recommend cuts.
Building it reliably takes a few controls that are not optional.
- Point-in-time correctness. When the model reconstructs last year’s spend to justify this year’s zero base, it must use the ledger as it stood at close, not as it reads today after reclassifications and late accruals. Pull the current view and you leak information backward, and the baseline you are supposedly rebuilding from is quietly wrong.
- Lineage on every number. Each figure in a package traces to the transactions or contract lines that produced it. An analyst who distrusts a total can expand it to the source rows in one step. A number the model cannot source is dropped, not shown, because a confident wrong number is worse than a visible gap.
- Reconciliation before review. Package totals sum back to the general ledger before a human sees them. If the packages for a cost center do not tie to that center’s reported cost, the cycle stops there. Nobody should spend an afternoon debating a package that was never going to add up.
- Entity resolution on suppliers and cost centers. Vendors, contracts, and internal cost centers get resolved to canonical entities so spend aggregates correctly. This is the difference between one honest package and three misleading fragments, and it is where quiet errors hide.
Driver analysis is the piece that turns a pile of costs into something an owner can argue about. For each package the model attaches the driver that best explains the spend: headcount for a team’s tooling, transaction volume for payment processing fees, seat count for software. It reports the relationship and stops there. When a cost has grown faster than its driver, that gap is the conversation. Cloud spend up forty percent against a workload that grew eight does not settle anything on its own. It lands on the desk of the person who has to explain it, with the number already attached.
Keeping owners in the loop, and keeping an audit trail
ZBB only works because someone owns each cost and has to defend it. Automation that removes the owner removes the point. So the model’s output routes to the accountable person, who confirms the package, adjusts the driver if the model picked a weak one, and records the justification in their own words. The system captures who approved which package, when, against which version of the numbers.
That audit trail earns its keep twice. It is what a controller or an auditor needs to see that the budget was actually rebuilt rather than rolled forward with a fresh coat of paint. And it becomes the eval set for next cycle. Where owners consistently overrode the model’s driver choice, that is a labeled correction you feed back in. Where the model flagged a gap that turned out to be a timing artifact, that is a false positive you tune against, because a review process that cries wolf gets ignored inside two cycles.
Watch for drift between cycles. Charts of accounts get restructured, cost centers merge, a new ERP lands. The mapping that resolved spend into packages cleanly last year degrades silently against this year’s structure. Treat the mapping as versioned and monitored, not as a one-time setup, and check it against a sample of reconciled packages before you trust a full run.
The honest summary is narrow. AI makes zero-based budgeting affordable to do properly, and to do more than once a year, by taking the reconciliation and assembly off your people. It does not make the decisions, and you should be suspicious of anyone selling it as though it does. The cut still belongs to whoever has to live with it.
FAQ
Does AI decide which costs to cut in a zero-based budget?
No. The model assembles the evidence for each cost package and flags anomalies, but a named owner justifies and defends every line. The audit trail records who approved what, not what the model suggested.
How do you stop the model from double-counting or inventing spend?
Every figure it surfaces carries lineage back to a source transaction or contract, and package totals are reconciled against the general ledger before anyone reviews them. A number without a source is treated as a defect, not a value.
Is zero-based budgeting worth automating if we only do it once a year?
The assembly work is the same whether you rebuild annually or quarterly. Automating it is what makes a more frequent cycle affordable, which is usually the point of adopting ZBB in the first place.