Every week, a reconciliation analyst asks an AI assistant to match trade data against cash positions. It walks through the rules, checks each line, flags what’s off, and gets it right. The next week, same reconciliation, same assistant — it works the whole thing out again from scratch.
Multiply that across every exception and every close, and the driver behind the AI bill becomes clear: AI keeps redoing work it’s already done. The reason is simple: a model re-estimates the calculation every time instead of looking up an answer it already has, and each estimate gets billed as new work, with no guarantee it agrees with the last one.
That rework shows up in three ways. Teams feed AI large, unprepared datasets that skip past the business rules already in place, which burns tokens and produces answers nobody fully trusts. They reach for a high-cost reasoning model to do a routine job a cheaper one could handle, or send more context than the task actually needs. And most organizations have limited visibility into what any of this is costing until the invoice arrives. Each problem compounds the others: a wrong answer means rework, and rework means paying twice for the same task.
The bill and the return are the same problem
Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027. Escalating cost is one of the three reasons Gartner names, alongside unclear business value and weak risk controls. Those three reasons are related: an agent that has to reason through the same calculation every time gets more expensive and less trustworthy at the same time, for the same reason.
And even where projects survive, the payoff is thin. BCG’s Center for CFO Excellence found that median ROI from AI in finance is just 10%, and a third of finance leaders report limited or no gains at all. That’s not a story about AI being the wrong technology. It’s a story about what it’s being asked to do.
What changes when the calculation isn’t the job of AI
NextWave Consulting, an Alteryx partner, cut token consumption 20x on a front-office-to-back-office reconciliation once the calculation ran through an existing governed workflow instead of an LLM working it out from scratch each time. That’s the shift underneath the “AI is getting smarter” story most teams are chasing. Some of what finance asks AI to do was never a reasoning problem to begin with.
AI is strong at a specific set of tasks:
- Drafting a first-pass narrative or summary from data that’s already correct
- Spotting a pattern or anomaly worth a second look
- Answering a question in plain language instead of a formula
None of those require the model to rebuild your matching rules, your materiality thresholds, or your close calendar from raw data every time someone asks. That work already has an answer. It was built once, tested, and run the same way for years, usually in a workflow, sometimes still in a spreadsheet nobody’s gotten around to moving. Handing that logic to a model to reconstruct on demand is where both the token bill and the trust problem start.
Reuse the logic, don’t rebuild it
Use AI for what it’s good at, and stop paying it to reinvent work your team already completed. The matching rules, the thresholds, the close calendar — that logic was built once and proven. The fix is letting AI call it, instead of rebuilding it every time someone asks.
That’s also what makes the output defensible, not just cheaper. When the same governed logic produces the same answer on the same data, however the question was phrased, you get something a controller can sign off on: an audit trail showing which data was used, which rules were applied, and who approved the result. Visible, understandable, repeatable, auditable — Alteryx calls this the VURA standard, and it’s the difference between a pilot that had a good week and a close process you can actually trust.
Alteryx is hosting a session on exactly this on October 16 — how to bring LLM token costs down without giving up what AI is good at. Register here.