Your team just rolled out Claude Enterprise or ChatGPT for Business across the company. The enablement workshops went well. Adoption spiked. And then the invoice arrived.
This is the pattern playing out across mid-sized Australian businesses right now. AI tools are being adopted faster than finance teams can budget for them, and the pricing models are fundamentally different from anything most companies have dealt with before. According to PwC's 29th Global CEO Survey (January 2026), 56% of CEOs reported no measurable revenue or cost benefit from their AI investments. One contributing factor: companies are spending without governance, treating AI like just another SaaS line item when it behaves more like cloud compute.
This guide walks through how to properly budget for AI tools, set up governance that protects you without stifling adoption, and use cost levers most companies don't know about.
Traditional SaaS is predictable. You pay $X per seat per month, the cost is fixed, and you can forecast it twelve months out with near-perfect accuracy. AI tools break this model in three ways.
Anthropic's Claude Enterprise plan charges a per-seat fee (billed annually) that covers platform access only. All usage across Claude, Claude Code and Cowork is then billed separately at standard API rates based on actual consumption. There is no included token allowance with the seat fee. ChatGPT Enterprise operates on a similar model at scale, with fair-use throttling and usage-based billing kicking in for heavy users.
The current API rates (as of March 2026) for Anthropic's models are:
This means two users on the same plan can generate wildly different costs depending on which models they use, how much they use them, and what tasks they're performing. A software engineer using Claude Code heavily could consume 50 to 100 times more tokens than someone in marketing using it for occasional copy editing.
In any AI tool rollout, usage follows a steep power curve. A small number of heavy users (typically engineers, analysts and researchers) consume the vast majority of tokens, while a large portion of the user base barely touches the tool after the initial novelty wears off. This creates an uncomfortable exposure calculation. If you have 120 seats and set a flat overage cap of, say, $500 per user per month, your theoretical maximum exposure is $60,000 per month on top of your seat fees. That's $720,000 annualised. In practice most users won't hit it, but finance teams need to plan for the worst case.
Companies that run enablement workshops (as they should) typically see a 2 to 3x spike in usage in the first 60 to 90 days. People are testing prompts, regenerating outputs, building projects and generally figuring out what the tool can do. This is a feature, not a bug. But it means your first few invoices will be unrepresentative of steady-state costs. If you set your budget based on month one, you'll overshoot. If you panic at month one costs, you'll cut too early.
Based on what we're seeing work across Australian businesses scaling AI tools, here's a practical governance framework that balances cost control with adoption momentum.
Start with role-based soft caps and a single enterprise-level hard cap as a safety net. Don't set per-user hard caps yet because you don't have enough data to know where the right thresholds are.
Once you have four to six weeks of usage data, you can start tightening governance with confidence.
Usage normalises as people stop experimenting and start building repeatable workflows. At this stage, you should see cost per unit of output declining even if total usage stays steady or grows. Your governance shifts from controlling spend to optimising value.
Beyond governance and caps, there are three technical cost levers that can dramatically reduce your AI tool spend. These are primarily relevant if you're using Claude via the API (either directly or through Enterprise), but the principles apply to any usage-based AI platform.
The single biggest cost lever available to most companies. Haiku 4.5 costs US$1 per million input tokens. Opus 4.6 costs US$5. For straightforward tasks (drafting emails, summarising documents, answering standard questions), the output quality difference between the two is negligible, but the cost difference is 5x on input and 5x on output.
Teams that implement a 70/20/10 split (70% Haiku, 20% Sonnet, 10% Opus) instead of defaulting everything to Sonnet can cut their total token spend by roughly 60%. Even a simple rule like 'use Haiku unless the task requires complex reasoning' makes a significant difference.
Anthropic's prompt caching feature allows you to reuse previously processed prompt segments across API calls. Cached reads cost just 10% of the standard input price. The initial cache write costs 1.25x the base input rate for a 5-minute cache window, which means caching pays for itself after just one subsequent read.
For any workflow that uses a consistent system prompt, company knowledge base or standard set of instructions, this can reduce input token costs by up to 90% on repeated requests. If your team is running the same type of analysis across multiple documents or using a shared system prompt for customer support, prompt caching should be the first thing you implement.
For non-time-sensitive workloads (overnight report generation, bulk content analysis, data processing), the Anthropic Batch API provides a flat 50% discount on both input and output tokens. The trade-off is that processing is asynchronous, typically completing within minutes but with a 24-hour delivery guarantee.
Combining all three levers (model routing + prompt caching + batch processing) can reduce effective per-token costs by 80 to 95% compared to naively sending everything to the most expensive model in real time.
