Why Your Finance Team's AI Bill Keeps Climbing While Token Prices Crash
Why Your Finance Team's AI Bill Keeps Climbing While Token Prices Crash
Token prices have fallen roughly tenfold a year. Enterprise AI spend still doubled. Here's the actual cost driver — and the fix that doesn't involve cutting the tools your team relies on.
If you had asked a CFO eighteen months ago what would happen to AI costs once token prices came down, most would have guessed the bill would shrink. Reasonable guess. Wrong answer. Token prices for large language models have fallen at roughly the rate industry analysts have tracked since the technology went mainstream: about ten times cheaper for equivalent output quality, every year, according to Andreessen Horowitz's ongoing analysis of LLM inference pricing. And yet enterprise AI-native application spend grew 108% year-on-year in 2026, according to Zylo's 2026 SaaS Management Index, with 78% of IT leaders reporting AI charges nobody had budgeted for.
For a finance team running AI-assisted reconciliation, invoice matching or reporting workflows, this isn't an abstract industry statistic. It's the gap between the number in the budget and the number on the invoice.
The numbers behind the paradox
Two figures worth sitting with. Enterprise AI spend isn't creeping up — it's roughly doubling year-on-year even as the underlying compute gets cheaper. And a widely-cited industry estimate, reported by CX Today in February 2026 from a review of enterprise agentic AI deployments, put the share running over budget at 73%, with some blowing past their original estimate by more than 2.4 times. We'd treat that second figure as a directional signal rather than a precise benchmark — the underlying study behind it isn't independently published — but it lines up with what Zylo and others are separately reporting: AI spend is growing faster than most organisations are forecasting it.
|
+108% YoY
Growth in enterprise AI-native app spend in 2026, even as per-token prices kept falling. Source: Zylo's 2026 SaaS Management Index.
|
~73%
Share of enterprise agentic AI deployments reported to run over budget, per industry reporting (CX Today, Feb 2026) — treat as directional, not precise.
|
Where the money actually goes
The mechanism is straightforward once you see it. A year or two ago, "using AI" in a finance context typically meant one prompt, one response. You asked a chatbot a question, it answered, that was the transaction. Token pricing on that model was a genuinely reasonable proxy for cost.
That's not what AI usage looks like in most finance workflows now. An AI agent doing bank reconciliation or invoice matching doesn't send one call — it plans a sequence, checks its own output, retrieves supporting records, flags exceptions, and sometimes retries when something looks off. A single reconciled transaction can trigger a dozen or more model calls under the hood, each with its own context window, and that context tends to get fatter over time as more history, more rules and more prior exceptions get fed back in. On top of that, a growing share of AI spend now comes from monitoring agents that run continuously in the background, checking for anomalies whether or not anyone asked them to look.
Multiply calls per task by tokens per call by hours running unattended, and a falling per-token price stops mattering much. The unit price went down. The number of units per task went up faster.
Track cost per completed task, not cost per token
This is the practical shift finance leaders need to make. Auditing your vendor's per-token rate is the wrong question to be asking in 2026. The useful question is: what does it cost, end to end, to complete one bank reconciliation, match one batch of invoices, or resolve one flagged exception? That figure captures the planning calls, the retrieval calls, the retries and the background monitoring that a token price alone hides.
Once you're tracking cost per completed task, you can actually see waste. An agent that calls the ledger three times to confirm what it already knew the first time isn't efficient just because each of those calls is individually cheap.
The "compile the workflow" pattern
Earlier this month, developer Vivek Haldar published a detailed account of cutting one of his own AI agent workflows' token use by 94% and its latency by 87%, with no quality loss he could detect in his own runs — and without switching to a cheaper model. The workflow in question was a daily skill that searched his blog archive for older posts worth resurfacing, checked whether he'd mentioned them recently, and drafted a short LinkedIn post linking back to the original. It had originally been written entirely as natural-language instructions that an AI agent interpreted fresh on every single run.
His insight was that after running the same workflow repeatedly, most of it had stopped being exploratory. It searched the same sources, built the same inventory, and applied the same filters every time. Only two steps genuinely needed a model: choosing a good candidate, and writing the draft. Everything else could be ordinary, deterministic code. So he "compiled" the workflow — a thin layer still calls an AI model for the judgment steps, but the repeatable plumbing now runs as code instead of being re-reasoned from scratch on every execution.
The pattern he describes is a useful one for any recurring task, not just his: start in natural language because it's flexible and lets you discover what the workflow actually needs. Run it enough times to see which parts have "crystallised" into the same steps every time. Then move those stable parts into code, and keep AI calls only where genuine judgment is required.
What this looks like for a finance team
Recurring finance tasks tend to follow the same arc. Bank reconciliation or invoice matching often starts as ad hoc chatbot use: someone pastes a statement or an invoice batch into an AI tool and asks it to find the matches. That's a reasonable way to start, because it's fast and it surfaces what the real matching rules are.
The problem is when that ad hoc pattern never evolves. If the same categories of transaction are being matched the same way every week, that repetition is the signal the workflow has crystallised and is ready to move from "ask AI every time" to "AI handles the judgment calls, code handles the repeatable steps." The exception that genuinely needs a human-level read — an unusual narration, a vendor name that doesn't quite match — still goes to AI. The routine ninety percent of matches that follow a known rule don't need to be re-reasoned by a model every single time.
This is exactly the kind of pattern we've been applying at PFL as we build out our own tools. One example, still in early stages, is a RAG-based payroll checker designed to flag anomalies in payroll data before it's finalised. We're not ready to talk about how it works in detail yet. What we can say is why we're building it this way rather than as another "paste it into the chatbot" tool: the checks that repeat every pay run should run as repeatable logic, and AI should be reserved for the judgment calls that genuinely need it. It's early days, with more building and testing ahead of us.
Falling token prices were never going to fix your AI budget on their own. What fixes it is knowing which parts of a recurring workflow have stopped needing a model, and having the discipline to move them out of the token meter entirely.
Do you know what your AI tools actually cost per completed task?
PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations — including help making sense of what your AI-assisted workflows are really costing you.
Talk to PFL →Welcome to LLMflation — Andreessen Horowitz
Zylo's 2026 SaaS Management Index Finds AI-Native App Adoption Is Surging — Zylo
The Agentic AI Cost Problem: Calculating TCO for Agentic AI — CX Today
How I Cut an AI Agent's Token Use by 94% — Vivek Haldar
Why Your Engineers' Favorite AI Tools Are Wrecking Your 2026 Budget — Forbes
Next up: a look at how NFP boards are being asked to sign off on AI use policies — and what should actually be in one.
Comments
Post a Comment