The bill is the worst way
to find out an action was wrong.

AI coding agents misfire in cost-shaped ways: they spin up too many instances, they overscale, they delete things you had to rebuild, they pick the wrong region. Espada is the last gate before the cost hits.

The shape of agent cost

We track four buckets that show up in real AWS-shop pilots:

  • Direct waste. The agent provisioned an oversized RDS instance, a wrong-region EKS cluster, a never-shut-down ECS task.
  • Incident-driven. The agent did something destructive; you spent engineer-hours fixing it; the cost of those hours.
  • Misalignment. The agent did the right thing in the wrong env (prod when meant for staging); recovery cost + impact.
  • Approval-loop. The cost of the gate itself: approver time, response delay.

Espada catches buckets 1, 2, and 3 before the call runs. Bucket 4 is the cost of running the gate; we measure it explicitly so you can see net.

The cost-savings dashboard

Espada ships with a built-in cost-savings dashboard that aggregates per-call estimated-avoided-cost from the audit chain. Spec at docs/operations/cost-savings-dashboard.md; ROI breakdown at docs/operations/roi-dashboard.md.

espada admin roi --explain

Period: 2026-04-15 to 2026-05-15

  Blocked actions:                  142
    of which estimated-destructive:  47

  Top blocks by estimated avoided cost:
    1. iam:DeleteUser on production-app-role:   $48,000 (would have outaged for ~6h)
    2. rds:DeleteDBInstance prod-billing-pri:   $32,000 (recovery time × on-call hours)
    3. s3:DeleteObjects s3://prod-billing-arc:  $18,000 (24h re-derivation)
    ...

  Approver friction (gross cost of running the gate):
    218 approvals × median 4 min × $150/hr =     $2,180

  Subscription cost (3 AWS accounts × $1,000):   $3,000

  Estimated net (this period):                  +$93,820

Caveat: avoided-cost estimates use customer-provided incident-cost inputs
and our default 80% catch rate. The audit chain is the source of truth
for what happened; the dollars are derived from your assumptions.

Try the calculator

A back-of-envelope ROI calculator — your inputs, the math is shown — at our ROI calculator (also published in the repo at business/sales-assets/roi-calculator.html).

What we don't claim

  • 10x ROI. We don't quote it. Real numbers depend on your incident rate and your incident cost. Our calculator surfaces both.
  • "Saves you $X by default." Espada doesn't save anything you don't have at risk. If your agents never destroy anything, Espada is insurance you don't yet need.
  • That FinOps replaces other tools. CUDOS, Vantage, and AWS Cost Explorer do the post-hoc analytics well. Espada prevents the actions; CUDOS et al show you what they did. Composes well.

Agent budgets (preview)

Beyond preventing destructive actions, Espada's agent-budget gate can cap per-agent cost-impact over a window (e.g. "this agent may not provision more than $500/day of new RDS capacity"). See extensions/agent-budget-gate for the implementation. Configuration documented in the admin guide.