The 2027 budget spreadsheet has one AI line, labeled "AI tools," and someone sized it by multiplying seats by the 2026 license price. That line will be wrong before the second quarter, and price will not be the reason.
Most budgets answer "how much should we spend on AI?" The better question is which work the money changes, and what has to be true underneath that work for it to pay off. Our answer: fund AI as four lines, not one (usage, review capacity, foundation work, and one scoped pilot), protect the maintenance budget instead of raiding it, and defer platform commitments until a use case earns them. We run AI delivery pipelines for clients and for our own operations, and these four lines are where the money went in 2026.
What should a 2027 AI budget include?
A 2027 AI budget should separate the cost of using AI from the cost of making it work. Seat licenses are the visible part. The spending that decides whether AI pays off sits in the other three lines.
Usage, sized from logs, not seats. Token and API spend tracks how the work is built: how much context each request carries, how often a task retries, how many files an agent reads again. In our own delivery work, spend moved more with those choices than with headcount. Pull three months of real usage before you set the number. Then put an alert on it. AWS Budgets can notify you when spend exceeds, or is forecast to exceed, the amount you set. We explain why the bill is a design question in your AI bill is an architecture problem.
Review capacity. Someone has to read what the agents produce before it ships. That is senior engineering time, and it rarely appears in an AI budget because it looks like existing payroll. It is not existing capacity if those people were already full. Budget the hours as their own line. AI coding moved the bottleneck to code review, and the staffing has to follow it.
Foundation work. An agent is only as useful as the system it works in. Missing tests, stale dependencies, and service accounts with years of piled-up permissions all make AI output slower to trust. Fund the cleanup as a separate line with a named owner. Fix the foundation before you bolt on the agent covers what that work includes.
One scoped pilot with a kill date. Pick one workflow, one success measure, and one date when you either expand it or stop it. Spread the same money across five experiments and you get five demos. Our guide to running an AI pilot that doesn't waste the budget covers the scoping.
What not to cut to pay for AI
Do not fund AI by cutting maintenance. It is the trade we see proposed most often, and it works against the AI spend directly, because the agents operate inside the systems you stopped maintaining. We made the full case in why cutting maintenance to fund AI backfires.
Unsupported runtimes are the concrete version of this. Adobe ColdFusion 2021 reached the end of extended support on November 10, 2026, and its final update shipped on December 9, 2025. If a ColdFusion 2021 server still runs in 2027, the upgrade belongs in the budget ahead of any AI feature that touches that application. What ColdFusion 2021 end of support means lays out the first steps. The same logic applies to any runtime or framework that left support in 2026.
What to defer until a use case earns it
Defer any commitment that assumes a use case you have not proven yet.
- Company-wide licenses before adoption data. Buy for the teams that use the tool. Expand when usage says so, not enthusiasm.
- Your own agent platform. Most teams need one working workflow first. Build vs. buy a coding agent gives four questions to answer before you build.
- Custom model training. Retrieval over your own documents, plus a good evaluation set, covers most of what teams expect training to deliver.
How do you present an AI budget that survives the year?
Give every line an owner, a measure, and a review date. The memo that survives the first quarter says what each line buys and when someone will check. Schedule a quarterly review where money can move between the four lines. Usage will run over in one place and under in another, and a budget that cannot move becomes a budget people work around.
If you want a second look at the plan before it locks, our AI implementation framework starts with the same four lines and the foundation work under them.