Building an AI Cost Model Before the CFO Asks for One

The forecast request is coming

If your CFO hasn't yet asked for an AI spend forecast, they will. The line item is growing too fast and too visibly to stay in the "emerging technology" category where nobody asks hard questions. At some point in the next two quarters, someone in finance will say: "What does this look like in 12 months? In 24?"

The question is whether you have an answer ready or you're building it under pressure. The organizations that build their AI cost model proactively — before the board asks — earn trust, influence budget conversations, and get ahead of the inevitable cost scrutiny. The ones that scramble produce a number nobody believes.

Why traditional IT cost models don't work for AI

If you've built cloud cost models, you'll recognize the structure. But AI breaks several assumptions that make traditional models work:

• Usage-based pricing defies linear projection. A per-seat SaaS tool grows with headcount. An API-based AI service grows with usage intensity, which can spike 10x when a team deploys a new agent or changes a workflow. Headcount is a poor proxy.
• Adoption curves are S-shaped, not linear. AI tools spread through organizations in waves. A team pilots, validates, scales, and then usage plateaus until the next use case is found. Extrapolating from the growth phase overestimates. Extrapolating from the plateau underestimates.
• Prompt engineering changes cost profiles overnight.An engineer refactors a system prompt, switches from a reasoning model to a standard model for a subset of queries, or adds a caching layer. Costs shift 20–40% with no change in usage volume. There's no equivalent in traditional IT cost modeling.
• Model version changes reset price-per-unit. When a vendor releases a new model generation, the price-per-token often drops but the tokens-per-output may change. The net cost impact is unpredictable from the outside.

None of this means you can't model AI costs. It means you need to model them differently — with more scenario variance, more frequent recalibration, and a structure that accounts for organizational complexity.

The four inputs your model needs

A useful AI cost model has four inputs. Most people think about the first three and skip the fourth, which is the one that makes the model actually useful.

1. Adoption trajectory

How many users, teams, or workloads will be consuming AI services over the forecast period? This is the volume dimension.

Don't model adoption as a single company-wide curve. Model it per business unit. The data science team's adoption trajectory looks nothing like marketing's. Engineering may be at 80% penetration with Copilot while customer success is just starting a pilot. Aggregating them produces an average that's wrong for everyone.

2. Consumption intensity

How many tokens, API calls, or inference requests does each user or workload generate? This is the usage-per-unit dimension.

Consumption intensity varies enormously by use case. A developer using Copilot for code completion generates modest, steady token volume. A data pipeline running document extraction generates high, bursty volume. An AI agent serving customer inquiries generates variable volume that correlates with support ticket volume. Model each separately.

3. Unit pricing

What does each unit of consumption cost, accounting for tiered pricing, committed-use discounts, and model selection? This is the rate dimension.

Price per token is not static. Most AI vendors offer volume discounts at commitment thresholds. Some offer reserved capacity pricing. Model selection matters — routing simple queries to cheaper models and complex queries to more capable models can reduce average cost per inference by 30–50%.

4. Organizational mapping

This is the input most models miss, and it's the one that makes the difference between a number the CFO nods at and a model the CFO actually trusts.

Your AI cost model should be built on the same organizational hierarchy that your financial reporting uses: business units, cost centers, departments, projects, geographies. The rows in your model aren't "ChatGPT" and "Copilot" — they're business units crossed with use cases crossed with AI tools. That's the granularity the CFO needs to make budget decisions.

Without this mapping, your model produces a single company-wide number. With it, the model produces a forecast that each BU leader can validate against their own plans. The difference in credibility is enormous.

Start with scenarios, not point estimates

Given the volatility of AI adoption and pricing, a single forecast number is dishonest. Present three scenarios:

• Base case:Current adoption rates continue. No new major use cases. Pricing stays flat. This is the "nothing changes" projection and it's almost certainly wrong — but it's the floor.
• Expansion case: Two to three new business units adopt AI tools. One existing use case scales significantly. Volume discounts kick in at the next pricing tier. This is the most likely scenario for most organizations.
• Breakout case:Engineering deploys an agent framework. A business-critical process gets rebuilt around AI. Token consumption jumps 5–10x in a single quarter. This is the scenario the CFO needs to see, even if it's not the most likely, because it's the one that blows the budget.

The range between scenarios is the honest answer. Presenting it as a range rather than a point estimate signals sophistication, not uncertainty. Finance teams prefer an honest range over a precise number that turns out to be wrong.

The model killers

Even a well-built AI cost model can break. Watch for these:

• New use cases nobody anticipated.The marketing team decides to generate all product descriptions with AI. The legal team starts using AI for contract review. These weren't in the plan and they carry real cost. Build a buffer — 15–20% of the base case — for unanticipated adoption.
• Model upgrades that change pricing. When a vendor ships a new generation, your cost-per-token changes. Sometimes it drops (efficiency gains). Sometimes it rises (more capable but more expensive models become the default). Review your model quarterly and recalibrate.
• Viral internal adoption. AI tools spread through word-of-mouth. A team pilot turns into a department standard turns into a company-wide rollout in weeks, not months. The adoption curve you modeled over 12 months happens in 3.
• No organizational taxonomy.When a new department starts using AI and there's no cost center or BU structure to slot them into, their costs land in a catch-all bucket that nobody manages. The model breaks because the mapping breaks. Build the taxonomy first.

A template you can use this week

You don't need a complex financial model to start. We've built a free Google Sheets template that gets you 80% of the way. Make a copy and start filling it in.

The template is structured around the four inputs above:

• Rows:One row per business unit × use case × AI tool combination. If the sales team uses both Copilot (seats) and an internal AI assistant (API), that's two rows for sales.
• Columns: Organizational mapping (BU, cost center, geography), AI tool/vendor, pricing model, current monthly cost, and projected cost at 6 and 12 months.
• Roll-up tabs: Pre-built summaries by business unit, cost center, vendor, geography, and company-wide.
• Three scenario sheets (base, expansion, breakout) pre-wired so you can present the range, not a single number.

Build the base case first. The whole exercise takes a day if your AI showback datais already mapped to your org structure. If it's not, building the taxonomy first is time better spent than building a model on shaky data.

Where to start

Map your organizational taxonomy. Inventory your AI spend against it. Build the three scenarios. Present the range. You'll have more credibility than 90% of the AI cost forecasts being presented to boards right now — because yours will be grounded in real usage data mapped to real business structure.

Cloudsaver's free savings assessment includes AI spend forecasting with automated scenario modeling — built on your actual consumption data, mapped to your organizational hierarchy.

The best time to build an AI cost model is before anyone asks for one. The second best time is today. The worst time is in a board meeting with a number you made up on the way in.