LLM cost & reliability audit
You are already running AI in production and the bill is climbing faster than the usage. We audit the engineering — routing, caching, token efficiency, retries and failure handling — and hand you a prioritised list, each item costed against your own traffic.
The bill is an
engineering artifact.
Two teams can build the same feature and receive invoices that differ by a factor of five to ten. Nothing about the market explains that gap — it is entirely a function of decisions made inside the code: which model handles which request, whether the repeated instruction block is cached, how much context is resent on every turn, and what happens when a call fails.
The reason it goes unnoticed is that none of it breaks anything. A retry storm looks like a working system with a higher bill. An oversized model looks like a working system with a higher bill. Nobody gets paged.
The three largest levers are usually caching, routing and prompt discipline. Provider pricing is the one published number here: cached input reads bill at a fraction of the base rate, which is up to 90% off repeated input tokens. Routing and prompt discipline compound on top of that, but the size of each depends entirely on your own traffic mix — so the audit measures them against your bill rather than quoting you a range we cannot stand behind.
What we examine
Cost and reliability are audited together, because most of the expensive failures are reliability problems wearing a cost disguise.
Token accounting
Where the tokens actually go, broken down by feature and request type. Almost every audit finds a single code path responsible for a disproportionate share of the bill, and the team is surprised by which one.
Prompt efficiency
Context resent when it could be cached, instructions that grew by accretion, examples that stopped earning their tokens, and conversation history carried further than it is used.
Caching
Whether prompt caching is enabled at all, whether the prompt is ordered so the cacheable prefix is stable, and whether cache lifetimes match your real traffic pattern. This is the most commonly missed saving we find.
Model routing
Which requests genuinely need the strongest model and which are being over-served. We test the cheaper option on your real traffic rather than assuming — a routing change that costs accuracy is not a saving.
Retry & failure behaviour
Retry storms, missing backoff, unbounded agent loops, and calls that fail silently and get retried by a caller who does not know. This is where surprise invoices come from.
Reliability posture
Timeout handling, fallback behaviour when a provider degrades, rate-limit handling, and whether a partial failure surfaces to a human or is quietly swallowed.
Observability gaps
What you cannot currently see — per-feature cost, error rates by type, latency percentiles. Several findings are usually "you have no way to know this," and that is a finding.
A prioritised list, not a slide deck
Every finding arrives with four things: the evidence from your own traffic, an estimated monthly saving in Canadian dollars, the implementation effort in days, and any risk the change introduces. Sorted by return, so you can draw a line anywhere and know what you are giving up.
Findings we judge not worth doing are included and marked as such, with the reasoning. An audit that only reports the flattering findings is a sales document.
You can implement it with your own team — the report is written for that — or engage us for the changes separately. There is no obligation either way, and the audit fee is not credited against implementation work, so we have no incentive to inflate the list.
The full cost arithmeticWhat we need from you
Read access to the relevant source, your provider's usage and billing export for the last 60–90 days, and 90 minutes with the engineer who built it. Where source access is not possible, we work from logs and instrumentation — it costs some depth, and we will say which findings are affected.
We sign your NDA, or ours if you prefer. Nothing from your codebase is used anywhere else.
Fixed fee, no contingency
Prices in Canadian dollars, verified 23 August 2026.
Audit
from $9,500CAD · One production system, up to three AI featuresIncluded
- All seven areas above
- Findings costed against your own traffic
- Prioritised by return, with effort and risk
- Working session with your engineers
- Written report you own outright
Not included
- Implementing the changes
- Model or vendor contract negotiation
Implementation
Per findingCAD · Fixed price quoted against the agreed listIncluded
- Fixed price per finding, chosen by you
- Before-and-after measurement on real traffic
- Accuracy checked, so no saving costs quality
- Changes delivered as reviewable pull requests
Not included
- Any obligation to use us — the report stands alone
When this pays
$3,000+CAD / month · Current AI spendWhy the floor
- Below it, the audit costs more than it can return
- We will decline rather than take the engagement
Instead
- Under the floor, we will point you at the two or three changes that matter, at no charge
When you should not buy this
Below CAD $3,000 a month in AI spend. Even a 60% saving is under $1,800 a month, and the audit does not repay itself before the system changes underneath it. Ask us the free version of the question instead.
When the system is about to be rewritten. Auditing code that will not exist in a quarter is a waste of your money. Wait, and build the cost controls into the rewrite — which is much cheaper than retrofitting them.
When the real problem is that it doesn't work. If accuracy or reliability is the actual complaint, a cost audit answers the wrong question. That is an evaluation and remediation engagement, and we would rather scope it correctly than sell you this one.
Questions worth asking first
This is for AI you are already running in production, above $3,000 CAD a month in current spend. Under the floor, we will point you at the two or three changes that matter, at no charge.
What does the audit cost, and when is it worth doing?
From $9,500 CAD for one production system, up to three AI features, over 2–3 weeks. It pays at current AI spend of $3,000+ CAD a month — below that floor the audit costs more than it can return, and we will decline rather than take the engagement. Prices in Canadian dollars, verified 23 August 2026.
What do you need from us to run it?
Read access to the relevant source, your provider's usage and billing export for the last 60–90 days, and 90 minutes with the engineer who built it. Where source access is not possible, we work from logs and instrumentation — it costs some depth, and we will say which findings are affected. We sign your NDA, or ours if you prefer, and nothing from your codebase is used anywhere else. Access is least-privilege and per-engagement, and it is revoked at project close or on your request.
Do we have to hire you to implement the findings?
No — the report stands alone. You can implement it with your own team, and the report is written for that, or engage us for the changes separately. There is no obligation either way, and the audit fee is not credited against implementation work, so we have no incentive to inflate the list.
Could a cheaper model cost us accuracy?
We test the cheaper option on your real traffic rather than assuming — a routing change that costs accuracy is not a saving. Where we implement a finding, there is before-and-after measurement on real traffic and the accuracy is checked, so no saving costs quality.
Does the report include findings you think we should skip?
Yes. Findings we judge not worth doing are included and marked as such, with the reasoning. An audit that only reports the flattering findings is a sales document. Every finding is sorted by return and carries the evidence from your own traffic, an estimated monthly saving in Canadian dollars, the implementation effort in days, and any risk the change introduces.
Bring last month's invoice
Your provider's usage export and a description of what the system does is enough for us to tell you, in one call, whether an audit is likely to repay itself.
Book a scoping call