The problem finance brings
Token spend rises every month with no per-agent attribution. Finance starts asking which agent cost what, and the platform team has no way to answer. Cost is usually the first of three problems to arrive, ahead of governance and security, and it comes from the same gap as the other two: nothing sits at the moment the agent acts that can count, attribute or say no.
More than 40% of agentic projects are forecast to be cancelled by 2027 on cost and control (sources as printed on the live site: S&P Global and McKinsey, 2026; Gartner; Deloitte). A budget you only see at month end is not a control.
Enforced on the call
In KPATH a budget is a policy check, not a report. Identity, policy, budget and human approval are checked on every request before the call proceeds. A team’s spend limit can be enforced on the call itself, so the budget stops the spending instead of reporting on it afterwards.
A call that would breach the limit is refused at the call, and the refusal lands in the tamper-evident record with the same detail as an allowed call: which agent, which target, which action, which decision. No credential is added, so the agent has nothing to retry with.
This is the Enforce step of the Verify-Enforce-Stop-Prove test applied to money rather than permissions.
Attributed to the agent, rolled up to the request
Enforcement needs attribution, and attribution is worth having on its own. Spend is attributed to each agent either way, whether a limit is enforced or not.
Attribution runs in two directions.
| Level | What it answers |
|---|---|
| Per agent | Which agent cost what, against a named owner and a risk tier |
| Per originating request | Which piece of work caused the spend, across every agent in the chain |
The second is the one most estates cannot produce. Agents call other agents: an assistant hands work to a specialist, which calls a third. Each agent’s spend is real, but the cost that matters to finance is the cost of the piece of work that started it. KPATH governs the chain as one piece of work, sees the delegation chain in the request envelope, and rolls the whole chain’s spend up to the request that started it.
What gets metered
KPATH sits between every agent and everything it reaches: enterprise APIs, other agents, MCP tool servers, external SaaS proxied through it, and the models the agents call. It meters model calls too. So the number finance sees covers what a piece of work cost across every system it touched, not the model bill alone.
Discovery cuts spend too
An agent that can see every service will try every service. KPATH’s governed directory shows an agent only the services policy allows for its job, so the agent is pointed at what it needs, which also cuts the tokens it spends working out what to call.
Budget and the rest of policy
Budget is one check among five, and it sits with the others. A high-risk agent past its budget is refused before human approval is asked for, so nobody signs off a payment that policy would have stopped anyway. Every budget decision is recorded, streamed to your SIEM and covered by the same tamper-evident record as every other decision.
Owners, not just totals
Because every agent identity carries a named owner, a budget refusal has someone to go to. The conversation stops being “spend went up” and becomes “this agent, owned by this team, hit its limit doing this piece of work”. That turns a cost report into cost control.