How to measure the ROI of an AI agent

"Is the agent worth it?" deserves a real answer, not a vibe. Here's a framework that keeps the number honest.

Tie each agent to one outcome metric

Don't measure activity ("messages sent"); measure outcomes. One primary metric per function:

• Sales: speed-to-lead, qualified-to-meeting rate.
• Support: first-response time, resolution time, clean-escalation rate. (more)
• Ecommerce: revenue vs. goal, promotion lift.
• Operations: time-to-detect, incidents caught before customer impact.

Set a baseline before the agent so you can attribute the change.

Count the value honestly

• Time reclaimed — hours your team no longer spends on the routine work, valued at loaded cost.
• Revenue influenced — faster follow-up and resolution convert and retain more.
• Risk reduced — issues caught early, mistakes avoided, an audit trail you didn't have.

Count the full cost

• Usage — model and tool costs per task. Good platforms attribute cost per job so you can see what each workflow costs, not just a monthly lump.
• Build — time to create the agent (a day, not a quarter, on the right platform).
• Oversight — the human review on gated actions. Real, but small if scoped well.

The honest ratio

ROI = (value created − fully-loaded cost) / cost. The trap is counting the value and forgetting the oversight and usage; the other trap is counting the cost and forgetting the risk reduced. Count both sides.

Watch the leading indicators

Outcome metrics lag. Track leading ones too: escalation rate (falling = the agent is handling more), and per-job cost (stable or falling = it's getting efficient). Per-job cost visibility is one reason to run agents somewhere that attributes it.

Start by picking the one metric that matters and measuring the baseline. Then book a demo and watch it move.