Skip to main content
Next Step PM
Back to the library

Design · Add-on

Manage AI cost and speed, and decide when to fine-tune

Every AI call costs money and takes time. Estimate both early, pick the cheapest model that meets the quality bar, and use simple savings before anything complex.

About 2-5 daysFast-movingLast reviewed 5 November 2026

Who this applies to

Skills it builds
Technical and AI fluencyCommercial and business sense
PM types
AI
Product types
All
Starting points
All
Ways of working
All
Experience
All
Size of organisation
All

Steps, by experience

Beginner (under 1 year)

  1. Learn that cost depends on how much text goes in and comes out.
  2. Estimate calls per user per day and the cost per call.
  3. Multiply out to a monthly cost at your expected usage.
  4. Agree how fast a response must be for the feature to feel usable.
  5. Set a spending cap before launch.

Intermediate (1 to 4 years)

  1. Compare models on quality, speed and cost with your own test set.
  2. Use savings such as caching repeated instructions and batching non-urgent work.
  3. Decide whether fine-tuning is worth it only after simpler options.

Experienced (5+ years)

  1. Set cost and speed targets for AI features across products.
  2. Build unit economics into pricing and business cases.
  3. Review model choices regularly, because prices and models change quickly.

Who can do this

Owner: Product manager

Can be delegated to

  • Data Scientist / ML Engineer or Backend Engineer: Compare models on quality and speed
  • Business Analyst or Data Analyst: Model monthly cost

Reviewed by: Tech Lead

Tool examples

AWS
Set billing alerts for AI usage.
Google Cloud
Set a budget alert on the AI project.
Claude
Check current model pricing before you estimate.

Definition of done

  • Monthly cost estimate agreed
  • Speed target agreed
  • Spending cap set
  • Model chosen against a test set

Common pitfalls

  • Using the largest model by default
  • No spending cap
  • Fine-tuning before trying simpler options

If you lead the team

What good looks like: AI features have known cost and response times, with the cheapest model that meets the quality bar.

Ask your PM

  • What does each AI call cost us, and how will that grow?
  • Is this the cheapest model that meets our quality bar?

Warning signs

  • AI costs are discovered on the first big bill.
  • The largest model is used by default.

Do these first

Sources