Deterministic over generative: the economics of software that runs at ~$0 per request

There is a default move in AI product building right now: take a workflow, wrap a large language model around it, charge for access. It works as a demo. It is also, for a large share of knowledge work, the most expensive and least reliable way to ship the thing.

My claim is narrower than it sounds. Most knowledge work is rules plus a little judgment. A credit score, an eligibility check, a tax-residency test, a study schedule, an astrology chart - these are mostly procedure. There is a rulebook. There are inputs. There is a defined output. The judgment is real but it lives at the edges, not in the centre. When the centre is procedure, an LLM call per request is the wrong tool. You are paying a probabilistic engine to re-derive, from scratch, something that was knowable in advance.

The cost shape is the whole argument

Per-request inference is a recurring variable cost. Every single use is another call, another bill, forever. It scales linearly with success: the more people use your product, the more you pay, every month, with no end date. That is a fine shape for some things and a terrible shape for a calculator someone runs forty times in an afternoon.

Deterministic logic has the opposite shape. You spend effort once to encode the rules, and after that the marginal cost of one more request is effectively zero. It is a function call. It runs in the browser or on a cheap server. There is no per-use bill because there is no per-use inference. The cost lives at build time, not run time.

Use AI at build time to manufacture the rules and content once. Do not pay an inference bill at run time to re-derive them on every request.

This is not anti-AI. It is a question of where the AI earns its cost. We use AI heavily - at build time. AI helps structure the rulebook, draft the content, generate test cases, and stress the logic. That work happens once and gets compiled into deterministic software. The user never triggers a model. They trigger a function.

What this looks like in things that actually exist

This is not theory I am hoping to test later. There are free, in-browser tools running on this exact model today:

• A CRS calculator that scores Canadian Express Entry profiles against the published IRCC point tables. No model call - it is the rules, encoded.
• A TEF score estimator that maps practice performance toward the relevant bands. Arithmetic and lookup, not inference.
• A Kundli generator that computes a Vedic astrology chart from birth data using an ephemeris. Astronomy and rules, deterministic to the minute.

Each of those runs at roughly zero cost per use, returns instantly, and gives the same answer every time you give it the same inputs. There is no usage meter ticking in the background. That last property is not a nice-to-have. It is the second half of the argument.

Determinism is testability, and testability is trust

Same input, same output. That sentence is doing a lot of work. It means you can write a test for it. It means you can trace why a result came out the way it did - back to a specific rule, a specific table, a specific line. It means an auditor, a regulator, or a sceptical customer can be shown the logic and check it. None of that is available when the answer is a fresh sample from a probability distribution that may differ tomorrow, or differ on the same input twice.

In regulated and high-stakes domains - immigration, tax, law, education - this is not a preference. It is close to a requirement. "The model said so" is not an answer you can defend. "Here is the rule, here is the table, here is the calculation, and it will return this every time" is. Determinism turns a black box into something you can stand behind.

Where AI genuinely earns its cost

I am not pretending the rules cover everything. They do not. There is a real category of work where an LLM is the right tool and worth paying for at run time: genuine judgment on a novel case, messy unstructured input that has to be read and interpreted, and drafting where a human will review the output anyway. Summarizing a long document, turning a vague question into a structured one, generating a first draft - these are model-shaped problems. The mistake is not using AI. The mistake is using it for the procedural ninety percent that a rulebook already answers.

The human gate on the dangerous steps

For the high-liability moments - the ones where a wrong answer costs someone real money or a real visa - the right architecture is not "more AI" and it is not "fully automated." It is deterministic computation up to the edge of judgment, then a human approval gate. The software does the heavy, repeatable work and hands a clean, traceable case to a licensed professional for the call that should not be automated. The machine is fast and consistent. The human owns the liability. Neither pretends to be the other.

What it means for the business

Put the pieces together and you get a different business shape than the usual AI wrapper. Margin: when marginal cost per request is near zero, you are not handing a slice of every sale to an inference provider, so growth compounds instead of leaking. Reliability: deterministic software does not have a bad day, get rate-limited at the wrong moment, or quietly change its mind after a model update. Trust: you can show your work, which is the thing that actually closes regulated customers.

I should be straight about where this sits. This is a thesis-and-architecture piece, not a results piece. Million Dollar Studio is at zero monthly recurring revenue today. I am not claiming traction I do not have. What I am claiming is that the cost shape, the determinism, and the trust properties are real and already running in the free tools above - and that they point at a better default than wrapping a model around every request and hoping the bill stays small.

The crane builds the building, then comes down. AI is the crane. It belongs at build time, manufacturing the rules and the content. What ships and runs - what the customer touches every day - should be deterministic software that costs almost nothing to run, returns the same answer every time, and can be checked by anyone who asks. That is the bet.