What Is Deterministic AI? Deterministic vs Probabilistic AI in Banking

Deterministic AI is software that always produces the same output from the same input, by following fixed calculations and written rules. In banking, it means a debt service coverage ratio, a policy test or an approval limit gives one answer every time it runs, and anyone can trace exactly how that answer was reached.

The term is used loosely. Some people use it for rules engines and calculators that have been part of banking for decades. Others use it for systems that wrap a language model inside fixed rules so that its output cannot change a decision on its own. What they share is the property a credit team cares about: same facts in, same result out, with a clear reason.

A probabilistic system produces an output by weighing likelihoods, so the same question can lead to different answers. Large language models work this way. They generate text one piece at a time, choosing each piece from a range of likely options. That is what makes them flexible enough to read a messy tax return or draft a memo in the bank's style.

It is also why they vary. Ask the same model the same question twice and the wording, the emphasis and sometimes the numbers can change. Federal Reserve Governor Michael Barr made this point in an April 2025 speech, noting that because generative AI usually involves stochastic processes, answers can differ when the same query is asked at different times. A model can also produce statements that sound right but are not supported by any source, which is usually called hallucination.

Neither approach is better in general. Each is suited to different work. The table below compares them on the points that matter most in a credit process.

A credit decision has to be repeatable and explainable. If two analysts spread the same borrower, the coverage ratio should match. If a request exceeds an officer's authority, it should go to the next signer every time, not most of the time. In the same April 2025 speech, Governor Barr said that varying answers are tough to square with banking, "where decisions must be well-controlled, numerically and legally precise, explainable, and replicable."

There is a legal side too. Under Regulation B, a lender that declines credit must give the applicant the specific principal reasons for the adverse action. A reason that changes depending on when the model was asked does not meet that standard. And when a loan goes bad, the bank needs to show what it knew and which rule it applied on the day it decided. Only a deterministic record of calculations and rules can do that reliably.

None of this means language models have no place in credit. They are very good at the reading and writing that takes up much of an analyst's week. The trick is to give them jobs where variation is acceptable because a fixed calculation or a person checks the result.

The pattern that works in practice splits the job into three roles, each with a clear owner. A deterministic engine computes and enforces. A model drafts and challenges. A named person signs. The model never holds the pen on a number that drives a decision, and the engine never makes a judgment call that belongs to a person.

Partly, but not enough to rely on for a credit decision. Settings such as a temperature of zero make a model's output much more consistent, and fixing the model version helps. Even so, outputs can still shift when the provider updates the model, when the surrounding software changes or when a prompt is worded slightly differently. Google DeepMind CEO Demis Hassabis described today's systems in February 2026 as having jagged intelligence, strong at some tasks and weak at others, which is another way of saying their behavior is uneven.

The more reliable approach is to put determinism around the model rather than inside it. Keep calculations in code the bank controls, check every model step against fixed rules before it takes effect, and require a person to confirm anything that drives a decision. Then it does not matter if the model words a draft differently on Tuesday, because the numbers and the decision path are the same.

On April 17, 2026, the Federal Reserve, OCC and FDIC issued revised model risk management guidance (Federal Reserve SR 26-2, OCC Bulletin 2026-13), replacing SR 11-7 from 2011. The definition of a model in that guidance excludes simple arithmetic calculations and deterministic rule-based processes with no statistical, economic or financial theory behind them. It also states that generative AI and agentic AI models are novel and rapidly evolving and are not within its scope.

That does not mean generative AI is ungoverned. The guidance says a bank's own risk management and governance practices should decide the controls for tools it does not cover, and the agencies said they plan to issue a request for information on banks' use of AI, including generative and agentic AI. For a credit team, the practical reading is simple: keep calculations and rules deterministic and well tested, and wrap any language model in controls the bank sets and can show working.

Vendors often say their AI is accurate. The better questions are about what happens when it is not, and whether the parts that decide are fixed. Ask to see each answer on your own files.

Aarvion Risk OS uses this split. AI reads documents and drafts credit work, with every figure tied to its source page. Calculations, policy tests and authority routing run as fixed rules the bank sets. Every AI step is checked against those rules, which can allow it, hold it for the right signer, block it or stop all activity, and each step is recorded.

Decisions such as approve, grade, waive and amend stay with the bank's named people. Banks typically start with a 90-day pilot on one workflow and compare the results with their current process.