AI infrastructure · Serving · Routing · Residency
Most enterprise AI spend goes on sending easy requests to expensive models because nobody built the routing layer. The bill scales with usage; the value does not. That gap is an infrastructure problem with a known fix.
The cost
AI budgets do not fail on the headline rate. They fail on four line items with no owner.
Classification, extraction and routing tasks sent to the largest available model because it was the one wired up first. The same task on a correctly sized model is an order of magnitude cheaper and typically faster.
A timeout, a malformed response, a tool call that failed schema validation. Each retry is billed. Retry rate rarely appears on the invoice as a line, so it rarely appears in the forecast either.
Prompts that carry the whole policy document on every request because caching was never configured. You pay for the same tokens thousands of times a day.
A workload that could run on a cheap regional endpoint routed through an approved one instead, because the data was never classified by where it truly had to stay.
The build
A serving layer that makes the cost of a decision visible and then lets you govern it.
A task taxonomy mapped to model tiers, with a measured quality floor per tier, not a guess. Easy work goes to small models and stays there.
Spend attributed to the business action that caused it — an invoice classified, a document extracted — not to a token count no one can act on.
Stable context cached at the provider boundary, so the policy document is paid for once, not on every call.
A data classification that says which workloads may leave a jurisdiction, so the expensive compliant path is used only where it is genuinely required.
Retries counted, capped, and surfaced. A workload that retries 30% of the time is a defect, not a cost of doing business.
Provider outage does not stop the queue. Degraded service on a smaller model beats an incident.
The economics
These are the levers, with the order of magnitude each one operates at. What they are worth depends on your volume and mix, which is the first thing an assessment establishes.
Why this firm
This is the same discipline as a cutover gate, applied to spend: a number nobody can attribute is a number nobody can control. We instrument the decision before we optimise it.
Questions
Because a token count is not a unit a finance function can govern. Cost per invoice classified, per document extracted, per case triaged — those map to volume forecasts and to the value of the work being done. Token spend only tells you that the bill went up.
It does if you route on price alone. We establish a measured quality floor per task before anything moves tier, and the routing layer holds that floor. Tasks that genuinely need frontier reasoning keep it; the point is that most of them do not.
The opposite. A routing layer is what makes providers substitutable — the abstraction is ours, and swapping a tier is configuration, not a rewrite. Provider lock-in usually comes from having no routing layer at all.
Directly. The highest-value agent workloads sit against systems of record, which is where residency, audit and retry semantics stop being theoretical. We built the gate for ledgers first.
Next
We will map spend to business actions, identify the tier mismatches, and give you a written read on what is recoverable. You keep the analysis either way.