Spend attributed to workflows
A provider invoice is one number. We break it down by workflow, feature and user group, so you can see what each thing you built actually costs to run per month and per task.
Efficiency
Your AI spend is growing and nobody can say what it bought
Inference costs tend to grow faster than usage, and the bill rarely says why. Most of the systems we look at are running a frontier model on work a much cheaper one handles identically, sending the same context hundreds of times a day, and paying for three tools that overlap.
From$4,000
Fixed-fee, read-only: spend attributed per workflow and a ranked reduction list with measured quality risk.
Indicative starting price. The fixed fee for your scope is quoted after the two-day diagnosis, before any build begins.
What makes it work
Three things we insist on
Cheaper only where quality holds
Every proposed reduction is tested against an evaluation set built from your real cases. If moving a workflow to a smaller model costs accuracy, we say so and leave it. Cost work that quietly degrades output is not a saving.
The cheapest token is one you never send
Most of the reduction is structural rather than a model swap - caching near-duplicate requests, trimming context that was never read, routing easy cases away from the expensive path, and deleting work nothing consumes.
Attribute the spend first By workflow, model and step. Optimising the part you assumed was expensive is effort wasted
Find which of six causes it is Retrieved context grew, retries are invisible, an agent loop has no ceiling, everything runs on the expensive model, nothing is cached, or a batch job is billing at production rates
Measure quality before changing anything The baseline that makes a cost cut checkable rather than a hope
Change one thing, re-run the set Cost down and quality held, or the change goes back
Capabilities
What is actually included
- 01
Spend attribution
Provider and infrastructure cost broken down per workflow, with cost per task and month-on-month trend.
- 02
Model right-sizing
Candidate downgrades tested against your evaluation set, reported with the measured quality delta.
- 03
Caching and routing design
Identifying near-duplicate traffic and easy cases that never needed the expensive path.
- 04
Context and prompt efficiency
Finding the context being sent on every call that no output ever depends on.
- 05
Tool and licence overlap
Mapping what you pay for against what is used, including the seats nobody has opened this quarter.
In detail
What this covers, specifically
A category name is not a scope. These are the individual pieces of work inside this practice - take the two that apply to you and ignore the rest.
- 01
Spend attribution
Provider and infrastructure cost broken down per workflow, feature and user group, with cost per task and a month-on-month trend.
- 02
Model right-sizing
Testing whether a cheaper or smaller model handles a given workflow identically, and reporting the measured quality delta rather than assuming.
- 03
Context efficiency
Finding the context sent on every single call that no output has ever depended on, which is usually the single largest avoidable cost.
- 04
Caching strategy
Identifying near-duplicate requests, because in most production systems a meaningful share of traffic is a question already answered.
- 05
Request routing
Cascading easy cases to a cheap model and escalating only what needs the expensive path, with the escalation rule measured.
- 06
Batch versus realtime
Finding the workloads paying a realtime premium for output nobody reads until the next morning.
- 07
Licence and tool overlap
Mapping what you pay for against what is used, including the seats nobody has opened this quarter and the three tools doing one job.
- 08
Infrastructure review
GPU utilisation, instance sizing and idle capacity for anything self-hosted.
- 09
Budget alerting
Forecasting and thresholds so a runaway loop or a pricing change is caught in hours, not on the invoice.
What you receive
Concrete artefacts, not a slide deck
- Cost breakdown per workflow, feature and user group
- Ranked reduction list with measured quality risk against each
- Evaluation results for every proposed model change
- Caching, routing and context recommendations
- Implementation plan, if you want us to make the changes
01 Diagnose Days 1-2
02 Prove Week 1
03 Integrate Weeks 2-3
04 Operate Ongoing
Around three weeks end to end. That comes from scoping tightly to one workflow - not from skipping a phase. Each still ends in evidence you can check.
Further reading
Written up in more detail
- Measuring cost per resolved task Total AI spend tells you nothing. The unit metric that does, how to compute it including the human cost of escalation, and why it beats deflection rate.
- Cutting LLM costs without degrading quality Eight cost reductions ranked by quality risk, from the ones that are free to the ones that are a real trade - and how to prove which is which before shipping.
Let's talk
Fixed fee, read-only, one week
We need billing exports and read access to the systems. We change nothing. You get a written breakdown of where the money goes and a ranked list of reductions with the quality risk stated against each - yours to implement with us or without us.
Related practices