Fall Back to Deterministic When the AI Is Unsure

Keep a deterministic path underneath the model and route to it whenever the model is uncertain or overkill. The best AI-native products are not pure AI. They are hybrids: a model handles the ambiguous, open-ended cases, and plain rules handle the ones that do not need a model at all or that the model gets wrong too often. This is not a compromise or a lack of ambition. It is how you get a system that is both capable and reliable, instead of one that is impressive in a demo and flaky in production.

The thesis: a deterministic fallback is a strength, not a crutch. The question is not "AI or rules," it is "which cases go to which, and how do you decide."

Why pure AI is the wrong default

Routing everything through a model is expensive, slow, and less reliable than it needs to be. Plenty of inputs are unambiguous. A transaction from the same vendor you have categorized fifty times does not need a language model to guess, it needs a lookup. Sending it to the model anyway costs latency, costs money, and adds a chance of a fresh wrong answer to a case that was already solved. That is a bad trade, and it is the same waste as failing to route cheap steps to cheaper paths.

Worse, some inputs are ones the model is simply unreliable at, and you know which ones from your eval work. For those, a rule beats a model not because rules are smarter but because rules are consistent, and consistency is what production needs.

How to decide what goes deterministic

Route by confidence and by category.

Use the model where the input is genuinely ambiguous and rules cannot cover the space: free text, novel documents, natural-language intent. That is what models are for.

Use rules where the mapping is known and stable: exact matches, lookups against prior decisions, hard constraints that must never be violated. If there is a policy that says "this always does that," encode the policy, do not ask a model to re-derive it every time and hope it agrees.

And build the handoff on confidence. When the model runs but comes back uncertain, do not ship the guess, fall back to a safe deterministic default or escalate to a human. Deciding this well depends on having an honest confidence signal rather than a fake precise score, and it is the same instinct as designing the fallback before the model fails.

The hard constraint rule

Here is the line I do not cross: never let a model override a hard constraint. If money must reconcile, if a legal deadline is fixed, if a value has to be inside a range, that check is deterministic code, full stop. The model can propose. The rule disposes. You do not want a probabilistic system deciding whether a filing meets a statutory rule, you want it drafting inside guardrails that a deterministic layer enforces. This is the backbone of making AI decisions defensible: the constraints are auditable rules, not model output.

That is how AI-native bookkeeping actually works in practice. The model reads and suggests. Deterministic reconciliation checks that the numbers add up. Neither alone is enough. Together they are trustworthy.

When a hybrid is the wrong answer

The caveat, so this does not become dogma. Do not build a sprawling rules engine to avoid using the model where the model is clearly better. If you find yourself hand-coding hundreds of brittle rules to handle natural-language variation, you are fighting the wrong battle, and the model would do it better and cheaper. The hybrid is about routing each case to the tool that fits it, not about minimizing model calls for their own sake.

The mistake in both directions is treating this as an identity ("we are an AI company" or "we are a rules company") instead of an engineering decision made per case.

The bottom line

A deterministic path underneath the model gives you a floor. When the model is unsure, slow, unavailable, or simply not needed, the system still does something correct instead of something clever and wrong. We build that floor into every model-backed feature at Girard AI, because the point of AI-native is reliability with capability, not capability that falls over when it matters. Model for the ambiguous. Rules for the certain. A confident handoff between them.