MEASUREMENT GUIDE
Measure AI token savings without guessing.
Shared memory can reduce repeated context and avoid some model calls, but the result depends on the workflow. This guide shows how to build a fair baseline, run the same tasks with memory enabled, and report both savings and quality failures.
Token savings equal the baseline input and output tokens for completed tasks minus the tokens used for those same tasks with memory enabled. Report avoided calls, smaller remaining contexts, memory quality, and operating overhead separately.
Start with a stable baseline
Choose a repeatable workflow and a fixed set of representative tasks. Record successful completion, input tokens, output tokens, retries, latency, and model calls before changing the request path.
- Use the same task set in both tests.
- Keep model and quality criteria stable.
- Count retries and failed tasks.
- Measure per completed task, not per request only.
Separate the two sources of savings
A valid memory hit may avoid an entire model call. A memory miss may still use fewer input tokens because Governora builds a smaller context. Report these effects separately so the result can be reproduced.
Illustrative calculation
Imagine 1,000 completed support tasks use 2,000 input tokens and 300 output tokens each: 2.3 million tokens in the baseline. If 300 tasks reuse an approved answer without a model call and the remaining 700 tasks use 1,200 input plus 300 output tokens, model use becomes 1.05 million tokens. The illustrative reduction is 1.25 million tokens, or about 54%. This is arithmetic for explaining the method, not a measured Governora customer result.
Do not trade quality for a lower bill
A reused answer counts only when it passes the same correctness, approval, scope, and freshness checks as a newly generated answer. Report rejected memory hits, stale answers, human reviews, and additional storage or retrieval cost next to the token result.
- Answer acceptance rate.
- False or unsafe memory-hit rate.
- Human-review rate.
- Latency and non-model infrastructure cost.
HOW IT WORKS
Four clear steps.
Each step stays connected to the same execution. This makes authority, external work and the final outcome explainable.
- 01
Baseline
Run the fixed task set without shared memory and record complete cost and quality.
- 02
Enable
Run the same tasks with memory-first routing and lean context.
- 03
Compare
Separate avoided calls, reduced context, failures, and operating overhead.
- 04
Repeat
Repeat over time as memory, source data, prompts, and models change.
QUESTIONS
Simple questions. Direct answers.
What Governora does, what it does not do, and where it fits.
Does Governora promise a fixed token-saving percentage?
No. Savings depend on repetition, memory-hit quality, context size, model behavior, and workflow design. Measure one workflow against its own baseline.
Should cached-input discounts count as token savings?
Report them separately. Provider caching may lower price even when token volume stays the same; shared memory may reduce or avoid the request itself.
What is the most useful headline metric?
Use model tokens per successfully completed task, supported by memory-hit rate, avoided calls, answer quality, latency, and total operating cost.
Map one governed AI execution.
We will map the actor, decisions, capabilities, evidence, external calls and proof with you.
Plan a pilot review →