Definition
Return on Agent (ROA) is a metric that measures whether work done by AI agents creates enough business value to justify the full cost of running and governing it. It compares the value created by “agented work” with everything it takes to produce that value: software and model costs, but also monitoring, human oversight, data upkeep, and the cleanup when agents get something wrong.
Greg Kihlström introduces ROA in Marketing Operations 3.0 (De Gruyter Brill) as “the top-line economic measure for a hybrid human-agent system.” It’s meant to answer an executive question: is this system of people and agents creating enough value to deserve more budget, more autonomy, and a bigger role in the business?
See also: Return on Investment (ROI) · Total Cost of Ownership (TCO) · AI Agent · Key Performance Indicator (KPI)
Not to be confused with: Return on Assets, the standard financial ratio (net income ÷ total assets) that also uses the abbreviation ROA. In finance and investor contexts, “ROA” almost always means Return on Assets.
Why it matters for marketing
Marketing teams are deploying AI agents faster than they’re measuring them. Gartner predicted in June 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. All three come down to measurement.
The familiar metrics don’t fit well. Clicks, opens, conversions, and hours saved can all improve while the system quietly creates problems somewhere else: extra cleanup for the service team, compliance reviews for legal, or customers who trust the brand a little less. The book’s point is that “local metrics can look healthy while the larger system becomes more expensive, less coherent, or harder to trust.”
ROA puts value and full cost in the same view. It also stops costs from being hidden in other departments’ budgets, which the book flags as one of the most common ways AI programs look better than they are.
How to calculate ROA
The book’s canonical formula:
ROA = (Value Created by Agented Work − Total Cost of Agented Work) ÷ Total Cost of Agented Work
It’s usually expressed as a percentage. A positive ROA means the system creates more value than it costs. A negative ROA means it doesn’t, at least not yet.
What counts as value
| Value component | What to include | What not to include |
|---|---|---|
| Incremental revenue or margin | Higher win rates, better conversion quality, churn prevented, faster renewals, upsell and expansion | Revenue that can’t reasonably be connected to the agents |
| Capacity released and redeployed | Human time moved into strategy, exception handling, creative work, or customer-facing judgment | “Time saved” that disappeared into meetings, rework, or drift |
| Opportunity capture | Pipeline, retention, or upsell captured because agents acted at the right time or at a scale people couldn’t | Hypothetical upside with no credible link to results |
What counts as cost
| Cost component | What to include |
|---|---|
| Visible system costs | Model usage, API and token costs, platform fees, orchestration tools, infrastructure |
| Invisible operating costs | Monitoring, evaluation, governance, prompt and policy management, data engineering, human oversight |
| Exception handling and recovery | Cleanup, escalation handling, fixing errors, restoring customer trust, support after bad automation |
The book’s exclusion test is simple: “If the value cannot survive a skeptical finance conversation, it does not belong in the formula.”
Worked example
A retention squad of agents runs for a quarter. These figures are illustrative.
| Amount | |
|---|---|
| Incremental margin from churn prevented | $420,000 |
| Capacity redeployed to high-value work | $90,000 |
| Opportunity capture (expansion deals agents surfaced) | $60,000 |
| Total value created | $570,000 |
| Visible system costs | $120,000 |
| Invisible operating costs (monitoring, oversight, data work) | $150,000 |
| Exception handling and recovery | $60,000 |
| Total cost of agented work | $330,000 |
ROA = ($570,000 − $330,000) ÷ $330,000 = 72.7%
The same program measured only against its visible costs would look like this: ($570,000 − $120,000) ÷ $120,000 = 375%. The gap between 375% and 72.7% is exactly the hidden cost ROA is designed to surface.
How to utilize ROA
Driver and guardrail metrics
The book is clear that ROA shouldn’t stand alone. It’s the outcome metric, supported by two other classes of measurement:
- Driver metrics explain why ROA is moving. Examples: action relevance, escalation quality, trust preservation, value realization, decision quality, timing quality, and the Hybrid Squad Efficiency Ratio (HSER).
- Guardrail metrics show whether the system is safe enough to scale. Examples: escalation appropriateness, compliance exceptions, drift rate, hallucination frequency, offer integrity, trust degradation, error recovery time, and brand safety violations.
A positive ROA with failing guardrails means fix before scaling. Improving drivers with weak ROA means redesign. Weak drivers plus weak guardrails suggests the use case itself is unsound.
Turning ROA into decisions
The book ties ROA to five executive decisions:
| Decision | When |
|---|---|
| Scale | ROA is positive and strengthening, guardrails are stable, and results are repeatable |
| Fix | ROA is promising, but data quality, orchestration design, or governance gaps are holding it back |
| Reallocate | The capability works, but it’s aimed at the wrong use case |
| Constrain | Short-term returns look good, but trust, compliance, or safety burdens make expansion unwise |
| Kill | ROA stays weak, strategic fit is poor, or governance costs outweigh value |
The 90-day ROA review
The book recommends reviewing ROA every 90 days with five questions:
- Is ROA improving, weakening, or flat?
- Which driver metrics explain the movement?
- Are the guardrails stable enough to trust the trend?
- Is the current use case still the right one?
