You're spending too much on AI.

Most AI workflows contain large amounts of work that never needed AI in the first place: parsing, formatting, routing, lookup, and conditional logic. Code handles those tasks more cheaply, more quickly, and with identical output for identical input.

Before routing work to a model, ask whether a function can do it. If the answer is yes, write the function. Reserve inference for the edge cases that actually require reasoning.

Benchmark: in the greenchemistry.ai rebuild, 90% of the original workflow became deterministic Python. Cost dropped from $5.00 per run to under $0.005.

The frontier model is rarely the right default. Structured extraction, classification, constrained generation, and predictable formatting often perform better on smaller models when speed, consistency, and cost are the real metrics.

The McLaren principle applies: you do not put a tow hitch on a Formula One car to take the garbage to the dump. Match the vehicle to the job, and keep frontier capability for the small share of work that genuinely needs it.

Starting hypothesis: most enterprise workflows need frontier capability on less than 15% of their tasks.

AI systems often get built for the workflow as people imagine it rather than the workflow as it actually exists. Real audits expose redundant steps, hidden deterministic work, and unnecessary inference that was only there because nobody diagrammed the system first.

Use a determinism scorecard from 0 to 4: from fully deterministic work that should immediately become code to genuine frontier tasks that need explicit justification.

Typical finding: initial audits place 60–80% of organizational AI usage at score 0 or 1.

AI costs spiral when teams treat inference cost the way old cloud teams treated egress: a number they will optimize later. Later never comes. The workflows harden first and the waste becomes cultural.

Instrument cost and latency at the point of call. Track cost per workflow run, not just cost per API request. Review the trend alongside classification updates so cost discipline compounds over time.

Local inference eliminates per-token API costs for qualifying workloads and keeps sensitive data inside the organizational perimeter. For structured, high-volume tasks, that is often both the cheaper and safer architecture.

Apple Silicon is a strong current example because unified memory, efficient matrix hardware, and low idle power make intermittent inference practical. These are engineering claims, not brand claims.

Constraint: local-first is powerful, but only when the workload is sufficiently repetitive and can be handled by sub-frontier models.

If a workflow requires auditability, it cannot rely on a raw stochastic call as its core mechanism. The architecture has to be designed around determinism from the beginning.

That makes this more than a cost discipline. It is also a compliance, reliability, and governance discipline. The more deterministic code you extract, the more of the workflow becomes inspectable by default.

Convergence: the same architectural choices that make AI cheaper also make it easier to defend.