AI cost engineering
Cut your AI costs without cutting what your customers get.
We are an AI cost engineering firm for companies running large language models in production. We find where the money goes, ship the savings that cannot affect quality first, and prove the ones that can — before a customer ever sees them.
Everything is judged on one number: cost per successful task. Not cost per token — because a cheaper answer that causes a retry, an escalation or a lost customer is not a saving, and a per-token dashboard will tell you it was.
Fixed-fee, two-week audit. Credited in full against implementation work. UK-based, Azure-native.
The problem
Your invoice tells you the total. It does not tell you the truth.
Almost every team we speak to can read their AI bill and almost none can explain it. The provider dashboard groups spend by model and by day. Your business runs on features, customers and outcomes. Nothing in the default reporting connects the two, so the conversation stalls at “it went up again”.
That gap is expensive in a specific way. It makes the cheapest-looking change the most attractive one. Moving a workload to a smaller model looks like an instant saving per call — right up until the retry rate doubles, support volume rises, and the same task now takes three calls instead of one. Every one of those extra calls is a perfectly ordinary, correctly-priced, successful API request. Nothing in a per-token view catches it.
The fix is not a better dashboard. It is a better denominator.
What you see
Cost per token
Measures your infrastructure. Falls whenever you switch to a smaller model, regardless of whether the answer worked. Cannot see quality at all.
What is hidden
The cost of failure
Retries, rejected outputs, abandoned sessions and escalations are all billed and all look like successes to the provider. In most estates this is the first genuinely uncomfortable number an audit produces.
What decides
Cost per successful task
All variable AI cost, failures included, divided by the customer tasks actually completed. Denominated in units of business, so it survives a conversation with a CFO.
The method
Free savings first. Judgement calls last. In that order, deliberately.
There are roughly a dozen levers in an AI bill. Most advice starts with the one that carries the most quality risk. We work in strictly increasing order of risk, which means the engagement pays for itself before anybody has to spend political capital on a trade-off.
Changes that cannot alter your output
Prompt caching, batch endpoints, request deduplication and output discipline. The model receives exactly the same instruction and returns exactly the same class of answer — only the billing changes. No evaluation harness required, no product sign-off, no risk conversation.
- Reorder prompts so the stable prefix is cacheable
- Move latency-insensitive work to batch endpoints
- Deduplicate identical calls and fix double-billed retries
- Stop generating output that is truncated or discarded
Changes you can measure your way through
Retrieval trimming, conversation history strategy, calls per task and retry hygiene. These change what the model sees, so we validate each one against a golden dataset built from your own traffic. The decision becomes empirical rather than political.
- Trim retrieved context to where accuracy stops improving
- Replace quadratic history resends with summaries and windows
- Collapse multi-call chains that do not earn the extra call
- Make silent retries visible, then make them unnecessary
Changes that need a judgement
Model routing, prompt reduction, distillation and capacity commitments. The biggest numbers live here and so does all of the risk, which is why they come last — after the golden sets exist and after the free savings have already paid for the work.
- Route each task class to the cheapest model that passes
- Cut inert instructions out of prompts that are billed every call
- Consider distillation for narrow, high-volume tasks
- Size capacity commitments against measured demand
The full playbook is written up, free and in detail, in how to reduce LLM API costs. We would rather you tried it yourself than hired us for something you could ship in a fortnight.
Services
Six engagements. Every one of them starts with a measurement.
Most clients begin with the audit and take the implementation work in phases they approve one at a time. Nothing here requires the next thing to be bought.
AI Spend Audit
A two-week, fixed-fee forensic review of your AI bill. We instrument your calls, attribute every pound to a feature and an outcome, and return a ranked list of savings with the expected quality impact of each one.
2 weeksLLM Gateway Implementation
The single OpenAI-compatible endpoint every AI call in your estate goes through. It routes, retries, caches, enforces budgets and records what everything cost — deployed in your own cloud, not ours.
3–5 weeksModel Routing & Evaluation
The largest single lever in most AI bills is that one model serves every request. We segment your traffic by task, find the cheapest model that still passes, and prove it on your own traffic before it reaches a customer.
4–6 weeksPrompt, Context & Caching Engineering
The cheapest token is the one you never send. We restructure prompts for caching, trim context that adds cost but not accuracy, and move eligible work to batch — changes that are provably output-identical.
2–4 weeksAzure AI Cost Optimisation
Azure-specific AI cost work: choosing between provisioned throughput and pay-as-you-go, getting deployment types and regions right, and wiring model spend into the Azure cost tooling your finance team already uses.
3–4 weeksAI FinOps Retainer
Cost reductions decay. Prompts grow, traffic shifts, models get deprecated and new features ship without a budget. The retainer keeps the measurement running and the backlog of savings moving.
