The hours-saved survey
Ask people how much time the copilot saved them, multiply by an hourly rate.
Why is the return on AI so hard to measure?
Every company can say what it spends on AI to the dollar. Almost no one can say what they got out of it.
Spend is recorded by provider, revenue by customer, and nothing connects the two.
what the invoices say 30d
- Anthropic$5.2k
- OpenAI$3.9k
- Google$2.6k
- coding seats$1.7k
$13.4k by provider, by month. that is all an invoice knows.
?
which work · on which capability · for which customer
no ticket on the trace, no tag on the resource, no chain to the customer
what the customers say 30d
- Enterprise accounts$422k
- Self-serve teams$383k
$805k revenue by segment. no line on it for AI.
How companies measure it today
The vendor dashboard
Tokens, seats, acceptance rates, spend by model, spend by team.
The pilot business case
One project, a projected saving, a sign-off, a budget.
The cost-centre spread
Finance allocates the AI bill across departments by headcount.
The cost-per-successful-task scorecard
The full cost of the work, including human review, retries, and rework, divided by the tasks that met a quality bar, followed for one workflow over time. OpenAI's scorecard for the AI age.
The missing link is the value chain.
A customer pays for a need to be met. Meeting it takes a capability, and that capability depends on others: onboarding a business unit runs on the onboarding capability, which needs the core API, which needs search and the data platform.
Most businesses today can only really track revenue per headcount (people cost), total platform cloud cost, and AI spend. With Metacosm, you can map the value chain, and tag every resource (people, agents and tech) to a capability, and get a real picture of what each customer costs you vs. what they pay.
Understand your ROI on every customer
An example software company: one customer, the eleven capabilities that serve them, and a month with agents. Each card shows a capability's people, cloud, and AI cost, and the panel compares the customer's margin with the month before.
After the $11.1k of AI spent on this chain, margin is up $20.2k: a 182% return. That is $15.0k of extra revenue from faster onboarding and $17.5k of team time freed on support, reporting, identity, and search, less $1.2k of extra cloud and the AI itself. Across the business, $2.3k was spent without a known value. Those agent runs carried no ticket key, so they aren't attributable to a capability or a chain, and nobody can say what they were for.
Keeping the numbers honest
Share the cost.
The example has one customer, so it carries the whole cost of every capability on its chain. With more customers, a shared capability's cost has to be split: equally, by volume, or by revenue if you have the data. Whichever you pick, log the decision, and apply it the same way every month, so the change from month to month is real even though the split itself is a judgement call.
Freed capacity only counts if it goes somewhere.
When an agent adopts part of a capability's work, the team's allocation to that capability drops. That is a saving only if the people moved to other work. The example assumes they did. In practice, check before you count it as a saving.
Spend without revenue can be an investment, if you can see it.
The legacy admin console used $1.7k of AI this month and returned nothing, but the agents are documenting it so it can be retired. The spend is a migration cost you can follow to its end, because it's logged to a capability and a ticket. Without that link, it is a line in the AI bill nobody can explain.
Measure across several windows, not one moment.
Revenue moves for reasons that have nothing to do with agents, so one month's revenue gain proves little. Trust the cost side first, because tags and ticket keys put each cost on a capability without anyone's judgement, and count a revenue gain only if it holds across several months on the same chain under the same rule.
Metacosm holds the chain, and pulls the cost onto it.
Metacosm's world model already records the links this page relies on: which capabilities each customer needs, what each capability depends on, and which team owns it. It turns complex ROI calculations into simple addition.
How cost reaches a capability
- cloud
- Assigned by resource tag, from each provider's bill.
- ai agents
- Each agent trace carries a ticket key, and each ticket belongs to a capability, whichever provider the tokens came from. Retries and rework carry the same key, so a model that needs three attempts shows three attempts' spend, which a per-provider invoice cannot show.
- unattributed
- Spend with no tag and no ticket lands in a pool of its own. Its size is how much AI spend you cannot yet account for.
cost of running last 30 days
$3,492
cloud
$1,668
ai agents
$5,160
total
cloud · by bound resource
EC2 · api + job runners aws$2,041
Kubernetes · namespace opencost$698
RDS · reporting replica aws$431
Workers · edge cloudflare$322
ai agents · by ticket
TM-1042 Retire the legacy admin console$612
TM-996 Document the console permission model$402
TM-1038 Reporting replica for usage data$363
other tickets$291
trace → ticket → capability. pulled, never authored.
Two pieces of plumbing go in first.
Tags on cloud resources and a ticket key on every agent trace. Until both are in, the AI bill is one line.
The same view says where the next agent goes.
The return by customer is a view over the same model as the map and the team topology, so the three cannot disagree. It shows which chains cost the most and where agents have already paid back, so the next agent goes where the numbers point, not to whichever team asked first.
Make the transition with us.
Metacosm is being built with a small number of Australian scale-ups making the move from AI-assisted to AI-native. Design partners shape the roadmap, get senior attention from both practitioners, and pay early-partner terms.