LegalTech & Compliance
Know what each document workflow costs.
Long contracts, repeated instructions, and multiple extraction passes can make a fixed-price document expensive to process. Compare model and prompt-version costs, then check whether reported cache usage is helping.
Where margins shrink
The same context can be paid for repeatedly.
Sending dense document context to an expensive model on every pass can increase input cost. Longer outputs and retries add more usage. A customer paying per document may generate very different margins depending on document length and the number of steps.
Attribute
Connect every pass to one document job
Use one execution ID for the processing job and a unique step ID for each model call. Add customer, workflow, feature, prompt ID, and prompt version to connect spend to the work that generated it.
Use opaque job IDs, not document titles or client names. Keep document contents in your own systems.
Compare
Benchmark prompt-version economics
Send prompt IDs and version labels without prompt text. Filter the dashboard by version and inspect input tokens, output tokens, cost, and latency for comparable document workflows.
Compare similar document lengths and task types. A cheaper average can reflect easier documents rather than a better prompt.
Validate
Check reported cache usage
Send the provider's actual cached-token counts and any provider-specific usage facts needed for pricing. Inspect the execution cost breakdown to see how cached usage was rated.
Ganivra does not inspect document prefixes or enable provider caching. Reported cached tokens show cache use; they do not prove a particular prefix was reused. Cache-write and read rules depend on the provider.
Decide
Reduce cost while checking quality
Test shorter repeated instructions, eligible prefix caching, or a lower-cost model for simpler steps. Compare cost per document with your own extraction accuracy, missed-clause, and human-review results.
Keep quality evaluation in your application or evaluation system. Ganivra measures the economics of the supplied usage; it does not certify legal accuracy or guarantee unchanged quality.
Illustrative scenario · not a benchmark
What a 60% cost reduction would mean.
| Per document | Baseline workflow | Tested alternative |
|---|---|---|
| Customer revenue | $1.00 | $1.00 |
| Tracked AI cost | $0.50 | $0.20 |
| Contribution after AI cost | $0.50 | $0.80 |
| Contribution margin | 50% | 80% |
These invented figures illustrate the calculation: a $0.30 reduction from a $0.50 baseline is 60%. They are not observed customer results or a promise of 50–70% savings. Validate the alternative on representative documents and your quality criteria before rolling it out.
A practical measurement plan
Keep the comparison explainable.
workflow: contract_extractionKeep the task definition consistent across versions. Use a neutral feature and customer ID for attribution.
prompt_id + prompt_versionLabel the baseline and candidate. Send the actual model, token counts, and latency for each call.
cached_tokensUse the provider-reported value. Do not estimate it from how much repeated text you intended to send.
Turn document cost into customer economics
Connect your customer pricing plan or confirmed revenue to usage. See which tenants and document workflows have enough margin to cover tracked AI costs, then review costs outside the model bill separately.
Start with one document workflow
Measure the baseline before optimizing.
Capture a real execution, add version labels, and compare the next change using cost and quality together.