Waste Reduction Is Not the Goal. Visibility Is.
Most AI cost optimization conversations start in the wrong place. Teams want to know how to spend less, so they look for things to cut. Smaller models, shorter prompts, fewer agents. Sometimes this works. More often, it reduces capability without materially reducing cost, because the cuts happen to the wrong things. The teams that optimize AI costs well are not starting from a desire to spend less. They are starting from a desire to see clearly. The spending reduction follows.
Taiichi Ohno figured this out at Toyota in the 1950s, and his insight is worth understanding before you open a single API bill.
The Toyota Lesson Most People Get Wrong
Ohno's Toyota Production System is usually taught as a cost-reduction program. That is not what it was. It was a visibility program. Ohno's central observation was that waste is invisible by default. You do not notice it because it looks like normal operations. Inventory sitting in a warehouse looks like preparedness. Workers waiting for parts look like a staffing issue. Defects discovered at the end of the line look like quality control doing its job.
His taxonomy of the seven types of waste, which he called muda, was not a list of things to eliminate. It was a taxonomy for making the invisible legible. Once managers could name the type of waste they were observing, they could measure it, assign it a cost, and decide whether to fix it. Before the taxonomy existed, a lot of operational loss simply had no name and therefore no owner.
The sequence at Toyota was: make waste visible, give it a name, assign it a cost, then eliminate it. Teams that tried to skip to elimination without the visibility step kept optimizing the wrong things.
The Seven Types of AI Spend Waste
Engineering teams running production AI systems have their own version of Ohno's taxonomy. Here are the seven types of AI spend waste worth naming, because named waste is waste you can actually address.
Unnecessary Retries
Agent loops that retry failed API calls without an idempotency check generate redundant costs. A workflow that fails at step three and retries from step one has paid for steps one and two twice. In high-volume systems, retry waste compounds quickly and is almost never visible in cost dashboards because it looks like normal traffic.
Over-Provisioned Context Windows
Sending a 100,000-token context to a model that only needs to process the last 2,000 tokens is not a prompt problem. It is an architecture problem. Many systems were built when context costs were less of a concern and have never been revisited. The waste here is often large and addressable without changing what the model outputs.
Duplicate API Calls
Parallel agent architectures that make independent calls for the same input, or systems without caching that re-process the same documents on re-runs, generate duplicate costs. These are usually easy to find with call-level logging and disproportionately expensive to ignore.
Idle Agents
Background processes that poll APIs or maintain active sessions when no user is waiting are a form of standby waste. This is the electric motor left running overnight that Edison's engineers found first. The agent is doing something, it just is not producing anything while it does it.
Wrong Model for the Task
Using a frontier reasoning model for classification, extraction, or simple formatting is overproduction in Ohno's terms. You are producing more capability than the task requires, and paying for the excess. The fix is a routing layer, not a capability reduction.
Unattributed Background Processes
Scheduled jobs, batch reprocessing runs, and maintenance workflows that hit the API without tagging are a bookkeeping problem with cost consequences. When attribution is missing, these processes become invisible line items that grow over time and are only discovered when someone looks at a spike and cannot explain it.
Stale Caches
A caching layer that does not expire correctly will either miss too many requests (waste from redundant calls) or serve outputs from outdated model versions (waste from quality degradation that triggers human review and correction). Cache hygiene is an optimization task that is easy to defer and expensive to ignore.
The Playbook, In Order
Start with attribution instrumentation. You need cost tagged by workflow, by team, and by task type before any of the optimization moves below make sense. Without it, you are optimizing blindly. Once attribution is in place, rank your workflows by cost per output and prioritize the top three for the waste taxonomy review. Most teams find that two or three waste categories account for the majority of addressable cost, and fixing those first creates enough headroom to fund the rest of the work.
Ohno's insight was that waste does not reduce itself. It has to be made visible first, then named, then assigned a cost, then attacked in priority order. The Toyota production system is still the clearest articulation of that sequence, and it applies to AI infrastructure in 2026 with almost no translation required. Oberhahn is built to do the visibility step: instrument your AI spend, surface the taxonomy, and make the waste legible before you spend time on the elimination.