Guide · AI workload procurement
AI workload procurement: choosing models and providers on evidence
Most teams pick an AI model once, from a benchmark or a price sheet, and then pay for it on every request. Procurement treats that choice as a purchase: a defined unit of work, a price per successful result, a record of why one supplier was chosen, and a date to look again.
Buy per workload, not per model
A model that is the right buy for summarising support tickets can be the wrong buy for extracting fields from invoices. The unit of procurement is one workload: one kind of job, with its own success check, volume and constraints on latency, data handling and cost.
What a procurement decision needs
- A success check fixed in advance. The test your system already applies to an answer, written down before results are compared.
- Cost per successful result. The cost of every attempt, failures and retries included, divided by the successes. See why the cheapest model per request can cost more per result.
- Hard constraints kept separate from price. Latency, data-handling terms and security review are pass-or-fail conditions, not points traded against cost.
- A dated price table. List prices change; a decision is only reproducible with the prices it used.
- Evidence someone else can check. The inputs, rules and outputs behind the decision, each with a fingerprint. See what evidence to keep.
Where usage data comes in
The best evidence about a workload is the usage you already have: provider exports or gateway logs with tokens, reported cost, outcome and latency per request. Comparing models on that data, rather than on a public benchmark, measures them on your own work. The provider comparison checklist sets out the method step by step.
What AIYVO does today, and what is planned
Today, AIYVO is an offline analyzer. It reads a usage export you already have and reports, per provider and model, cost per successful request (or per completed task, when each row is a task), success rate, latency and coverage, with the numerator, denominator and source behind every figure. It withholds a comparison when the sample is too small or costs are missing, and it refuses files with prompt, response or personal-data columns. It does not call providers and does not verify provider billing.
Choosing a supplier per workload under hard constraints, routing work between providers and spend governance are planned, not available. The planned-platform simulation shows how they are intended to work, on synthetic data.