I’ve been sitting with a client situation that I think a lot of finance teams are going to recognize. They made the move from individual Claude subscriptions to Claude Enterprise. Smart decision — they needed the compliance controls, the audit logging, the admin oversight. All the things you need when you’re doing professional services work with client data.
And then the first full month’s invoice came in.
They went from roughly $20 a month per person to $20 a day for the organization. That’s a 30x increase. Same tool. Same team. Very different number. And the question that landed on my desk was the one finance always asks when a number goes up thirty times: what are we getting for this?
How AI Pricing Got Here
This isn’t a billing error. It’s a structural pricing shift that caught a lot of organizations completely off guard.
When Claude and ChatGPT first came out, the pricing was simple — $20 a month, all-you-can-use within your limits, flat rate. That pricing was never designed to be profitable at scale. It was designed to get users. The AI labs subsidized inference costs to drive adoption.
Then enterprise plans showed up with bundled token usage. Seat fees were higher, but they covered your usage costs up to a point. Total cost was predictable. You could model it.
Here’s what changed in 2025 and into 2026: Anthropic decoupled the seat fee from token usage. The seat fee actually dropped — down to around $20 per user per month. Looks like a win. But all token consumption — every prompt, every document analyzed, every workflow — now bills separately at full API rates. No bundle. No discount. And real-world costs for active users land anywhere from $60 to $250+ per user per month depending on usage patterns.
The teams that felt this hardest were the ones doing it right — heavy document analysis, automated workflows, using the tool all day for client deliverables. Because usage is the cost driver now. Not seats.
Why “Hours Saved” Still Doesn’t Cut It
I’ve written about this on Substack and I’ll say it again here: hours saved is a bad ROI metric. Not because it’s wrong directionally, but because cost doesn’t equal outcome. People overestimate how long tasks took before AI and overestimate how much time they save after. The incentives are messy and the math is easy but inaccurate.
And the data backs this up. A 2026 Writer/Workplace Intelligence survey found that 97% of executives say AI benefits their work — but only 29% report seeing significant organizational ROI. A Deloitte survey found 74% of enterprises want AI to drive revenue growth. Only 20% have actually achieved it.
Everyone agrees AI is doing something. Almost nobody can point to where it shows up in the P&L.
The Four-Layer Framework Finance Teams Need
There’s a framework I’ve found useful for structuring the internal AI ROI case. Think of it as four layers, stacked by credibility with your CFO.
Layer 1 — Activity: Tokens consumed, queries handled, uptime. Your IT team tracks this. It’s necessary, but it’s not ROI.
Layer 2 — Productivity: Hours saved, tasks completed faster, speed improvement. This is what most AI ROI presentations show. It’s the minimum to justify continued use — but it’s still not what a CFO calls financial impact.
Layer 3 — Financial Impact: Revenue influenced, costs avoided, risk reduced. This is where CFOs start to listen. If your team used AI to turn a proposal in two days instead of five and won that deal, Layer 3 is the revenue from that deal. If AI flagged an error before it went to the audit committee, Layer 3 is the cost of the restatement you avoided.
Layer 4 — Strategic Capability: New services you can now offer, competitive positioning, headcount leverage. This is board-level conversation — partially quantifiable, but directionally what most AI investments are ultimately buying.
If you’re building the ROI case for a CFO, you need to be operating at Layer 3. And you need to identify — before the tool goes live — what specific P&L lines you’re trying to move.
Key Takeaways
- The 30x bill is structural, not a mistake. Token unbundling is now the default for enterprise AI contracts. Model it before you sign or renew.
- “Hours saved” is Layer 2. CFOs need Layer 3 — financial outcomes that appear somewhere in the P&L.
- Set your baseline before deployment. You cannot prove ROI retroactively if you didn’t document the starting point.
- Cost management is now ongoing. Audit model tiers, govern Fast Mode (6x the standard rate), push async work to Batch API. The AI meter runs continuously.
- The CFO owns this. 48% of finance chiefs say AI ROI accountability sits with them. If you’re in finance or advising finance clients, this conversation is yours to lead.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply