Why AI Costs Break Traditional SaaS Budgeting
Traditional SaaS budgeting is built on a simple assumption: you know what you're going to pay before you pay it. A seat-based tool costs $X per user per month, and your forecast is a multiplication exercise. AI tools have shattered that assumption. Across the Cledara platform, over 73% of companies now pay for at least one AI tool, with the average company subscribing to 3.1 AI vendors. And roughly 76% of those AI subscriptions use soft (variable) budgets rather than fixed ones, which tells you everything about how unpredictable this spending category has become.
This matters because AI is not a niche line item anymore. It spans engineering, marketing, design, data, and company-wide productivity. When spend is variable, spread across departments, and growing fast, your traditional SaaS budgeting process will not keep up. You need a dedicated AI cost allocation framework.
Fixed-seat SaaS vs usage-based AI: a fundamentally different cost model
With fixed-seat SaaS, your cost is determined at the point of purchase. Ten Slack licences at $12.50 per user per month equals $125 per month, full stop. You can budget annually with high confidence. Usage-based AI tools flip this entirely. Your cost is determined by consumption: tokens processed, API calls made, images generated, or minutes transcribed. The same tool can cost $20 one month and $2,000 the next depending on how your team uses it.
This creates a forecasting problem that most finance teams have never dealt with in their software stack. Infrastructure teams are familiar with variable cloud costs (AWS, GCP), but application-layer AI tools bring that same variability into every department's budget. An engineering team running Anthropic API calls and a marketing team generating content with Jasper are both creating unpredictable spend, but in completely different ways.
The three AI pricing models (per-seat, per-token/API call, hybrid)
To budget effectively for AI, you first need to understand how each tool charges you. AI pricing falls into three distinct models, each requiring a different allocation approach.
Per-seat pricing is the most familiar. ChatGPT Team at $25 to $30 per user per month, Cursor Pro at $20 per user per month, and Perplexity Pro at $20 per user per month all follow this model. These are predictable: multiply the per-seat cost by headcount and you have your budget. The allocation challenge here is departmental: which team owns the subscription, and who approves new seats?
Per-token or per-API-call pricing is where forecasting gets difficult. The OpenAI API charges per million tokens (ranging from $0.25 to $150 depending on the model), Anthropic's API charges similarly, and Mistral AI follows the same pattern. Costs scale directly with usage, and usage can spike without warning when a developer ships a new feature or a data team runs a large batch job. Based on Cledara platform data, Anthropic API subscriptions show wide payment variance, with individual transactions ranging from under $1 to nearly $1,000.
Hybrid pricing bundles a base subscription with usage-based overages. Notion now includes AI features in its Business plan at $18 per user per month, with usage limits baked in. Cursor offers a $20 per month Pro plan that includes a credit pool for AI completions, with higher-usage tiers at $60 and $200 per month. ElevenLabs charges a monthly fee that includes a set number of characters, with overage charges beyond that. Hybrid models are the trickiest to forecast because you need to predict not just adoption but intensity of use.
Building an AI Budget: The Framework
A workable AI budget starts with categorisation, moves to departmental envelopes, and builds in explicit variance buffers for the tools that refuse to behave like traditional SaaS.
Categorise AI spend: productivity tools vs developer tools vs embedded AI
Not all AI spend is created equal, and lumping it together makes allocation impossible. Break your AI stack into three categories.
Productivity AI covers tools used broadly across the organisation for day-to-day work. Think ChatGPT Team, Claude, Perplexity, Grammarly, and Otter.ai. On the Cledara platform, Claude is used company-wide, while Perplexity AI and Grammarly similarly serve cross-functional teams. These tools are typically per-seat, relatively predictable, and should be budgeted centrally or allocated per headcount.
Developer AI includes tools used primarily by engineering and data teams: Cursor, GitHub Copilot, Replit, Lovable, and API access to OpenAI, Anthropic, or Mistral. Cledara data shows that Cursor is used exclusively by engineering teams, while OpenAI and Anthropic API access is concentrated in engineering and data. These range from per-seat (Cursor) to fully usage-based (API calls), and budgets should sit with the engineering or data team that owns the consumption.
Embedded AI refers to AI features built into tools you already pay for. Notion AI, SaaS platforms with AI add-ons, and AI-powered analytics are examples. The cost is often invisible, bundled into a higher-tier plan you might have upgraded to specifically for AI features. These are the hardest to isolate and allocate because the AI cost is inseparable from the platform cost.
Set budget envelopes by department and use case
Once you have categorised your AI tools, assign budget ownership. Each department should have a defined AI budget envelope that covers the tools their team uses. The data supports this approach: on the Cledara platform, AI tools map cleanly to functional areas. Midjourney and ElevenLabs are used by marketing and design teams. Cursor and Replit sit with engineering. Fireflies and Otter.ai serve product and operations teams.
A practical starting point is to audit your current AI spend by department (Cledara's expense code mapping does this automatically), then set envelopes 10 to 15% above current run-rate to allow for organic growth. For departments that are just beginning to adopt AI, create a smaller exploratory budget with a quarterly review trigger.
Build in 20 to 30% variance for usage-based tools (with caps)
Here is the most important piece of practical advice in this guide: for any usage-based AI tool, budget 20 to 30% above your projected spend, and pair that buffer with a hard spending cap.
The buffer acknowledges reality. Usage-based AI costs are inherently variable, and trying to forecast them to the penny is a waste of time. The cap prevents a runaway scenario where an API integration goes haywire and racks up thousands in charges overnight. On the Cledara platform, you can set spend limits on the virtual card assigned to each AI tool, which means the cap is enforced automatically at the payment level, not just in a spreadsheet.
