Most AI applications make one model decision for the entire product.
A support ticket classifier, an invoice extractor, a refund decision, and a security response may all be sent through the same powerful model.
That feels safe, but it creates two problems:
The real question isn't:
Which model is cheapest?
It's:
Which model is appropriate for this specific workflow, given its complexity, risk, and observed behaviour?
I built ** MargIQ** to help answer that question using evidence from actual application traffic.
MargIQ identifies recurring AI workflows and evaluates them against the models already available in your application.
For each workflow, it can show:
The customer-facing unit is the workflow, rather than an individual prompt or a global model setting.
Reducing model cost is only useful if the application remains reliable.
When MargIQ does not have sufficient evidence, it keeps the requested model.
It also protects workflows where multiple outputs may all be defensible because the application has not clearly defined an important taxonomy, priority rule, or expected response structure.
Instead of repeatedly testing models or silently choosing a cheaper option, MargIQ explains what needs clarification before making a recommendation.
MargIQ is designed for server-side AI applications using compatible model-provider clients.
Getting started is as simple as:
npm install margiq
You keep your existing provider credentials and model configuration.
MargIQ works with the models your application already uses rather than requiring a specific provider.
The free plan starts in Report-only mode. It observes recurring workflows and reports potential savings without changing production routing.
When you're ready, workflow controls let you choose how optimization is applied:
MargIQ is now live, and I'd love feedback from founders and engineers running recurring AI workflows in production.
In particular, I'm interested in:
You can check it out here:
I built MargIQ because I believe AI cost optimization should be based on workflow evidence and business risk, not a blanket instruction to "use a smaller model."
I'd genuinely appreciate any feedback or suggestions from people building production AI systems.