Analytics And Costs
Understand the administrative surfaces that summarize usage and cost behavior in MARCUS.
MARCUS includes organization-level analytics and cost surfaces for teams that need operational oversight. These pages are not primarily for every end user. They are for the people responsible for adoption, sustainability, governance, and platform health.
If you are a program lead, administrator, product owner, or technical reviewer, these surfaces help you answer: "Is the system being used, is it being used well, and what is it costing us?"
What These Pages Are For
- review broad usage patterns
- inspect cost summaries and event streams
- support budgeting and model-governance conversations
- understand whether a product change altered user behavior
- relate source growth, usage growth, and spend growth
Who Should Use Them
These surfaces are generally intended for:
- administrators
- operations leads
- program owners
- technical reviewers
- developers watching benchmark or health behavior
Most resident or clinician users do not need to live in these pages daily.
What You Will Typically See
| Surface | Purpose |
|---|---|
| Analytics | Organization or workflow-level visibility into usage patterns |
| Costs | Administrative cost summaries, time-series views, and raw events |
| Developer benchmarks | Engineering visibility for benchmark runs and health-oriented metrics |
These surfaces tell different parts of the same story. Costs without usage context are easy to misread. Usage without quality context is also easy to misread.
How To Read Analytics Well
Analytics helps you understand activity patterns such as:
- which product areas are being used
- whether adoption is rising or falling
- whether usage is concentrated in a few teams or spread broadly
- whether a new feature changed behavior
Questions analytics can answer well:
- Are users actually asking questions after uploads?
- Is one program adopting MARCUS while another is stalled?
- Did a rollout or training effort change engagement?
Questions analytics alone cannot answer:
- Are the answers clinically good?
- Is the corpus high quality?
- Are users satisfied?
For that, you need to combine analytics with corpus review, answer review, and benchmark signals.
How To Read Cost Data Well
Cost surfaces are most useful when you interpret movement in context.
Cost increases can happen because:
- more users are active
- more documents are being ingested
- more retrieval or generation volume is occurring
- configuration changed
- a new feature increased usage intensity
A cost increase is not automatically a problem. It may reflect healthy adoption.
A Better Way To Think About Cost Changes
When cost changes, do not ask only "why is spend up?"
Also ask:
- Did source volume increase?
- Did user volume increase?
- Did question volume increase?
- Did a more complex workflow become available?
- Did retrieval or enrichment depth change?
This turns cost review from a blame exercise into a system-behavior review.
A Good Review Pattern For Admins
When you review these pages, look at four things together:
- usage volume
- source growth
- cost trend
- benchmark or quality signals
Reviewing these together gives you a more honest picture than any one panel by itself.
Common Misreadings
| Misreading | Better interpretation |
|---|---|
| "Costs went up, so the system got worse." | Cost may have risen because the system is being used more or processing more evidence. |
| "Usage is high, so quality must be good." | Adoption does not prove answer quality or corpus quality. |
| "Benchmark changes do not matter because users are still active." | Users can remain active even while quality drifts; health signals still matter. |
| "One expensive day means we need a drastic policy change." | Look for sustained patterns, not isolated spikes. |
When To Investigate More Deeply
Investigate when you see:
- sustained cost growth without matching usage growth
- large drops in activity after a release
- unusual shifts in which features are being used
- benchmark or health regressions that coincide with behavior changes
- organizational complaints that are not visible in the raw activity numbers
One Useful Administrative Habit
Review analytics, costs, and corpus health together on a recurring cadence. MARCUS works best when governance is based on evidence from product usage, source quality, and answer quality together rather than from any one dashboard alone.