Close-up of green network cables plugged into server ports, showcasing technology setup.
FP&A

AI cost is landing in gross margin, not the IT budget

Inference, support agents and embedded copilots are migrating from fixed IT opex into cost of revenue, breaking the margin models FP&A teams are using to lock their 2027 plans.

The most consequential line item in the 2027 plan is not the headcount request or the software renewal. It is the quiet migration of AI spend out of IT opex and into cost of revenue, where inference calls, support agents and embedded copilots scale with customer activity rather than sitting inside a fixed envelope. AI cost in gross margin forecasting is no longer a technology question. It is a unit economics question, and most planning models still treat it as the former.

Why the 2027 plan is the one that breaks

Timing is the problem. CFO Dive, citing Deloitte research, reported that CFOs are bracing for AI costs to keep climbing into fiscal 2027. That is precisely the year most finance teams are locking down now, which means the steepest part of the cost curve lands inside a plan built on last cycle's margin assumptions.

The revenue side is not moving in lockstep. CFO Dive covered Intuit guiding to decelerating 2027 revenue even as AI products scale, a clean illustration of AI cost growth arriving ahead of AI revenue growth. Forrester's 2027 budget planning guidance, meanwhile, frames next year as the moment leaders stop being cautious and invest again. Put those together and you get plans built on spend-up assumptions against an uncertain top line.

The optimism is real too. A recent CFOtech item reports CFOs forecasting record profit growth attributed partly to AI returns. The gap between that confidence and the cost trajectory is exactly what makes this a planning question rather than a theoretical one.

Teams that leave AI in a fixed IT line will show a clean budget and a missed margin.

Fixed opex thinking versus variable cost to serve

When AI lived in an IT budget line, it behaved like software: negotiate once, amortize, forget. Inference does not behave that way. Every support ticket resolved by an agent, every copilot invocation inside a product workflow, every retrieval against a document store carries a marginal cost that rises with usage. If that usage correlates with customers, sessions or transactions, it belongs in cost of revenue, and the finance consequence is immediate.

The old planning instinct says gross margin improves with scale because fixed costs spread across more revenue. A volume-driven AI cost base inverts that logic for the portion of the stack that is genuinely variable. Growth now buys you more cost alongside more revenue, and whether margin expands depends on whether price and mix outrun usage per customer.

This also creates a classification risk that auditors and investors will eventually ask about. Teams that leave AI entirely in a fixed IT line will present a budget that looks disciplined and a gross margin that misses, because the variance will surface in a line nobody was watching.

Three model changes before the plan is signed

First, build a cost-to-serve unit metric at the workflow or active-user level. Not an allocation of total AI spend divided by customers, which hides everything, but a measured cost per meaningful action: per resolved ticket, per generated document, per automated reconciliation. That number is what tells you whether a pricing tier is profitable at volume.

Second, rebuild the gross margin bridge so AI usage is its own component. A bridge that separates price, mix and AI usage lets you answer the only question that matters in the board meeting: did margin move because we discounted, because the customer base shifted, or because people are using the expensive feature more than we modeled. Collapsing those three into one variance line makes the answer unavailable.

Third, change the reforecast trigger. Quarterly calendars are the wrong clock for a cost base that moves with usage. A trigger tied to inference volume, token consumption or agent session counts crossing a threshold gives FP&A a reason to reopen the model in week six rather than discovering the problem at quarter close.

The competition for the same opex dollar

AI run rate is not arriving in a vacuum. Payscale's 2027 salary budget work notes pay increases and inflation both rising, which means the merit pool and the AI envelope are competing inside the same operating expense line. Finance leaders who fund AI by quietly deferring compensation decisions are making a retention trade without labeling it as one.

Deloitte CFO Signals coverage and CFO Dive's reporting on the widening CFO mandate point the same direction: more scenario work, less tolerance for a single-point plan. The practical response is to model at least two AI usage paths, a base case tied to current adoption and an upside case where product adoption succeeds and cost to serve rises faster than price realization. The second scenario is the uncomfortable one, because commercial success is what creates the margin problem.

Key takeaways

  • AI spend is moving from IT opex into cost of revenue, making gross margin volume-driven rather than fixed.
  • Deloitte research reported by CFO Dive shows CFOs expect AI costs to rise into fiscal 2027, the year most plans are being locked now.
  • Build a cost-to-serve unit metric per workflow or active user, not a total-spend allocation.
  • Separate price, mix and AI usage in the gross margin bridge so variance is explainable.
  • Tie reforecast triggers to usage volume, not the quarterly calendar.