How to forecast AI spend in your FY2027 budget

Forecast AI spend line by line from the driver behind each cost, instead of putting one growth rate on the total. Seats come from the headcount plan, internal usage from active users and how heavily they use the tools, customer-facing AI from customer activity, and agents from the number of workflow runs. Then put tighter planning ranges on the predictable lines (seats) and wider ones on the usage-driven lines (agents and new workloads).

That's more work than adding 25% to this year's number, but a blanket growth rate misses in both directions: it over-budgets seats and badly under-budgets agents. And AI is now big enough to matter. In BCG's AI Radar 2026, companies said they plan to spend about 1.7% of revenue on AI in 2026, roughly double the year before, so the FY2027 AI line is one your board will read closely.

Bottom line: Start from one AI spend number reconciled to the general ledger, split it into seats, internal usage, customer-facing AI, agents and supporting infrastructure, and forecast each from its own driver. Most of the uncertainty sits in usage and new agent workloads, so that's where the planning range should be widest.

Why one growth rate on AI spend gets it wrong

Because each AI cost moves with a different part of the business. Say you put 25% growth on the whole AI budget. The drivers underneath might say something very different: seats grow 10% with the headcount plan, customer-facing AI grows 40% with customer activity, and agent spend grows 80% as workflows move into production. The blanket rate over-budgets the first line and under-budgets the other two, badly in the case of agents.

The totals can even land close together while every line is wrong. Here's an illustrative company with $1.2M of AI spend this year:

• Seats and subscriptions, $480K this year: +10% with headcount gives $528K. A blanket +25% gives $600K, so $72K over.
• General internal usage, $120K: +20% with adoption gives $144K. Blanket gives $150K, $6K over.
• Customer-facing AI, $300K: +40% with customer activity gives $420K. Blanket gives $375K, $45K under.
• Agents and workflows, $150K: +80% with workflow runs gives $270K. Blanket gives $187.5K, $82.5K under.
• Supporting infrastructure, $150K: +35% with the workloads it serves gives $202.5K. Blanket gives $187.5K, $15K under.

The driver-based total is about $1.56M against $1.5M for the blanket rate, only 4% apart. But $78K of seat and usage budget is sitting in the wrong place and $142.5K of customer and agent cost is missing, so the variance shows up in the wrong owners' P&Ls all year. The numbers here are illustrative, not benchmarks.

Start from one reconciled AI spend number

You can't forecast what you can't count. Before building drivers, get this year's AI spend into one total that ties back to the general ledger: provider exports tell you what was consumed, and the GL tells you what was booked in each period. Reconcile the exports to the ledger, not the other way round.

The spend hides in more places than the obvious vendor invoices. Look in six places:

• Seats and subscriptions for employee tools, in AP and software invoices.
• API and model usage, in each provider's billing or usage console.
• Coding tools and development, in engineering software, cloud and expense data.
• Other AI tools, in AP, procurement and employee cards.
• AI infrastructure (compute, storage and data platforms), in cloud and warehouse bills.
• Agents and workflows, in agent platforms, workflow logs and provider usage.

The categories someone has to assemble by hand are the ones that get missed. In research by Mavvrik with Benchmarkit cited in our budgeting for AI spend guide, 67% of companies include public cloud in their AI cost reporting, but only 42% include AI coding tools and 36% include AI agents. AI bundled into software you already buy, as credits or per-action charges, is even easier to miss. Our step-by-step guide to getting spend data out of Claude, ChatGPT and Cursor covers the exports.

How to forecast AI spend line by line

Forecast each line from the business activity that moves it, and set the planning range by how much that driver can swing.

A few notes on each driver:

• Seats are the easy line. Take the headcount plan by department, multiply by the share of each team that gets a seat and the tier, and add known price changes at renewal. If you plan hires by start date, seats follow them; our guide to reconciling the headcount plan with actuals covers the timing.
• Internal usage follows adoption, and adoption is concentrated: the ebook cites LayerX data where the median user had 12 AI conversations and the top 5% had 144 or more. Forecast power users separately, because a few people moving teams can move the line.
• Customer-facing AI belongs with cost of revenue and should come straight from the revenue model's volume drivers (accounts, transactions or documents processed) times a unit cost you sample from recent months.
• Agents are the hardest line. Forecast runs per workflow per month and a sampled cost per run, then add each new workflow from the month it's planned to go live. A pilot that moves into production can multiply this line within a quarter.

How much detail does your AI forecast need?

Match the effort to how material AI is. Three levels cover most companies:

1. AI is a small line. Roll this year's run-rate forward: current annualized spend, plus known additions (new hires using AI at the same rate, a new tool, a pilot going live), minus known reductions (unused seats, retired tools), then a reasonable range around the result. That's still better than a flat 3%.
2. AI is a material budget line. Forecast each line from its driver, as in the table above.
3. AI is core to the product or operating model. Add low, base and high cases by line, a named owner for each driver, and monthly refreshes of the assumptions that move (pricing, adoption, consumption and new workloads).

