What is AI Agent Cost ROI?

AI agent cost ROI is accepted outcomes per euro when you price failure, retries, and human cleanup into the ledger — not token spend alone or a generic TCO calculator pitch.

Practitioner ROI for production agent orgs looks like this: each task class routes through the cheapest capable model and role, outcomes pass a quality bar, and you compare outcomes per euro week over week. A free model that silently fails and burns an hour of recovery is expensive; a mid-tier model that succeeds once is often the better ROI.

This is agentic cost engineering applied to the question CFO guides and SaaS calculators gloss over: what did each agent lane actually deliver for the money?

Start with Operating Note 002 for task classes and positioning. The cost/ROI series hub collects routing tables, eval fixtures, and operating notes.

What practitioner ROI includes (and excludes)

In scope:

• Failure-priced accounting — retries, silent failures, and approval/cleanup minutes in the denominator (when cheap models get expensive)
• Routing discipline — promote/demote lanes from eval fixtures, not gut feel (model watch)
• Per-workflow comparison — same formula every week so lanes are comparable (failure-priced ledger week)
• Role-aware spend — class D (irreversible) stays expensive on purpose (Note 003)

Out of scope (generic TCO / CFO SaaS framing):

• Vendor bill totals without outcome attribution
• One model for every task class
• Dashboards that crown the lowest $/1M tokens without pass rate
• Hire-funnel “book a demo” ROI calculators with invented benchmarks

A minimal ROI formula

For one agent workflow over a week:

ROI_proxy = accepted_outcomes / (model_spend + infra_share + human_hours × hourly_rate)

• accepted_outcomes — units that passed your quality bar (published post, green PR, closed ticket)
• model_spend — API cost attributable to that workflow
• infra_share — portion of your stack (e.g. the €35/mo content org band)
• human_hours — approval, cleanup, and incident time

You do not need perfect finance. You need the same formula every week so routing changes produce comparable numbers.

How routing connects to ROI

Routing tables assign each task class to a role and model tier. ROI tells you whether that assignment still holds:

1. Log outcomes, spend, and human minutes for seven days.
2. Compute ROI_proxy once per lane.
3. Promote one lane and demote one lane based on eval fixtures — not token price alone.

Downloadable starting point: Routing table v1 with A–D task classes and example rows.

Next step

If agent cost ROI, model routing, or operating an agent org is your bottleneck — contact me with a short note on what you are building. Browse the full cluster at the cost/ROI series hub.