Trimio Field Notes

Where ROI Arguments Die: The Attribution Gap in AI Agents

May 21, 2026 6 min read roiagentsfinops

Two questions determine whether AI agents survive a CFO's budget review. The first: did the agent run successfully? Most teams can answer this. The second: did running it create business value? Almost none can. That gap is where AI ROI arguments die — and it's widening as agentic workloads scale.

The asymmetry is structural. Operational success metrics — completion rate, error rate, latency, token count — are generated automatically by every AI infrastructure layer. Business impact metrics — churn prevented, pipeline generated, engineering hours saved, support cost reduced — require integration with downstream systems that most teams haven't built yet.

Did it run?
Answerable
Completion rate, error rate, latency — generated automatically by every AI infrastructure layer.
Did it prevent churn?
Usually Not Answerable
Business impact requires downstream system integration most teams haven't built yet.
Cost attribution
The Prerequisite
Before ROI, you need a precise cost denominator. Attribution is where the accounting starts.

The Completion Illusion

A customer support AI agent that handles 2,000 tickets per day has a 94% completion rate. That is a real number. It means 1,880 tickets were processed without error. What it doesn't tell you:

The 94% completion rate looks like success. Whether it is success depends on the answers to questions the completion metric doesn't capture. A team with only completion metrics is making a bet: that completing tasks at high rates translates to business value. It often does. But it's a bet, not a measurement.

The CFO's budget review is where that bet gets called. "We spent $180,000 on AI agents last quarter" requires a response that isn't "we processed 180,000 tickets." It requires "we processed 180,000 tickets, which prevented an estimated X churn events at $Y average contract value, generating Z in retained revenue." The chain from cost to value has to close.

The Completion Illusion
Completion rate is an operational metric, not a business impact metric. The question a CFO asks isn't "did the agent complete?" It's "did completing it create value proportional to its cost?"
Before Attribution
Black box spend
  • $180K on agents last quarter
  • 94% completion rate
  • No cost per agent type
  • No link to business outcomes
  • CFO asks "what did we get for this?"
After Attribution
Defensible ROI
  • Support agent: $0.12/ticket × 180K = $21.6K
  • Escalation agent: $2.40/event × 5K = $12K
  • Estimated churn prevented: $180K retained ARR
  • ROI: 6.1× on support agent spend
  • CFO can sign off on next quarter's budget

Why Attribution Chains Are Hard to Build

The attribution chain from AI agent action to business outcome involves several structural challenges:

Time delay. A support ticket resolved by an AI agent today may influence renewal likelihood three months from now. The signal and the outcome are separated by a quarter, multiple touchpoints, and noise from other factors. Connecting them requires a longitudinal data model that most teams don't have.

Counterfactual reasoning. ROI attribution requires knowing what would have happened without the agent. If an AI agent escalated a churning customer to a human CSM who saved the account, the ROI is the contract value. But was the escalation the causal factor, or would the CSM have identified the churn risk through other signals? Counterfactual questions are answerable at the population level (A/B testing over time) but not at the individual event level.

Multi-system data. The attribution chain crosses multiple systems: the AI platform, the CRM, the support system, the billing system, the product analytics system. Joining these systems for attribution analysis requires data infrastructure investment that isn't part of the AI deployment story.

Definitional disagreements. Finance, product, and engineering often disagree about what "ROI" means for AI. Finance wants revenue impact. Product wants feature engagement. Engineering wants hours saved. These are different measurements, and consolidating them into a single ROI number requires an organizational alignment that rarely happens before the CFO asks for it.

The Attribution Gap — Who Can Answer What

QuestionWho Can Answer It NowWhat's Needed
Did the agent run?Any AI infrastructure layerAlready available
What did it cost?Provider billing (total only)Per-agent attribution layer
Cost per agent type?Almost no oneTagged API calls + attribution proxy
Did it generate business value?Almost no oneDownstream system integration + causal model
ROI on agent deployment?Essentially no oneAll of the above, plus alignment on "value" definition
Start With the Denominator
Per-agent cost attribution closes the cost side of the ROI equation. It doesn't close the business value side — but it gives you a precise cost foundation and identifies which agents have the highest ROI uncertainty to instrument first.

The Half the Gap That Is Solvable Today

Cost attribution closes half the ROI attribution gap — the denominator. Before you can compute ROI, you need a precise numerator (value created) and a precise denominator (cost incurred). The denominator is largely unsolved for AI today, but it's the easier problem.

With per-agent cost attribution, you can answer: this agent run, on this task, for this team, at this moment in time, cost $X. That gives you a cost foundation for the ROI calculation. It doesn't give you the numerator — that still requires business outcome integration. But it does give you a precise cost to defend against.

The practical value of cost attribution before full ROI attribution is in triage. When you know that your customer churn prediction agent costs $0.12 per customer scored, and your escalation agent costs $2.40 per escalation, and your onboarding agent costs $0.08 per user, you can prioritize the ROI instrumentation work in order of financial exposure. Start with the $2.40 agent — that's where imprecise ROI claims are most expensive.

Building Toward Full Attribution

The path to full ROI attribution requires integrating AI action data with business outcome data. The practical approach for most teams:

  1. Instrument agent runs with business context tags. Every agent invocation should be tagged with the relevant business entity (customer ID, opportunity ID, ticket ID) so the action can be joined to downstream business events.
  2. Define outcome events in downstream systems. Identify the business events that represent value: renewal, upgrade, churn prevented, ticket self-served, code deployed without incident. These need to be instrumentable events in your CRM, support system, or product analytics.
  3. Build the join. A simple table that connects agent_run_id to business_entity_id, joined over time to the outcome events for that entity, is the minimum viable attribution model.
  4. Separate correlation from causation. Report the correlation while building toward causal measurement. Don't conflate the two in reporting — but correlation is useful evidence even before causation is established.
The Prerequisite
Full AI ROI attribution is an unsolved industry problem. Cost attribution is the prerequisite — you can't compute ROI without a precise cost denominator. Start there.
Building Toward Full Attribution
Tag every agent run with a business entity ID. Define downstream outcome events. Build the join. Separate correlation from causation in reporting. Start with the denominator — it's the prerequisite for everything that follows.

Trimio closes the cost side of the ROI gap — per-agent attribution, cost per task type, and budget governance for every agent invocation. The rest of the attribution chain is still your engineering problem, but we make the denominator defensible. Start with cost attribution.

Trimio
Stop guessing. Start governing.
trimio is the LLM API gateway purpose-built for AI cost governance — visibility, routing, caching, and budget enforcement in one layer.