The bill nobody budgeted for
CFOs were told AI would save them money. Then the invoices arrived. In Pigment's Q3 CFO Index, 83% of finance leaders say consumption-based AI costs pushed their spending above expectations. Uber reportedly burned through its entire 2026 AI budget for coding tools in four months.
So the question I hear from finance teams isn't "should we use AI?" any more. It's "how do I justify this token spend?"
My take: before you go shopping for a cheaper model, look at what the tokens are actually being used for.
You're paying for the wandering
With agents, you're not just paying for the result. You're paying for all the wandering it took to get there.
Every time an agent has to figure out how your system works, it burns tokens. Which custom fields on a sales order actually matter. Which script fires when. What that odd status from 2019 means. Unless someone has written it down, the next run often does the same archaeology all over again. And the one after that.
Picture it with people. You hire a contractor who spends three days learning your account, fixes one thing and leaves. Next week you hire another contractor, who spends three days learning your account. Nobody would run a team that way. But it's roughly how a lot of agents work on an ERP today.
A messy account is a cost, not just a risk
For years, a messy NetSuite account was a risk problem. Zombie workflows, scripts nobody remembers writing, fields that made sense to someone once. Scary before a big change, easy to ignore on a normal Tuesday.
Not any more. A messy ERP isn't just a risk problem. It's a line item. The more confusing your account, the more you pay for AI to figure it out, and you pay again every time it has to figure it out from scratch.
The hourly report that ate the budget
I saw a post on Reddit that made me chuckle. A finance team built live NetSuite reports as Claude artifacts that refresh every hour. Each refresh fired off a bunch of saved searches... and they hit their Claude usage limits by midweek. The top reply suggested having Claude build a Suitelet or a dashboard that shows the same numbers instead. I couldn't agree more.
And it's not just a Claude thing. NetSuite's own assistant bills in AI units, and Oracle's documentation puts a simple question at around 10 of them and a research question at 50 to 200. Ask the same question every hour and it adds up fast.
This is a pretty good example of how CFOs end up with bills that came in way higher than promised. Using AI to build the report is great. Using AI to run the same report every hour, all week, is overkill when a plain old dashboard does it for free.
Build it once, then let it be boring
The more your system can handle deterministically, the fewer tokens you burn.
So use AI where judgment is needed: understanding what's there, designing the change, writing the script. Then hand the repeat work to the boring, reliable machinery your ERP already has. Saved searches, dashboards, scheduled scripts. They don't get any smarter, but they don't send you a bill either.
Build it once with AI, then let the dashboard do the boring repeat work. Your CFO will thank you.
The cheapest optimization isn't a cheaper model
The best AI cost optimization might not be a cheaper model at all. It might be retiring the zombie workflows and documenting what things are for, so the agent spends its budget on the work, not the digging.
That's a big part of why we built Splotch the way we did: map the account once, keep the map current as it changes, and let every agent (and every human) start from what's already known instead of from zero.
If AI spend has quietly become its own budget line on your team, start by looking at what your agents are digging through. That's usually where the money is.
From Kevin Lorenz
← Back to all posts