Budget season: fund AI by the work each team does, not by the seat

Planning season puts a VP in the usual spot. Each team sends up its own request for AI tools, written its own way. The executive team wants one number and a reason to believe it. You sit between them, and this year the numbers are moving in two directions at once.

Total spend is rising fast

CIO.com reported this week on Gartner’s latest forecast. Worldwide AI spending is expected to reach $2.7 trillion in 2026, up 49.5%. Within that total, spending on AI agents and assistants is forecast to grow 77.3%, AI security spending nearly doubles, and Gartner expects another 36.2% growth in 2027.

By 2030, Gartner’s John-David Lovelock said, “every dollar is going to be an AI dollar in one way or another.”

For a VP, the useful part is where the growth sits. Spending on agents and assistants is forecast to grow faster than AI software or AI infrastructure, and agents are billed differently from the tools most budgets were built around.

Unit prices are falling at the same time

On September 22, MacRumors reported OpenAI’s new GPT-6 Sol and Luna models. Sol is priced at $2 per million input tokens and $10 per million output tokens, which MacRumors describes as half its predecessor’s price. Luna fell from $0.20 and $1.20 to $0.10 and $0.50. OpenAI also says cached input reads get a 90 percent discount.

Cheaper units do not guarantee a smaller bill. An agent that works on a task for hours, re-reads its context and calls tools along the way can use far more units than a person typing into a chat window. And a lower price per unit makes it tempting to hand agents more work. The price cut is real; whether your total goes up or down depends on how much work you hand over and how you control it.

That is why a seat count is the wrong frame for most of next year’s asks.

Five lines for every team’s request

Ask each manager to answer the same five lines before an AI request reaches your budget. It turns a pile of different asks into a portfolio you can compare.

1. The work. One sentence on the recurring task, such as triaging inbound tickets or drafting the monthly forecast notes.
2. The volume. How many times a month that task happens today.
3. The unit cost. What one run costs in model usage and tools, measured in a pilot rather than taken from a price page.
4. The cap. The monthly ceiling, and who gets alerted when usage reaches 80% of it.
5. The owner. The named person who reviews the output and can switch the agent off.

A request without a measured unit cost goes back as a pilot, not a budget line.

A worked example

The numbers below are hypothetical. Replace them with your teams’ measured figures.

Your support team wants an agent to triage inbound tickets. In a two-week pilot it handled 1,500 tickets, and the usage bill for that period was $60. That is 4 cents a ticket. At 3,000 tickets a month, the run rate is $120 a month.

Now suppose the team switches on a feature that re-reads each customer’s full history for every ticket, and the cost per ticket triples. Volume is unchanged, and the line becomes $360 a month. Nothing is wrong with the agent; the design changed the unit cost. The cap and the owner are what catch it in the first week instead of at quarter end.

Seat licenses still make sense for tools people use at their own desks, like a writing assistant. Budget those per person. Budget agents per unit of work.

Brief up, explain down

Up. The executive team needs one page: spend by team, cost per unit of work, what each line replaced or sped up, and the caps. Two or three lines of each are enough. They are funding a portfolio, not reviewing every tool.

Down. Managers need three things: the cap for their team, how usage will be reported back to them, and what happens when they approach it. Say plainly that falling prices do not raise the cap automatically. A good pilot result is how a team earns a bigger one.

What to do this week

• Send your managers the five lines above and ask for answers before the next planning meeting.
• Ask finance whether your current reporting can show AI usage by team. If it cannot, that is the first thing to fix.
• Pick one agent already running in your organization and work out its cost per unit of work. That is your worked example for the executive team.
• Put caps and alert thresholds on anything billed by usage before the quarter starts, not after the first invoice.

Forecasts and prices will keep moving. A budget built on units of work and named owners holds up when they do.