Most agent steps are routing, extraction and formatting. Test right-sized models for those against your client evals, and keep frontier models on the steps that need them.
Prefer email? hello@azitalabs.com, subject “Workload evaluation: AI agents for clients”
You build multi-step agents for clients. Most of the work is routing, extraction and formatting, yet every step calls a frontier model because that is what the prototype used. Each client has its own eval suite and its own residency clauses, and you cannot tune the endpoints you rent. Rate limits and model changes arrive on someone else's schedule, disrupting your client deployments.
An example workflow
It builds document agents for enterprise clients, and most steps are extraction and formatting. An evaluation could run one client's eval set on right-sized models and compare pass rates, latency and total cost with the current provider.
An example workflow
Its workflows route leads and tickets for several clients. An evaluation could profile one production workflow step by step and compare pass rates and total cost with the frontier calls it makes today.
An example workflow
Some client contracts restrict where an agent may process data. An evaluation could run one agent where the client's contract allows, on the Azita fleet or an installation, and compare the evals and the total cost.
Every step is judged against your client's eval suite. A step that does not pass stays where it is.
Routing, extraction and formatting steps are matched to models sized to them, and steps that need a frontier model keep it.
Every step's total cost compared with what you pay today, so the margin on each client is visible before you switch.
Run your workload on capacity from Azita’s installations. Scope the workload and data requirements.
Discuss your workload on the fleetScope the room and workload together. Azita designs, installs and commissions the deployment.
Discuss an installation in your buildingWhere the work runs, which models are used, who has access and how long anything is kept are agreed in your proposal before any work starts. The Trust page sets out custody and residency by side.
Agree a representative job and the acceptance criteria. Compare the output, turnaround, data requirements and total cost before deciding.
Available to evaluate on your own workload.
On the Azita fleet, or on an installation in your building.
We'll review what you need to run and contact you to discuss the right setup.
There is no price list. Your proposal states what is being quoted, the commitment and what is included. Human review, software licences and integration are listed wherever they apply.
Add what you pay for today. It travels with your enquiry as a note, so the comparison starts from the same work.
Prefer to read first? How pricing works
Your eval suite decides. We map each step to the smallest model that meets your pass criteria and run the suite on both stacks side by side. A step that does not pass stays where it is.
Send one client agent in production with its eval set and last month's usage. We agree the pass criteria, run the suite on both stacks, and set out pass rates, latency and total cost for each step.
Your evals run in CI, and you cut over step by step behind your own flags. Any step that does not pass on the new stack stays with your current provider.
On the Azita fleet, or on an installation in your own building. Where your client data is processed, who has access and how long anything is kept are agreed in your proposal before any work starts.
Proposal based on your workload and deployment. It states what is being quoted, the commitment and what is included, and lists human review, software licences and integration wherever they apply. There is no price list.
Term, what is included, exit and data handling at the end are set out in your agreement before you sign.
Related
Engineering and product
Agents, support, extraction and product features, tested on your own evals.
The page for every kind of work in this group, and how an evaluation works.
Read
Also for
Chatbots & support AI
Intent, retrieval and response steps on right-sized models, tested on your own support evals. Compare resolution quality, latency and total cost first.
Read
Also for
SaaS product
Routing and extraction steps behind your AI features on right-sized models, tested on your evals. Compare pass rates, latency and total cost first.
Read
A production workflow and its eval set. We agree the pass criteria, then compare output, latency and total cost on the same input.
Arrange a workload evaluation