Enterprises are discovering an uncomfortable truth about generative AI: a lot of the token budget going into “agentic” workflows is being spent on tasks that were never a good fit for an LLM in the first place. Comparing files, reconciling numbers, checking balances — this is deterministic work, and asking a stochastic model to do it burns tokens, adds cost, and introduces risk that shouldn’t exist for tasks that already have a right answer.
NextWave, a digital acceleration consultancy for financial services and an Alteryx partner, set out to test that theory directly — and the results are now shaping how both companies talk about responsible, cost-controlled enterprise AI.
When Alteryx handles the deterministic work first, such as comparing, reconciling, and validating data, organizations can reduce the cost of paying a model to redo work that already has a right answer. In our verified benchmark testing with NextWave, that approach cut token consumption by up to 95% (a 20x reduction) on a defined reconciliation workload. Here is how the test worked.
The Test: Office of the CFO reconciliation
NextWave bench-tested a common Product Control workload: a daily front-office-to-back-office reconciliation, validating consistent views of open trades, valuations, risk, and cash balances across a data set of 120 trades, then explaining any differences with suggested next-best steps.
They ran the task two ways.
The first approach was purely stochastic: hand an LLM the front-office and back-office files directly and ask it to compare trades, valuations, risk, and cash balances, flag differences, and suggest next steps.
The second approach used an Alteryx One workflow, published for the LLM to call using the Model Context Protocol (MCP). Instead of asking the LLM to do the comparison itself, the query asked it to execute the Front Office Back Office Data Reconciliation workflow on Alteryx One — letting Alteryx handle the comparison logic deterministically, with explainability and audit built into the workflow execution, while the LLM stayed focused on what it’s actually good at: interpreting the output and recommending next steps.
The results
The gap between the two approaches was stark:
- 20x fewer tokens consumed by the Alteryx One + LLM workflow versus the purely stochastic approach
- 100% accuracy on the reconciliation, versus 2 missed breaks out of 120 trades when the LLM worked unaided
- Enterprise-grade governance, explainability, and auditability built into the deterministic workflow layer — something a pure LLM comparison can’t guarantee
LLMs are a fantastic step forward in technology, but care needs to be taken in knowing when to use them and when not to. It’s about knowing what tool to use for what problem and having those tools in your tool bag.”
It’s best to let deterministic workflows do the work they’re designed for, and let LLMs focus on reasoning, summarisation and next best actions. This is how enterprises can scale AI while controlling GenAI costs and improving trust.
To see how deterministic logic enables cost-effective, accurate AI responses, we encourage you to explore the Month End Close Acceleration AI-Ready Starter Kit.
Why this matters beyond one reconciliation task
This isn’t an isolated result. Alteryx’s own internal testing, published alongside its recent AI capability launches (Ask Alteryx, Agent Studio, an MCP Server, and Alteryx Skills), found similar patterns at scale: up to 93% token reduction and an 85% speed increase on tasks involving raw, ungrounded data, and up to 83% token cost reduction with a 65% speed increase on tasks using clean, already-governed data.
The common thread is what Alteryx calls the VURA framework — making business logic visible, understandable, repeatable, and auditable — so that AI agents draw on governed, deterministic logic instead of re-deriving it stochastically every time. NextWave’s reconciliation test is a concrete, numbers-backed example of that principle in production, not just in theory.
The bigger lesson
As more enterprises move agentic AI from pilot projects into production, the cost and accuracy of “let the LLM figure it out” approaches will become harder to ignore. The work of NextWave and Alteryx points to a more sustainable pattern: use deterministic, governed workflows for the deterministic parts of a process, and reserve the LLM for reasoning, summarization, and recommending what to do next.
NextWave and Alteryx are actively helping financial services clients apply this pattern — using LLMs alongside the Alteryx One platform to solve reconciliation and other Office of the CFO problems, and using LLMs to accelerate adoption of the Alteryx One platform itself.