For per-seat AI tools, a 5 to 10% buffer (for unplanned seat additions) is sufficient. For hybrid tools, budget at the tier that includes your expected usage, plus 15% for potential overages.
AI Chargeback Models That Work
As AI spend grows, finance teams need a systematic way to charge costs back to the departments that generate them. According to McKinsey's research on AI monetisation models, companies that implement departmental chargeback for AI create accountability and reduce wasteful consumption. Here are three models that work in practice.
Per-user allocation (for seat-based AI tools)
This is the simplest model. If your company pays $25 per seat per month for ChatGPT Team and the marketing team has 12 seats, marketing's chargeback is $300 per month. The cost is deterministic and the allocation is straightforward.
Per-user allocation works well for: ChatGPT Team, Cursor, Perplexity Pro, Grammarly Business, and similar per-seat tools. Assign each subscription to the department that owns the majority of seats, and split the cost proportionally if multiple departments share the tool.
Per-API-call allocation (for developer AI tools)
For usage-based AI tools like the OpenAI or Anthropic API, allocate costs based on actual consumption. This requires tagging API usage by team or project, which is where an AI cost management layer becomes essential.
The practical approach: use Cledara's AI Dashboard to track daily spend per AI provider, then allocate monthly costs based on each team's share of total API consumption. If the data engineering team consumed 60% of your Anthropic API tokens and the product team used 40%, the chargeback splits accordingly. This creates transparency and incentivises teams to optimise their usage (choosing cheaper models for simple tasks, caching responses, batching requests).
Departmental budget ownership with central governance
The most effective model for mid-market companies combines departmental autonomy with central oversight. Each department owns its AI budget and makes purchasing decisions within that envelope. A central finance or IT function sets the overall AI budget, defines approved vendors, and monitors for shadow AI (unapproved tools purchased outside the governance framework).
This model works because it balances speed with control. Departments do not need to file a procurement request every time they want to add a ChatGPT seat, but finance retains visibility into total AI spend and can intervene if a department consistently exceeds its envelope. On the Cledara platform, this is enforced structurally: each AI tool gets its own virtual card with a spend limit, and approval workflows trigger automatically when a team requests a new tool or a budget increase.
Forecasting AI Spend: A Practical Approach
Forecasting AI spend is harder than forecasting traditional SaaS, but it is not impossible. The key is to accept imprecision and focus on directional accuracy rather than false precision.
Trend-based forecasting from historical usage data
Start with what you know. Pull three to six months of AI spending history and look for patterns. Most companies will see a ramp-up curve as AI adoption grows, followed by a more stable run-rate once teams settle into their workflows.
Cledara's Budgeted Spend feature projects future AI costs based on historical payment patterns, automatically accounting for the variability of usage-based pricing. This is more reliable than manual forecasting because it uses actual transaction data rather than vendor estimates or department guesses.
For API-based tools, look at month-over-month growth in consumption (not just spend, since per-token prices drop frequently). If your OpenAI API usage grew 15% month over month for the last quarter, project that forward, but remember that LLM inference costs have been dropping roughly 80% per year, which means your unit costs are falling even as usage rises.
Scenario planning: what if AI adoption doubles next quarter?
Build three scenarios into your AI budget forecast.
Conservative: current adoption holds steady, usage-based costs grow at the trailing three-month rate, no new AI tools added. This is your floor.
Base case: adoption grows 10 to 20% (new seats, one or two new tools), usage-based costs grow at 1.5x the trailing rate. This is your planning target.
Aggressive: a new AI initiative launches (a company-wide rollout of an AI coding assistant, a new AI-powered customer support tool), adoption jumps 30 to 50%. This is your contingency scenario, and the 20 to 30% variance buffer should cover the gap between base case and aggressive.
The companies that get AI budgeting right are not the ones that forecast perfectly. They are the ones that plan for variability and have the tools to see what is actually happening in real time.
How Cledara Makes AI Cost Allocation Automatic
Most of the framework described above can be implemented manually with spreadsheets and monthly reconciliation. But manual processes break down as your AI stack grows, especially when you have 3+ AI vendors (as the average Cledara customer does) with different pricing models and billing cycles.
Cledara automates AI cost allocation across the entire workflow.
The AI Dashboard connects to AI providers via API keys and tracks daily spend per provider. Instead of waiting for a monthly invoice to find out what your OpenAI or Anthropic bill was, you see spend accumulating in real time, with the ability to set alerts and budgets before costs spiral.
Virtual cards per AI tool are the foundation of automatic allocation. Each AI subscription gets its own Cledara virtual card, which means every payment is automatically attributable to a specific tool and team. Usage-based costs are captured the moment they hit the card, not when someone remembers to log them in a spreadsheet. And because each card has its own spend limit, your variance caps are enforced at the payment layer.
Accounting integrations with Xero, QuickBooks, and NetSuite complete the loop. Every AI expense is automatically mapped to the correct GL code and cost centre, so your books reflect accurate departmental AI allocation without manual journal entries. When your CFO asks "how much did engineering spend on AI last quarter?", the answer is already in your accounting system.
Expense code mapping lets you assign each AI tool to the department that owns it. Combined with Cledara's budgeted spend forecasting, you get a complete picture: what each department spent on AI last month, what they are projected to spend next month, and whether they are on track against their budget envelope.
For companies managing AI spend across multiple departments and pricing models, the combination of virtual cards, real-time tracking, and automated accounting turns AI cost allocation from a quarterly fire drill into a continuous, governed process. Book a demo to see how it works with your AI stack.