Rolling, 3-month minimumHow we work
Four steps, and you can stop after any of them.
Measure
Two weeks. We trace every model call and join it to whether the customer's task actually succeeded. You get a cost per successful task per journey, and a waste inventory.
Bank the free savings
The output-identical work ships first — caching, batching, deduplication. It cannot change your product, so it does not need anyone's permission, and it usually covers the cost of the audit.
Prove the trades
Golden datasets from your traffic, candidates replayed in shadow, quality scored before a customer sees anything. The routing decision stops being an opinion.
Gate it, then keep it
An evaluation gate in CI blocks regressions, budgets cap runaway spend, and the unit economics stay on a dashboard. Savings that are not defended decay within two quarters.
How we operate
Three commitments that make the advice worth taking.
No provider incentives
We have no reseller relationships and take no commission from any model provider or cloud. A recommendation to move traffic off a product costs us nothing to make.
Your cloud, your keys, your data
Everything we build is deployed into your subscription. Prompts and completions stay inside your network boundary. There is no Rise10x-hosted middleman holding your traffic.
Numbers reconciled, not estimated
List prices change constantly and negotiated rates differ. Every projection is reconciled against your actual invoice before it is reported to anyone, including by us.
Rise10x runs its own products on this stack — a LiteLLM gateway on Azure Container Apps as the single AI egress point, cost telemetry joined on a task ID, and an evaluation gate in CI. The tooling we bring is the tooling we operate. More about how we got here.
Guides
The whole playbook, published.
None of this is proprietary. The value we add is doing it accurately on your traffic, in two weeks, with the measurement infrastructure left behind — not knowing something you cannot read.
How to reduce LLM API costs
The complete playbook, in the order we actually apply it: the changes that cannot affect your output, then the ones that can — and how to tell the difference before you ship.
14 min readCost per successful task
Per-token dashboards will tell you a change was a success while your retry rate doubles. Here is the denominator that catches it, and how to build it so no single service can quietly fake the number.
11 min readPrompt caching explained
Caching is prefix-based, which means one timestamp in the wrong place can cost you the entire discount. Here is how it actually works and how to structure prompts so it fires.
10 min readModel routing explained
Everyone knows a smaller model would do for some of the traffic. The hard part is proving which part, before a customer finds out you guessed wrong.
12 min readWhy AI agents cost more than you modelled
A chat turn costs what a chat turn costs. An agent run costs whatever it decides to spend. Here are the four structural reasons, and the controls that make agent spend predictable.
10 min readAzure OpenAI cost optimisation
Caching and routing apply everywhere. These are the decisions that only exist because you are on Azure — and the ones teams most often inherit from a proof of concept and never revisit.
11 min readFAQ
Questions we get asked before the first call
What does an AI cost optimisation consultancy actually do?
Three things, in order. It measures — instrumenting your model calls and joining them to whether the customer's task succeeded, so cost can be attributed to features and outcomes rather than to models and days. It reduces — shipping the output-identical savings first, then proving the quality trade-offs against golden datasets before anything reaches a customer. And it defends — leaving budgets, evaluation gates and unit economics in place so the saving does not decay.
How much can we expect to save on our LLM costs?
It depends entirely on how the system was built, and anyone quoting a single percentage without seeing your traces is guessing. Systems that resend full conversation history, send un-cacheable prompts and run one flagship model for everything have a great deal of room. Systems that have already been optimised once have much less. The two-week audit exists to answer this with your numbers rather than an industry average.
Will cutting AI costs make our product worse?
Not for the first two thirds of the work. Prompt caching, batch scheduling and deduplication cannot change a single token of your output — they change only how the provider bills the work. Retrieval trimming and history strategy can change output, so they are validated against a fixed dataset. Only model routing involves a genuine trade, and even there the trade is per task class rather than across the board.
Do you need access to our production systems?
Less than most teams expect. The audit needs read access to billing and model logs plus a short engineering session to add tracing to the call path. We work inside your cloud account and your network boundary, no customer data needs to leave your environment, and de-identified samples are used for any replay work.
What size of company is this for?
Teams spending roughly £3,000 a month or more on model APIs. Below that the arithmetic rarely justifies an engagement and we will tell you so on the first call rather than sell you one — the guides on this site cover the same levers, and most of the early wins are things a competent engineering team can ship themselves.
Which providers and clouds do you work with?
Azure OpenAI and Azure AI Foundry are where we are deepest, because that is where our own estate runs. The measurement discipline and the gateway pattern are provider-agnostic and we regularly work across OpenAI, Anthropic and open-weight models served from your own infrastructure.
Find out what your AI actually costs per completed task.
Two weeks, a fixed fee, and a ranked savings plan with the quality risk of every move stated up front. If the numbers say an audit is not worth it for you, we will say so on the first call.