Token costs, uncontrolled AI spend, and the governance gap that will cost your organisation — if you don't act now.
Every CFO I've spoken to in the last 12 months has signed off on at least one AI initiative. Most have signed off on several. Claude, Microsoft Copilot. ChatGPT Enterprise. AI features bundled into your ERP, your FP&A platform, your HR system. It's everywhere, it's accelerating, and for most organisations the cost of it is currently invisible.
That invisibility is about to end.
I'm not writing this to slow down your AI adoption. I'm writing it because the CFOs I respect most are the ones who price risk honestly — and right now, AI token cost is an unpriced risk sitting inside almost every finance function.
Let me give you the CFO version, not the technical one.
Every interaction with an AI system — every question answered, every document summarised, every forecast commentary generated, every journal entry proposed — consumes a unit of compute called a token. Tokens are the currency of AI. And like any currency, someone is paying for them.
Right now, for most organisations, that cost is bundled. It's absorbed into a Microsoft 365 licence fee, or included in a per-seat SaaS price. It feels free because it's not yet a separate line item. But the economics underneath it are not free. The hyperscalers — Microsoft, Google, Amazon, OpenAI — have collectively spent hundreds of billions of dollars building the infrastructure that runs these models. That capital has to be recovered. And as AI usage scales, the pricing models will adjust to reflect actual consumption.
The shift is already underway
What happens when 'included in licence' becomes metered usage
Microsoft Copilot, Azure OpenAI and others are already offering consumption-based pricing tiers — the bundled era is ending
Picture this: it's 2027. Your organisation has deployed AI across finance, HR, procurement, and customer service. Every workflow touches AI. Your agentic finance stack is processing thousands of transactions per day — reconciliations, journal proposals, variance commentary, consolidation runs, audit checks.
Each of those interactions consumes tokens. Each of those tokens costs money. And the bill arrives — not as a budget line you planned for, but as a usage overage on a platform contract nobody reviewed carefully.
This is not speculation. It's the natural trajectory of a market where AI infrastructure costs are currently being subsidised to drive adoption. The subsidies don't last. They never do.
Before we know it, all those co-pilots and ChatGPTs running around the organisation will start quoting you every time someone asks something. It is going to cost you. The question is whether you see it coming.
In my experience working across finance transformations in Australia, New Zealand, and the Pacific, I see three patterns of token waste that are already costing organisations money they're not tracking:
1. Conversational AI without governance
When AI tools are deployed without usage policies, employees use them for everything — including tasks that don't require AI at all. Long, rambling prompts. Repeated questions because the first answer wasn't quite right. Document uploads of files that are thousands of tokens long when a summary would have served the purpose. Every one of those interactions has a cost.
2. Poorly designed agentic workflows
Agentic AI — where AI systems take sequences of actions to complete a task — is significantly more token-intensive than a simple question-and-answer interaction. An agent that loops back to reconsider, that makes unnecessary tool calls, that passes large context windows repeatedly through the model — this is expensive compute. Well-designed orchestrated workflows minimise token consumption by design. Poorly designed ones burn through compute at a rate that will surprise you when it's metered.
3. Shadow AI
Your teams are using AI tools you haven't sanctioned. ChatGPT personal accounts. Free-tier tools. Unofficial browser extensions. Some of this is genuinely benign. Some of it involves your financial data being processed by models you haven't vetted, under data agreements you haven't reviewed, with no audit trail of what was sent or what came back.
The token cost isn't the biggest risk here. The data governance risk is. But both are real.
Ask yourself honestly: does your organisation have a framework for AI cost governance? Not a policy document that says 'AI should be used responsibly' — an actual framework that tracks consumption, attributes cost to business units, establishes usage thresholds, and reviews AI-generated outputs before they become business decisions?
If the answer is no — or 'sort of' — you are in the same position as organisations that deployed cloud infrastructure in 2012 without tagging resources or setting spend alerts. The AWS bill shock of that era is now a well-documented corporate cautionary tale. The AI bill shock of the late 2020s is still preventable.
Most of organisations have no formal AI cost governance framework
The AI equivalent of untagged cloud spend — invisible until it isn't
For those of us in finance, there's an additional layer of risk that goes beyond cost.
When an AI system generates a journal entry, proposes an accrual, or produces variance commentary — and that output is posted or published without human review — you have a governance failure. The question the auditor will ask is not 'did AI do this?' The question is 'who approved it, and is there a record of that approval?'
Ungoverned AI in a finance context isn't just a cost risk. It's an audit risk, a compliance risk, and in some jurisdictions, a regulatory risk. The framework for AI governance in finance needs to address not just spend, but accountability for every decision the AI makes.
This is why in everything we build at Octane, the principle is non-negotiable: the agent proposes, the human approves. Every journal, every cost-centre assignment, every period lock is presented to a named controller before it posts. Nothing happens without a human decision. And every decision — approval, rejection, revision — is logged with a timestamp.
That isn't a constraint on AI effectiveness. It's what makes AI trustworthy in a finance context.
Audit your AI tool landscape. Get a complete picture of every AI tool your organisation is using — sanctioned and unsanctioned. Understand what data is flowing through each one and on what contractual basis.
Understand your current token cost exposure. Work with your IT and procurement teams to identify where AI usage is currently bundled versus metered. Get a baseline consumption number, even if approximate.
Build token cost into your AI business cases. Going forward, every AI initiative should include a consumption cost model alongside the efficiency benefits. The ROI calculation isn't complete without it.
Design governance into agentic workflows from the start. If you're evaluating or deploying agentic AI in finance, make human-in-the-loop approval a design requirement, not an afterthought. The audit trail this creates is not just good governance — it's the evidence your auditors will want.
The AI bill shock is preventable. But only if you see it coming and act before consumption scales.
The next blog in this series covers the one I get asked about most: governance. Specifically, how do you build an AI-powered finance function that your auditors will actually respect?
Amendra Pratap is the Founder and Managing Director of Octane Solutions, an IBM Gold Partner specialising in AI-powered finance transformation across Australia, New Zealand, and the Pacific.