Voice AI · Underwriting · Claims · Patient care
Would another model have earned you more?
Ganivra takes the calls you already made — same tokens, same workflows — and prices them on GPT, Claude and Gemini models. See what each step would have cost, which models fit your request sizes and capabilities, and what to test before you switch.
How the comparison works
From what you spent to what you could test next.
Your spend, repriced
Compare against GPT, Claude and Gemini
For each workflow, see your actual spend next to the estimated cost of the same calls on other models, and the total for the view if only those calls changed.
Fit, not just price
Rule out models that cannot do the job
Your largest recorded requests set the minimum context and output size. Add requirements such as structured JSON, tool calling, image or audio input, and a latency target to narrow the list.
Know what to measure
Get the quality check for the task
Choose what the workflow does — routing, extraction, summarization, support, reasoning — and Ganivra names the quality check that matters before a trial.
Connect to margin
See savings against what the customer pays
Add customer pricing or execution revenue to see whether a cheaper step turns a loss-making call, claim or patient into a profitable one.
Voice AI agents
Every turn of the call is a model call. Price them per minute.
A six-minute call can make dozens of model calls: understanding the caller, routing, looking up the account, answering and summarizing. Most of those turns are short classification steps. Running them on the same model as the answer step sets your cost per minute — and your margin on per-minute pricing.
| Workflow step | Main task in Ganivra | Requirements to keep | What the comparison tells you |
|---|---|---|---|
| Detect intent / route | Classify or route requests | Tool calling, latency target | Often the largest share of calls. Strong candidate for a smaller model. |
| Answer the caller | Answer customer support questions | Tool calling, latency target, audio output if native | Test carefully: answer quality and turn latency both matter. |
| Post-call summary & CRM notes | Summarize reports or documents | Structured JSON output | Not live, so latency rarely matters. Good candidate to move. |
- Tracked AI cost today
- $0.42
- With lower-cost steps
- $0.21
At $0.12 per minute, margin moves from 42% to 71%.
$2.5M a year left on the table by staying on one model for every turn.
Check before you switch
- Intent accuracy and missed handoffs on labeled call samples
- Turn latency at p50 and p95 against your target
- Correct escalation to a human agent
- Summary field accuracy in the CRM
Send speech-to-text and text-to-speech charges as tool costs. Model comparison reprices token-based calls; per-minute speech charges stay as recorded.
Insurance underwriting
Not every page of a submission needs your most expensive model.
A commercial submission brings applications, loss runs, schedules and broker emails. Extracting fields from them is high-volume and repetitive; the risk recommendation is not. Comparing each step separately shows where a lower-cost model is worth testing and where the large context of loss runs rules a model out.
| Workflow step | Main task in Ganivra | Requirements to keep | What the comparison tells you |
|---|---|---|---|
| Extract application & schedule fields | Extract fields from documents | Structured JSON output, image input for scans | High volume and well defined. Strong candidate for a smaller model. |
| Summarize loss runs | Summarize reports or documents | Required context tokens from your largest submissions | Models that cannot fit your recorded request sizes are excluded automatically. |
| Risk recommendation for the underwriter | Analyze and solve complex problems | Your current model until a referral review says otherwise | Lower volume, highest stakes. Usually the last step to change. |
- Tracked AI cost today
- $1.60
- With lower-cost steps
- $0.60
At $4.00 per submission processed, margin moves from 60% to 85%.
$3M a year — before counting the submissions you can now afford to quote.
Check before you switch
- Field accuracy and missing values against underwriter-corrected submissions
- JSON schema validity for your policy admin system
- Referral and decline decisions compared with underwriter outcomes
- No unsupported statements in loss summaries
Ganivra compares cost and documented capability. Underwriting decisions still need your own model governance and fairness review before any change.
Insurance claims
Know what each claim costs to handle, and which steps can cost less.
