Deterministic vs. Probabilistic Systems: How They Inform the Use of AI in Accounting

Deterministic vs Probabilistic Systems & how it informs the use of Artificial Intelligence in Accounting

Deterministic System - Classical Computing

“...a deterministic system is a system in which no randomness is involved in the development of future states of the system. A deterministic model will thus always produce the same output from a given starting condition or initial state.” [wikipedia]

What we now call classic computer systems are deterministic.

“A deterministic algorithm is an algorithm which, given a particular input, will always produce the same output, with the underlying machine always passing through the same sequence of states.”

Probabilistic Systems - Large Language Models & AI

Probabilistic systems include randomness, meaning for a given set of inputs the output will not necessarily be the same.

In LLMs there are 3 primary sources of randomness

Temperature parameter

0.0 - 1.0 value that determines how often to take the next token with a lower score.

Next Token Prediction

The next token retrieval is based on a model that internally is doing a high dimensionality probability calculation.

Underlying Model Changes

As users we have no control over changes shipped by OpenAI to chatgpt or by the other labs to their respective large language models.

Why does it matter?

Tax Law and Accounting more generally demand deterministic systems. This is because deterministic systems have the following characteristics

• Accuracy: How close can we come to the theoretically perfect Cost Segregation Study, Tax Return or Audit?

• Consistency: How often can we produce a highly accurate Cost Segregation Study, Tax Return or Audit?

• Reproducibility: For a given set of inputs can we produce the same report time & time again?

• Debuggability: If the report is incorrect can we pinpoint the exact reason why?

• Traceability: For a system to be able to withstand an audit it must be traceable. It has to be clear how each value was determined, and how each line of text was produced.

• Extensibility: Can we add features in a predictable way? Can we handle new types of inputs without breaking existing features?

Where and when to use Artificial Intelligence in Accounting

Classical computing consistently struggles to interact with the real world and unstructured data but are great at building deterministic models when operating on structured data.

LLMs are great at interacting with unstructured data straight from the real world, but are unable to act as deterministic models on their own.

Use the right tool for the job

LLMs and other machine learning tools should be used to gather and interpret raw real world data and produce structured data from them. A human in the loop is necessary to guide these models and make them more consistent over time using reinforcement learning. Examples of real world raw data include images, pdfs, 3D point clouds, satellite maps, drawings and video.

Once the structured data is gathered together and vetted for accuracy then it can be fed into a classical computing system like a web application to execute the deterministic accounting calculations.

With the appropriate use of the available tools many accounting tasks can be automated.