Shopping agents scale with traffic. Make sure their cost scales with orders, not with retries.
Audon reads your OpenAI, Anthropic and xAI (Grok) cost data by project, workspace and model, separates seasonal demand from spend that doesn’t fit, and connects cost to successful sessions and jobs when you send outcome telemetry.
Free 30-day audit. Read access only, no card.
Observed spend
$20,604.56
30 of 30 days
Findings
3
1 anomaly to review
Largest project
47.7%
search-assistant
Verified savings
$1,467.04
1 approved experiment
- search-assistant47.7%$9,840.20
- shopping-assistant30.1%$6,214.75
- catalogue-enrichment16.7%$3,441.19
- review-summaries5.3%$1,108.42
Mapped as one project per shopper-facing agent or batch job.
Peaks, batches and assistants. Three shapes of spend, three ways to waste it.
Seasonal traffic and campaign peaks
A sale, a newsletter or Black Friday doubles traffic, and spend doubles with it, legitimately. A broken agent can hide inside the same spike.
Spikes are labelled anomalous spend, never waste, and your team notes the campaign on the finding. With outcome telemetry you compare cost per successful session in the peak against the weeks before it.
- spend_spike
- Evidence class on every amount
- Outcome telemetry
Catalogue and enrichment batch jobs
Re-enriching 400,000 SKUs is one job that can cost more than a week of shopper traffic, and a run that fails halfway costs the same as one that finishes.
Give batch jobs their own project. The spike detector catches a re-run the next day; with telemetry, the failed-run detector estimates the cost of runs whose failure was verified, labelled potential saving.
- spend_spike
- failed_run_spend
- Daily spend alerts
Assistant cost per order, not per token
A search or shopping assistant is worth what it costs per successful session. Tokens per call says nothing about that.
Send a run id, a status and how it was verified for each session; Audon divides attributable cost by verified successes. We don’t read order data; you decide what counts as a success.
- Outcome telemetry
- outcome_data_gap
Model swaps that only look cheaper
A smaller model cuts cost per call. If shoppers then rephrase three times, it can cost more per successful search.
Run the swap as an experiment on production traffic after a second person approves it. It counts as a verified saving only if cost per verified success falls without a success-rate drop over 2 points or a p95 latency rise over 20%.
- Verification workflow
- Breakdown by project, workspace and model
A month of a marketplace, audited. Demand, a failed batch and slow growth, kept apart.
- FND-0702High
Spend spike on project catalogue-enrichment
Anomalous spend$748.85 above median12–13 Oct at 5.6× and 4.9× the trailing 14-day median · 61 cost records
- FND-0703Medium
Spend on failed runs in workflow enrich-sku
Potential saving$612.35 estimated1,486 of 6,020 verified runs failed (24.7%) · 1.9 retries per run on average
Uses outcome telemetry (paid plans)
- FND-0704Medium
Sustained spend growth on project search-assistant
Observed spend$1,271.70 weekly increaseWeekly spend $1,640.00 → $1,990.25 → $2,402.90 → $2,911.70
FND-0702
Spend spike on project catalogue-enrichment
- Classification
- Anomalous spend
- Amount
- $748.85 above median
- Detector
- spend_spike 1.0.0
- Status
- Awaiting approval
Detected by scheduled audit14 Oct
Excess over median $748.85 on catalogue-enrichment.Note from the catalogue team14 Oct
Supplier feed re-sent 180,000 SKUs. Enrichment ran twice; the second run failed at 60%.Proposal: skip unchanged SKUs15 Oct
Hash product data before enrichment. Rollback: disable the skip flag.
Synthetic example: generated for this page, not a customer result. Anomalous and observed amounts are not savings, potential savings are not validated, and nothing becomes a verified saving without an approved experiment.
What changes for each role. Same data, three different questions answered.
E-commerce operations
analystToday
Plans peak capacity for traffic, but not for the agents behind search.
With Audon
- Spend per assistant and batch job, daily
- Peak spend noted, not mislabelled as waste
- Cost per successful session with telemetry
Engineering lead
adminToday
Finds a doubled enrichment run in the invoice, three weeks later.
With Audon
- Batch re-runs caught the next day
- Failed-run cost per workflow
- Model swaps proven by experiment
Finance
billing_adminToday
Sees AI spend jump in November and can’t tell demand from mistakes.
With Audon
- Observed spend reconciled to the provider
- Monthly alerts below budget
- Savings counted only once verified
Your first audit takes about fifteen minutes. Thirty days of real spend, read-only.
About 2 minutes
Connect a key (read access only)
Paste an OpenAI organization admin key, an Anthropic Admin API key or an xAI management key. We validate it against the cost and usage endpoints before saving it, encrypted.
Usually under 10 minutes
Import and reconcile 30 days
Daily cost records by project, workspace and model, reconciled against what is already stored. Recent days stay provisional until the provider finalises them.
Detectors run in seconds
Read the report
Six versioned detectors run on a hashed snapshot. You get an executive summary, coverage, spend profile, findings with evidence, next steps and the method.
Shopper data stays in your stack. We only need the bill and, if you choose, an outcome.
No shopper content
Connectors read cost and usage aggregates. Telemetry carries run ids and statuses, and rejects metadata keys that look like messages, emails, phone numbers or other personal data.
Nothing in the checkout path
Audon reads billing APIs out of band. It adds no latency and can’t take your search or checkout down.
Guarded experiments
Changes are approved by a second person and measured on production traffic, with success-rate and p95 latency guards before anything counts as saved.
Revocable access
The admin key is used read-only, encrypted with AES-256-GCM and never shown again. Revoke it at the provider and syncing stops.
Audon doesn’t make you compliant with consumer-protection or data-protection law; it keeps shopper content out of cost analysis by design. Nothing on this page is legal advice. Read the security overview and privacy policy for the details.
Questions e-commerce and marketplace teams ask before connecting.
Will Black Friday be flagged as an anomaly?
Do you need our order or product data?
Can it catch a failed batch job?
Can it block spend during a peak?
Which providers do you support?
What’s free?
Know what a successful session costs. Before the next peak.
Free 30-day audit. Read access only, no card, revoke any time.