- What decision follows: scale, fix, constrain, reallocate, or stop?
Common use cases
- Deciding whether an AI pilot should move into production
- Comparing agent deployments across marketing, sales, and service on one shared measure
- Building budget requests for AI programs that a CFO will accept
- Serving as a shared KPI across functions under Commercial Operations Orchestration (CO2), so no team can claim success while pushing costs onto another
Comparison: ROA vs. related metrics
| Metric | Formula (simplified) | What it measures | Key difference from ROA |
|---|---|---|---|
| Return on Agent (ROA) | (Value from agented work − total cost) ÷ total cost | Economic return of a governed human-agent system | — |
| Return on Investment (ROI) | (Gain − cost) ÷ cost | Return on any investment | Same structure. ROA specifies what counts for agent systems, especially governance and recovery costs. |
| Return on Ad Spend (ROAS) | Revenue ÷ ad spend | Revenue efficiency of advertising | Revenue-based, not margin-based, and limited to media spend |
| Return on Assets | Net income ÷ total assets | How efficiently a company uses its assets | A corporate financial ratio. Unrelated despite the shared acronym. |
| Total Cost of Ownership (TCO) | Sum of all costs over a lifetime | Full cost of owning a technology | Cost only. TCO thinking informs ROA’s denominator. |
| Hybrid Squad Efficiency Ratio (HSER) | Productive output per human lead managing agents | Leverage a human gets from agents | A driver metric that feeds ROA, not a substitute |
Best practices
- Count all the costs. Pull oversight, cleanup, legal review, and data work back into the denominator, even if they sit in other budgets.
- Only count redeployed time. Hours saved count only if they went into higher-value work.
- Use margin, not revenue, wherever you can, so discounting agents don’t look artificially strong.
- Keep a baseline. Measure the pre-agent process so incremental value is real.
- Never scale on ROA alone. Check guardrails first.
- Make every scorecard line lead to an action. The book’s test: “If the scorecard does not change a decision, it is reporting, not management.”
- Be careful in pilot mode. Small pilots often look cheap because governance costs haven’t shown up yet.
Future trends
- Agent-level cost accounting. Expect finance teams to track token, platform, and oversight costs per agent or per squad, making ROA easier to calculate.
- Standardized AI value frameworks. As more agentic projects face budget scrutiny, outcome-based frameworks like ROA are likely to replace adoption and usage metrics in executive reporting.
- Machine-mediated value signals. Measures such as Share of Model, citation inclusion, and agent negotiation conversion will increasingly feed the value side of ROA for brands selling through AI intermediaries.
- Governance cost visibility. Regulation and internal AI policies will make oversight costs harder to ignore, which will lower many naive ROI estimates.
- Shared cross-functional KPIs. ROA-style metrics are likely to show up in joint marketing, sales, and service scorecards.
FAQs
What does ROA stand for in AI? Return on Agent, a measure of whether AI agent work creates enough value to justify its full cost. In finance, ROA usually means Return on Assets, which is unrelated.
Who created the Return on Agent metric? Greg Kihlström defines it in Marketing Operations 3.0 (De Gruyter Brill).
What is the Return on Agent formula? ROA = (Value Created by Agented Work − Total Cost of Agented Work) ÷ Total Cost of Agented Work.
How is ROA different from ROI? The math is the same structure. The difference is in what counts. ROA requires including governance, monitoring, human oversight, and error recovery costs, and it only counts saved time when that time is actually redeployed.
What is a good ROA? There’s no universal benchmark. Any sustained positive ROA with stable guardrails is a case for scaling. A high ROA with failing guardrails is a reason to fix things first.
What are driver metrics? Measures that explain why ROA is moving, such as escalation quality, action relevance, value realization, and the Hybrid Squad Efficiency Ratio.
What are guardrail metrics? Measures that show whether the system is safe to scale, such as drift rate, compliance exceptions, hallucination frequency, and brand safety violations.
How often should ROA be reviewed? The book recommends a 90-day review cycle, alongside shorter operational check-ins.
Can ROA be negative? Yes. A negative ROA means total costs exceed the value created. That can be acceptable briefly in an experimental phase, but it’s a signal to fix, reallocate, or stop.
Why do pilots often overstate ROA? Because governance, monitoring, and exception-handling costs tend to be small or invisible at pilot scale and grow once the system goes into production.
Related Terms
- Return on Investment (ROI)
- Return on Ad Spend (ROAS)
- Total Cost of Ownership (TCO)
- Key Performance Indicator (KPI)
- Incrementality
- Business Value Realization Formula
- Share of Model (SoM)
- Hybrid Squad Efficiency Ratio (HSER)
- Guardrail Metrics
- Commercial Operations Orchestration (CO2)
Listen: related episodes of The Agile Brand
- From PegaWorld: Pega’s Tara DeZao on marketing ROI with agentic AI
- Expert Mode: The AI ROI trap and why your data is the real goldmine
Read: Measuring AI adoption is one thing; measuring impact will get real results (Martech Futurist)
Sources
- Kihlström, G. Marketing Operations 3.0. De Gruyter Brill. https://www.degruyterbrill.com/document/doi/10.1515/9783112218839/html
- Gartner. “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.” June 25, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027