Your business is going to spend more on AI this year. Three-quarters of finance leaders raised technology budgets for 2026, nearly half of them by 10% or more, and financial services firms led every other sector with increases around 15%.
The decision to spend is made. The harder question is the one your budget process is not built to answer: How do you forecast a cost that changes based on how your people use a tool?
Most firms get this wrong in two specific ways. Let’s review both before you build your AI budget.
Sign up for Kiplinger’s Free Newsletters
Profit and prosper with the best of expert advice on investing, taxes, retirement, personal finance and more – straight to your e-mail.
Profit and prosper with the best of expert advice – straight to your e-mail.
Mistake one: Budgeting AI as a fixed cost
You are used to software that costs the same every month. You buy a number of seats, you pay a flat fee, and the bill is the bill.
AI does not work that way. The cost tracks consumption, not headcount. Two advisers on the same license can generate wildly different bills because one runs long reports through the tool all day, and the other asks it a question twice a week.
This breaks the annual budget. You cannot set one number in January and hold it, because usage climbs as your people get better at the AI tool and find more uses for it.
A fixed line item guarantees one of two outcomes. You overfund it and waste capital you could have deployed elsewhere, or you underfund it and face an overage conversation in a near future quarter.
Treat AI the way you treat a variable operating cost, not a license. Build it as a range with a floor and a ceiling, then forecast on a rolling basis and revise quarterly. A firm running this as a static annual figure is budgeting a variable cost with a fixed budget, and the budget will lose.
Mistake two: Budgeting only the token bill
The number the vendor invoices you is the visible part of the cost. It is not the whole cost.
Every AI workflow carries expenses that never appear on the vendor’s bill, namely the human in the loop.
- Someone must review the output before it reaches a client, because a fiduciary cannot send unreviewed AI work product to the people who trust the firm with their money
- Someone must train staff on the tool
- Someone must maintain the governance, the acceptable use policy, the compliance records that an examiner will ask for
These are real costs, they scale with adoption, and they belong in the same budget line as the tokens.
This is why ownership matters as much as the number. Research on AI return shows that firms where technology teams own AI spend alone capture less value than firms where finance and compliance share the decision.
The token bill is a technology expense (although it should be a business expense). The review time and the regulatory exposure are not. Budget them together or you will underfund the part that keeps you out of trouble.
Budget against honest return, not the promise
Here is the number that should govern how aggressively you fund this. Less than 1% of executives report AI returns of 20% or greater, and a majority report returns in the range of 1% to 5%.
Meanwhile, Gartner expects 30% of generative AI projects to be abandoned after the proof of concept stage. The spending is real. The proven return, for most firms, is not yet.
This does not argue for sitting out. It argues for funding against measured outcomes rather than the vendor’s promise. Tie each AI budget line to a specific result you can measure, such as hours saved in a named workflow or a reduction in a particular operational cost.
Fund the use cases that clear that bar and reject the ones that do not. The firms that win the next three years will be the ones that fund AI where the business case is strong and measurable and refuse to fund it anywhere else.
How to build the number
Start with a pilot you meter. Run one real workflow through the tool for a full month and read the bill. That gives you a true cost per task, which is the only honest input to a forecast. Multiply by realistic adoption, not best-case adoption. Then add the costs the vendor never invoices: The time to review output, the time to train staff, the work to maintain governance.
Build the result as a range. Set a floor that covers committed usage and a ceiling that absorbs the growth you know is coming. Attach a pay-as-you-go overflow so that crossing the ceiling slows you down rather than cutting off a workflow your advisers now depend on. Then revisit the whole figure every quarter, because the underlying prices are moving and your usage is moving faster.
AI is now a permanent line in your operating budget. Treat it like one. Forecast it like a variable cost, fund it against measured return, and revise it on a schedule. The CFO who does this will deploy capital where it earns its keep. The CFO who sets a fixed number in January will spend the year explaining variances.
