Know what each customer’s agents cost you. Before it shows up in gross margin.
Audon reads your OpenAI, Anthropic and xAI (Grok) cost data by project, workspace and model, flags spend that doesn’t fit, and keeps every amount labelled for what it is. Map projects to tiers or features and see margin pressure while you can still act on it.
Free 30-day audit. Read access only, no card.
Observed spend
$24,808.58
30 of 30 days
Findings
4
1 anomaly to review
Largest project
69.7%
copilot-pro
Verified savings
$1,766.37
1 approved experiment
- copilot-pro69.7%$17,312.55
- free-tier16.8%$4,174.78
- doc-summarise9.3%$2,330.75
- eval-ci3.9%$990.50
Mapped as one project per product feature or tier.
Where SaaS margin leaks into the model bill. And what Audon does about each.
Gross margin per customer is a guess
Providers bill per organization, not per customer. When an enterprise account’s usage triples, you find out from the invoice a month later, blended with everyone else’s.
Give each tier or feature its own OpenAI project or Anthropic workspace and you get daily spend per project and model. The concentration detector flags one project carrying 60% or more of attributed spend; sustained growth flags three straight weeks above 20%.
- Breakdown by project, workspace and model
- spend_concentration
- sustained_growth
Free-tier usage you can’t see until month end
Trial and free accounts run the same agents as paying ones. A scraping script or a looping integration on a free plan burns the same tokens your revenue pays for.
Put the free tier in its own project. The spike detector flags any day above 2.5× its trailing 14-day median, and a daily spend alert on that project fires on the next complete UTC day.
- spend_spike
- Daily spend alerts
Model routing changes ship without a cost check
A router tweak or a new default model is a one-line change. Its cost shows up weeks later, mixed in with growth, and nobody can say whether it paid for itself.
The per-model breakdown shows the shift on the day it lands. With outcome telemetry you compare cost per verified successful task before and after, and a routing change only counts as a verified saving after an approved experiment on production traffic.
- Breakdown by project, workspace and model
- Outcome telemetry
- Verification workflow
AI cost of revenue that finance can defend
The board asks for AI COGS. The spreadsheet mixes invoices, forecasts and hoped-for savings into one number.
Observed spend reconciles to provider billing. Potential savings stay separate from verified ones in the data, the API and every export, so the number you report has one meaning.
- Evidence class on every amount
- Report and data export
- Detector versions and snapshot hash
What a first audit tends to surface. Four findings, four different kinds of money.
- FND-0311High
Spend spike on project free-tier
Anomalous spend$556.48 above median12–13 Oct at 3.6× and 3.1× the trailing 14-day median · 96 cost records
- FND-0312Info
69.7% of attributed spend is on project copilot-pro
Observed spend$17,312.55 project totalShare of attributed spend over 30 days · 4 projects with attribution
- FND-0313Medium
Sustained spend growth on model gpt-4.1
Observed spend$2,392.44 weekly increaseWeekly spend $3,120.40 → $3,790.12 → $4,561.30 → $5,512.84
- FND-0314Info
No outcome data: cost per successful task is unknown
Observed spendUnavailable cost per success$24,808.58 of spend with no matching outcome telemetry
FND-0311
Spend spike on project free-tier
- Classification
- Anomalous spend
- Amount
- $556.48 above median
- Detector
- spend_spike 1.0.0
- Status
- Open
Detected by audit26 Oct
Excess over median $556.48 on free-tier.Note from the platform lead26 Oct
One free account called the summarise endpoint 41,000 times in two days.Daily spend alert created26 Oct
free-tier above $250.00 per day, sent to the on-call webhook.
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.
Finance
billing_adminToday
Reconciles one provider invoice against a revenue model by hand, once a month.
With Audon
- AI spend per tier and feature, daily
- Observed spend kept apart from projected savings
- CSV or JSON export for the COGS model
Engineering lead
adminToday
Learns about a routing regression when finance asks why the bill moved.
With Audon
- Per-model shifts the day after they land
- Spike and growth findings with their evidence
- Experiments that prove a change saved money
Product
analystToday
Prices tiers on token estimates from a launch spreadsheet.
With Audon
- Real cost per feature to price against
- Free-tier usage watched by alerts
- Cost per verified successful task with telemetry
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.
Controls that fit how SaaS teams ship. Without putting anything in the request path.
Read access, out of band
An admin key, used read-only, reads cost and usage reports. Nothing proxies your traffic, so there is no latency cost and no new failure mode in production.
Agents can run it, people approve
Your own agents can start audits and create alerts over MCP with scoped OAuth. Approving a change that moves money still needs a signed-in person.
No customer content
Connectors never read prompts or completions. Telemetry rejects metadata keys that look like prompts, content or personal data.
Evidence you can hand to finance
Every finding carries its detector version, baseline and assumptions, and every audit records a hash of the data it analysed.
Audon supports your internal controls; it doesn’t certify anything. We hold no SOC 2 or ISO 27001 certification. Read the security overview and privacy policy for the details.
Questions AI-native SaaS teams ask before connecting.
Can it show cost per customer?
Do you sit in our inference path?
How do we measure whether a model change saved money?
Will it flag a big customer’s legitimate growth as waste?
Can our agents query it?
What does the free audit include?
Price your tiers on what they really cost. Start with 30 days of your own spend.
Free 30-day audit. Read access only, no card, revoke any time.