A question your board may already have put to you.
What did AI cost last quarter, broken down by product? A CFO can tell you the contribution margin of a product line and the payback period on a piece of capital equipment. Ask what a single AI feature costs to serve, and whether it is margin-positive. The answer you get is an estimate, hedged, and slow. That is not negligence. Nothing in the stack was built to do it.
The problem today
The spend is scattered
Five or more bills (Anthropic, OpenAI, Bedrock, Azure, Google) each counting usage a different way.
Nobody owns it
One shared login hides every team behind a single invoice line.
You find out too late
Overspend shows up on the invoice, not while it is happening.
Value-chain attribution
Five steps from consumption to a decision.
1
Understand AI Consumption
AI Workload Telemetry
See what was consumed, where, by which workload and for what purpose.
- Tokens in and out
- Models and providers
- Tools and actions
- Retrievals and data
- Workflows and steps
2
Connect spend to ownership
OWNERSHIP AND ATTRIBUTION
Attribute AI activity to the application, team, customer, cost centre and outcome it supports.
- Business owner
- Use case or process
- Customer or segment
- Team or application
- Cost centre or project
3
Build the true cost picture
PLATFORM VERSUS WORKLOAD COST
Combine direct AI consumption with the shared platform cost required to deliver it.
- Direct workload cost: tokens, tools, compute
- Shared platform cost: infrastructure, models, services
- Allocation rules agreed up front
4
Know the cost per outcome
FULLY LOADED COST PER OUTCOME
Calculate the fully loaded cost of delivering a transaction, claim, interaction or business result.
- + Shared allocation
- + Human verification
- + Exception and recovery
- = Fully loaded cost per successful outcome
5
Answer the business
BUSINESS IMPACT
Understand margin, cost-to-serve, profitability and unit economics, and know where to act.
- Product margin
- Cost-to-serve
- Project profitability
- Customer profitability
- Pricing and unit economics
Working Example
An insurance claims agent.
Direct workload cost
tokens, tools, compute
Shared platform allocation
infrastructure, models, services
Human verification
review, validation, approval
Recovery cost
exceptions, retries, rework
= Fully loaded cost per settled claim
The goal is decision-grade transparency, not forensic accounting perfection.
The same invoice, two ways
Today
Today: one shared login, no owner
$48,210
total AI spend for the month. 100% unattributed. Two logins cover five capabilities and sixteen applications.
Questions you cannot answer
- Which team caused the increase?
- What does one enriched job cost?
- Is this live traffic or testing?
- Which spend is safe to cut?
With Arco
With Arco: the same $48,210, now addressable
94.5%
has an owner
$0.38
per enriched job, 6c cheaper than last month
$52,900
month-end forecast, 5.8% over budget
$9,400
savings found per month, four actions
Figures are illustrative, taken from the sales flyer. How much detail you capture decides what the reports can ever show.
Know the owner
Every dollar of AI consumption has business context.
Know the true cost
Direct consumption, shared platforms and operational effort are brought together.
Know where to act
Know where to act.