First notice of loss, damage photos, invoices, coverage checks and adjuster notes each trigger their own model calls. Claims vendors and TPAs often charge a fixed fee per claim, so a costly model on a high-volume step quietly erodes the margin on every file.
| Workflow step | Main task in Ganivra | Requirements to keep | What the comparison tells you |
|---|---|---|---|
| FNOL intake and triage | Classify or route requests | Structured JSON output | High volume, clear labels. Strong candidate for a smaller model. |
| Invoice and estimate extraction | Extract fields from documents | Image input, structured JSON output | Filter to models with documented image input, then compare field accuracy. |
| Coverage reasoning and review flags | Analyze and solve complex problems | Your current model until leakage and false flags are measured | Errors here cost far more than tokens. Change only with evidence. |
- Tracked AI cost today
- $2.40
- With lower-cost steps
- $1.10
At a $6.00 per-claim fee, margin moves from 60% to 82%.
$7.8M a year across a book of that size.
Check before you switch
- Field accuracy on invoices, estimates and policy numbers
- Coverage citations checked against the policy
- False review flags and missed flags on closed claims
- Cycle time compared with your current model
Keep claim documents and photos out of telemetry. Send claim IDs as opaque execution IDs with token counts and charges.
Health & patient care
Price care workflows per patient without sending PHI.
Patient messaging, visit summaries, prior authorization packets and discharge instructions run every day for every patient. Per-member or per-patient pricing is fixed, so model choice decides whether high-need patients are served at a loss. Ganivra compares cost using model names and token counts only.
| Workflow step | Main task in Ganivra | Requirements to keep | What the comparison tells you |
|---|---|---|---|
| Draft visit summaries and patient letters | Summarize reports or documents | Required context tokens, structured JSON output | High volume with clinician review. Candidate to test on a smaller model. |
| Prior authorization packets | Extract fields from documents | Image input for faxed records, structured JSON output | Compare field accuracy and payer rejections, not just cost. |
| Symptom triage and urgent routing | Classify or route requests | Your current model, unless clinical validation supports a change | Safety-critical. Treat savings as a reason to evaluate, never as the decision. |
- Tracked AI cost today
- $1.35
- With lower-cost steps
- $0.55
At $3.00 per patient per month, margin moves from 55% to 82%.
$9.6M a year on a fixed per-patient contract.
Check before you switch
- Clinician-scored accuracy and unsafe omissions on sample summaries
- Urgent-case recall for any triage step
- Prior authorization field accuracy and payer rejections
- Reading level and instruction following in patient-facing text
Ganivra does not need prompts, transcripts or patient records. Use pseudonymous patient or program IDs. Clinical use still needs your own validation and approvals.
Margin calculator · at production volume
What would the switch be worth?
Start from an industry example, then replace each number with your own. In Ganivra, the current and candidate costs come from your recorded calls in Model comparison.
- Margin today
- 42%
- Margin with candidate
- 71%
$210,000 more gross margin every month
$2.5M a year. Staying put costs you that much — but switch only if the candidate passes the quality checks for this workflow.
Illustrative estimate. It covers tracked AI cost only, not staff time, telephony, data subscriptions or infrastructure.Set it up
Label the step, and the comparison follows.
execution_idUse an opaque ID so every step of one call, submission, claim or care episode stays together.
workflowModel comparison is grouped by workflow label. Give each step you want to compare on its own a label such as voice_intent_routing or loss_run_summary.
customer_id + revenueAttach a pseudonymous customer and your per-minute, per-claim or per-patient price to turn savings into margin.
What the estimate does and does not tell you
Estimates apply each model's list price to your recorded token counts. They do not measure output quality, and a different model may use a different number of tokens for the same task. Negotiated discounts, batch pricing and cached-token rates can change the result. Only calls with a recorded model and token counts can be compared; other costs in the view stay as recorded.
Start with one workflow
Find the step that costs more than it needs to.
Send a week of usage from one voice, underwriting, claims or care workflow. Open Model comparison and see what the same calls would have cost elsewhere — then multiply by the millions you run each month.