How do you offer “Unlimited AI” in B2B SaaS without exploding AI costs?

I recently gave myself a UX/System Design challenge:

Imagine a multi-tenant B2B HR SaaS platform that wants to introduce conversational AI

The initial problem was simple:

How can we offer conversational AI without restricting user usage while keeping AI costs under control?

While exploring the problem, I started challenging my own assumptions

1. What if I can't rely on credits?
2. What if I don't want visible usage limits?
3. What if every interaction needs a fresh AI-generated response?
4. What if caching isn't part of the solution?

Those questions made the problem much more interesting.

Instead of asking:

“How do I limit AI usage?”

I started asking:

“How do I reduce the cost of fulfilling each interaction?”

That led me to explore several areas:

• Input and context optimization
• Output/token optimization
• Model selection and routing
• Processing heavy tasks asynchronously
• Tenant-level AI cost monitoring
• Detecting abnormal usage
• AI cost vs. SaaS subscription economics

The key UX principle I took away:

Don't restrict the user's freedom. Optimize the cost of fulfilling the usage.

What started as a simple AI feature quickly became a broader problem involving:

UX × Product × System Design × Business Economics

I'm documenting the thinking process rather than presenting this as a “perfect architecture.”

If you were solving this whiteboard challenge yourself, what would you question first?