This is the section that nobody in the AI productivity space is writing about, because most of them aren't accountants. But it matters, particularly if you're trying to forecast cash flow or present a clean P&L to your board.
AI tool subscriptions and API usage sit entirely as operating expenditure. There is no asset being created that you own or depreciate. It's consumption-based spend, similar to your AWS or Azure bill.
Do not bury AI tool spend under 'Software Subscriptions' or 'IT Expenses'. Create a dedicated general ledger code (something like 'AI Tools & Compute') so you can track it separately. This gives you visibility for board reporting, makes it easier to compare against the productivity gains you're measuring, and prevents it from getting lost in a larger line item where nobody notices it creeping up.
The biggest accounting challenge with AI tools is that they're truly variable costs. One heavy sprint or project can swing your monthly AI bill by 30 to 50%. This makes traditional annual budgeting inadequate.
Best practice is to use driver-based forecasting:
Present all three scenarios to your board or leadership team. The range gives them comfort that you've thought about the downside, and the driver-based approach shows you understand the mechanics rather than just guessing.
Here's a simplified example for a 50-person Australian tech company rolling out Claude Enterprise.
Seat costs: 50 seats at Claude Enterprise pricing (custom, billed annually). Let's assume roughly $30 per seat per month as a baseline for this example. That's $1,500 per month or $18,000 per year for platform access alone.
Usage costs (unoptimised): If your 50 users average $100 per month each in token consumption (a reasonable estimate for a mixed team), that's $5,000 per month or $60,000 per year.
Total unoptimised: $78,000 per year.
With cost levers applied: Model routing (60% savings on token spend) + prompt caching (further 30 to 50% reduction on remaining input costs) could realistically bring usage costs down to $15,000 to $25,000 per year. Total optimised: $33,000 to $43,000 per year.
That's a significant difference, and it's the difference between a line item that raises eyebrows and one that's easily defensible.
Claude Enterprise uses a single seat type priced per user per month, billed annually, with a minimum of 20 seats. The seat fee covers platform access only. All usage is billed separately at standard API rates. Anthropic does not publicly list the exact per-seat price for Enterprise, as it varies by contract. Self-serve and sales-assisted options are both available.
AI tool subscriptions and API usage are operating expenditure (OpEx). You are paying for consumption of a service, not acquiring or creating an asset. It should be treated similarly to cloud compute (AWS, Azure) in your chart of accounts.
Prompt caching allows you to reuse previously processed portions of your prompt across API calls. Cached reads cost 10% of the standard input token price (a 90% discount). The initial cache write costs 1.25x the base rate for a 5-minute cache, meaning caching pays for itself after just one subsequent read of the same content.
Anthropic's Enterprise admin console allows admins to set spend limits at both the organisation level and the individual user level. There are no per-seat usage limits on the current Enterprise model, so usage must be managed through these admin-configured spend controls.
Haiku 4.5 costs US$1 per million input tokens and US$5 per million output tokens. Opus 4.6 costs US$5 input and US$25 output. That makes Haiku 5x cheaper on both input and output. For straightforward tasks where Haiku's output quality is sufficient, this is the single biggest cost lever available.
Monthly. AI tool spend is variable and can swing 30 to 50% month to month depending on project intensity. Set an annual forecast range (conservative, base, aggressive scenarios) but review actuals monthly and adjust your outlook quarterly.
AI tool spend is a new category of variable cost that most finance functions aren't set up to manage properly. At Scale Suite, we build the forecasting models, chart of accounts structures and management reporting that gives you visibility and control over this spend. If you're scaling AI tools across your team and need help getting the financial governance right, get in touch at hello@scalesuite.com.au.
Disclaimer: We review and check articles periodically. At time of writing (March 2026), the information above is accurate to the best of our knowledge. AI tool pricing changes frequently. Always verify current rates against official provider documentation before making budget decisions.
Scale Suite delivers embedded finance and human resource services for ambitious Australian businesses.Our Sydney-based team integrates with your daily operations through a shared platform, working like part of your internal staff but with senior-level expertise. From complete bookkeeping to strategic CFO insights, we deliver better outcomes than a single hire - without the recruitment risk, training time, or full-time salary commitment.
Considering hiring finance staff?
We'll show you the full cost of an internal hire vs our embedded team, and exactly how much you'd save.
We'll reply within 24 hours to book your free 30-minute call.
No lock-in contracts and 30-day money-back guarantee.
Prefer to book directly? Schedule your free 30-minute call here
