“Our AI unit economics are getting worse, and nobody can say why.”
Last verified: October 2026· engagement
The bill grows with usage, it arrives as one line on a cloud invoice, and every proposed cut is a guess because nobody can see cost per route or per completed task. The first question is not how much you can save. It is whether the workload is big enough, and visible enough, for optimization to be worth an engineer’s time.
Is it worth fixing?
It usually is when several of these are true:
- Eligible spend is large enough that a meaningful reduction is worth more over a year than the work costs. Below roughly $10k a month, a guide and a day of instrumentation usually beat an engagement.
- Spend is growing faster than usage or revenue, so unit economics are moving the wrong way.
- Nobody can attribute cost to a route, feature or customer, so cuts are being argued rather than measured.
- There is a planning moment coming: a margin review, a budget cycle, credits expiring, or a price rise to customers.
The decision, one question at a time
Each step links to the page that answers it. You can get most of the way without talking to anyone.
- 1
- 2
Understand the options
Should we change models, optimize prompts, or self-host?
Self-hosted vs API LLMs → - 3
- 4
- 5
Get it approved
What access, terms and acceptance criteria are involved?
The cost engagement, terms included →
What a first engagement looks like
Free read
A written read on where the spend is concentrated and whether it looks worth optimizing, from your rough spend and top routes. No call required.
Paid diagnostic
A fixed-price, read-only baseline week: cost per route and per accepted task, quality and latency, and a sized estimate per lever. Fixed price, quoted after the free read.
Stop or go
If the baseline shows the saving would not cover the work, we say so and you keep the attribution layer.
The accepted outcome is the test above: lower cost per accepted task on the in-scope routes, within agreed quality and latency limits, measured against a week-one baseline.The full acceptance criteria, exclusions and pricing terms are written out onthe engagement page, so you and procurement can see the finish line before any call.
Access and security
- Read-only access to traffic logs or a gateway export for the baseline week; redacted is fine.
- No production credentials, raw customer records or full prompt dumps are needed to assess fit.
- Mutual NDA before any data access.
When we’d tell you not to buy
Our scarce resource is senior engineering attention, so we would rather say no early. We decline or refer when:
- Spend is too small for any plausible saving to pay for the work.
- The bill is dominated by a fixed commitment or reserved capacity, so the lever is commercial, not technical.
- Nobody can grant read access to traffic or prompts.
- The real goal is to cut quality to cut cost, with no way to measure what that does to users.
Not there yet? Start withLLM cost calculator, How to reduce LLM API costs or The 7-lever cost playbook.
The engagements behind it
LLM Cost & Performance Optimization
$15–40k, or outcome-priced (25% of first-year savings) · 4–6 weeks
Self-Hosted LLM Infrastructure
$75–250k build, $10–20k/mo ops · 8–16 weeks
Questions
How much will we save?
Nobody can tell you honestly before seeing your traffic. We baseline cost per route and per accepted task in a read-only first week, size each lever against it, and only then set a target with you.
Is our spend big enough to bother?
As a rule of thumb, below about $10k a month of eligible spend, an engagement costs more than a year of any plausible saving. Use the calculator and the playbook instead; we will tell you the same thing if you ask.
Do we need to switch providers?
Rarely. Most of the saving is in routing, caching, batching and context control on the providers you already use.
Related guides
How to reduce LLM API costs
Most production LLM spend is avoidable: it comes from sending the wrong model too many tokens, too many times. Cutting it is an engineering problem — routing, caching, batching, and context control — not a matter of waiting for prices to drop.
LLM costWhy are LLMs so expensive?
LLMs are expensive because you pay per token on every call, output usually costs more than input, and agents re-read a growing context on every step — so cost compounds far faster than the work appears to grow.
LLM evalsLLM observability: what to instrument in production
LLM observability is the practice of recording enough about every model call that you can answer, after the fact, what it cost, how long it took, whether the output was any good, and what the model was actually shown. Standard APM answers none of those, because a call that returns HTTP 200 in 900ms can still be a complete failure.
LLM costHow to reduce LLM latency
LLM latency is two separate problems — the wait for the first token and the time to stream the rest — and each has its own set of fixes.
Hiring & partnersHow to hire an LLM cost optimization consultant
Hire an LLM cost consultant when your bill is large enough to matter but your team is too busy to instrument, route, and eval their way out of it themselves — and only hire one who leaves you owning the result.
Find out whether it’s worth fixing
A few structured questions, no call required. You get a written decision brief back — including whether we think you should buy anything at all.
Start the diagnostic