Global – October 27, 2026 -- Alteryx argues that unpredictable output, not token cost, is the main barrier to scaling AI agents in the enterprise: the same request to a large language model can return a different answer each time.
Token burn rises with agent complexity, but variance is the bigger problem
Agents must reason through business logic, infer context, read results and choose the next action, sometimes several times in a single instance. Restarting a process or re-deriving business logic adds to token costs. Alteryx says the larger obstacle is that LLMs generate outputs from a probability distribution. In its words, "you cannot scale a business on randomness."
Audit-grade calculations must meet a four-part VURA standard
Alteryx says calculations presented to auditors and board members must be visible, understandable, repeatable and auditable (VURA). Given the same inputs and business rules, they must return the same result every time. Writing tasks such as emails tolerate variability.
LLMs can write code for deterministic tasks. But finance leaders who know revenue-recognition rules best often cannot validate a Python script. Alteryx calls the result a black box.
One-off AI builds leave finance teams unable to verify the logic
Alteryx describes this as the "single-player problem." An AI-built revenue-recognition agent may produce high-level metrics from enterprise data in minutes using models such as Gemini or Claude. If finance leaders cannot parse the code, they cannot trust or operationalize it. A Git repository would not fix this, because finance teams are unlikely to adopt Git-based version control and code review.
Alteryx workflows replace code as the shareable, reviewable artifact
Alteryx proposes having the AI model build the solution as an Alteryx workflow instead of code. Business users can then inspect the data used, the logic applied and the calculations performed at each step. They can also use AI to adjust the workflow.
A workflow named "Revenue Recognition" would hold a machine-readable record of how a company recognizes revenue, including formulas and internal business logic. AI agents can read it, so the model does not need to re-derive the logic and burn tokens each time.
Workflows can also serve as executable tools for AI models. Users can call a workflow through MCP from their LLM interface whenever a calculation is needed.