Most mid-market companies building an FY2027 budget now sit at level 2. If AI spend is still small at your company, don't build a level 3 model for it: false precision costs more board trust than a directional answer with a clear note on what's in and out.

How to set low, base and high cases for AI spend

Build the range line by line, then sum it, rather than putting one percentage band around the total. Using the illustrative $1.2M company from earlier and the planning ranges in the table:

• Seats: base $528K, range about $502K to $554K (±5%).
• Internal usage: base $144K, range about $115K to $173K (±20%).
• Customer-facing AI: base $420K, moving with your revenue cases. If the revenue plan's low and high cases sit 10% either side of base, this line runs about $378K to $462K.
• Agents: base $270K, range $135K to $540K (−50% to +100%), because new workloads are the least predictable spend in the budget.
• Infrastructure: base about $203K, moving with customer AI and agents.

Two things stand out. First, the agent line alone accounts for more than half the total range, which tells the board exactly where the uncertainty lives. Second, adding every line's low case or every line's high case overstates the spread, because lines rarely all miss in the same direction. Present the base case, a realistic low and high, and name the one or two lines that drive the width. If the high case would breach a spending limit, agree in advance what triggers a review, such as agent spend running 25% over plan for two months.

How do you forecast usage against a committed AI contract?

Treat the commit as a floor and forecast consumption against it month by month. Many enterprise AI contracts are prepaid or committed amounts that usage draws down, with overages billed at a contract rate once the commit runs out.

1. Forecast consumption first, from the usage and agent drivers above, as tokens or credits per month.
2. Burn it down against the commit. Cumulative consumption tells you the month the commit runs out.
3. Price the overage from that month onward at the contract's overage rate, not the committed rate.
4. Budget the floor anyway. If consumption lands below the commit, the unused amount is still spend; plan the higher of committed and forecast usage.
5. Flag the true-up. If the contract settles overages annually or at renewal, the cash timing differs from the expense timing, so carry both into the cash forecast.

For a multi-year commit, run this per contract year. Model choice matters too: routing routine work to cheaper models moves the usage line, and our guide to the best LLMs for finance teams compares current models and pricing.

Where should AI spend sit in the budget?

Classify each cost by what caused it, not by vendor. Customer-triggered inference and agents belong in cost of revenue, employee seats and internal tools in the operating expense of the team using them, and experimentation and model work in R&D. The same vendor can land in all three, which is why the driver for each line matters more than the invoice. Splitting one provider bill across departments is a standard cost allocation problem, and AI's share of R&D is now its own benchmark: the median SaaS company puts 10% of R&D toward AI. For how to classify each line, see whether AI spend is COGS or opex.

Customer-facing AI also changes gross margin, so model it alongside the rest of cost of revenue; see SaaS gross margin benchmarks for where that lands.

Common mistakes when forecasting AI spend

• Forecasting from the invoice total. One provider bill can hold seats, API usage and agent runs. Forecast from what's inside it, or a shift in the mix looks like a price change.
• Leaving embedded AI out of the base. AI credits and tier upgrades added to existing software at renewal are AI spend even when the vendor name hasn't changed. Check the renewal calendar for the year ahead and add the increases to the right line.
• Cutting power users to hit the number. Heavy usage can be waste, but it can also be someone who has found a much faster way to work. Set a controlled experimentation budget and judge it by what moves into real workflows.
• Ignoring model price changes. Model prices move during the year, in both directions, and new models change which tier a workload needs. Re-price the usage lines when you refresh them rather than locking in December's rates.
• Treating a pilot's cost as the production cost. A pilot's run rate rarely survives production. Model the production volume from the month it goes live, not the pilot's monthly bill.
• Hiding the range. A single number with no range invites the question you can't answer. Show the base case with low and high cases, and say which lines drive the width (usually agents).

Who should own each AI spend driver?

Give every line an owner who controls the driver as well as the budget. Finance owns the model and the reconciliation, but the assumptions belong to the people who move them:

• Seats: the department heads who approve access, working from the headcount plan.
• Internal usage: IT or the AI platform owner, who sees adoption and usage depth.
• Customer-facing AI: product and engineering, who know which features call which models and how often.
• Agents and workflows: the function that runs each workflow, with engineering confirming cost per run.
• Infrastructure: engineering or the platform team.

Named owners make the monthly refresh a short conversation about what changed, not a forensic exercise at quarter-end.

When should the AI forecast be updated?

Monthly for the usage-driven lines, quarterly for seats. Usage, pricing and new workloads change faster than any annual budget, so treat the AI line like a driver-based budget that rolls with actuals rather than a number fixed in December. Put it in the FY2027 budget calendar as its own step, with owners for each driver, so it isn't rebuilt from scratch in the last week of the cycle.

Build your AI forecast in Aleph

Aleph connects your general ledger, HRIS and provider spend so AI actuals refresh automatically, classification mappings apply across periods, and AI spend becomes a driver in the plan with scenarios instead of a trended line. Watch the From AI spend to AI budget replay to see a driver-based FY27 AI forecast built live, get the full framework and model in our budgeting for AI spend guide, or book a demo to build yours on your own data.