An AI agent pays for itself when the value of a completed workflow unit exceeds the full cost of producing and governing that result. Token spend supplies one line in that calculation. Human review, exception handling, rework, recovery and the verified business outcome complete the record.
Unit economics memory joins those elements at run level. A leader can see which agent version handled the work, what it consumed, where a person intervened, which action reached the operating system and whether the expected result materialised. That record supports a commercial decision: expand, redesign or stop the workflow.
Agent ROI starts with a defined unit of work
McKinsey published a practical guide to the economics of agentic workflows on 24 August 2026. The timing reflects a practical shift in buyer attention: production agent programmes now face the same capital discipline as any other operating system.
A useful calculation begins with a named unit. “Customer support” is too broad. One correctly resolved billing query, one qualified sales opportunity accepted by an account executive or one invoice matched and posted gives the team something countable.
The unit needs a completion rule. A drafted response remains work in progress until it reaches the customer and survives the escalation window. A sales record has limited value until the CRM contains accepted data and the next step occurs. An accounts-receivable action earns credit when cash or a verified payment commitment changes the ledger state.
This boundary prevents activity from masquerading as value. Agent messages, tool calls and generated artefacts describe production. The business system confirms whether production became an outcome.
Variable execution cost needs run-level attribution
OpenAI’s API pricing separates input, cached input, cache writes and output rates across models. Agent workflows add further consumption through searches, code execution, retrieval, voice, computer use and external tools. Routing a case through several specialised agents compounds those charges.
Record cost against each workflow run:
- model, version and processing tier
- input, cached input and output usage
- paid tool calls and third-party API charges
- retries, loops and duplicate work
- storage, retrieval and execution infrastructure
- licences allocated to the workflow
- elapsed runtime and queue delay
Aggregate invoices cannot explain why one workflow became expensive. Run-level attribution can show that a weak retrieval step doubled context size, a timeout created three retries or an exception bounced between agents before reaching a person.
The economic consequence follows the mechanism. Better source scoping reduces context and review burden. A smaller model can handle a bounded classification step while stronger reasoning stays reserved for ambiguity. Caching helps when stable instructions and source material recur. Each intervention should be tested against quality and recovery, since a cheaper run that creates more corrections raises the total cost per accepted outcome.
Human review belongs inside the cost model
Review is productive work, and its cost should remain visible. An approval taking 90 seconds across thousands of routine cases carries a different burden from a specialist spending 25 minutes reconstructing missing context.
Capture the reviewer role, wait time, active handling time, decision and reason for intervention. Separate expected checkpoints from preventable review caused by weak evidence, unclear permissions or poor confidence calibration.
This evidence makes review design sharper. Repeated approval of complete, low-risk cases can justify a narrower sampling regime. Concentrated reversals around one source or action indicate a workflow defect. High queue time with low handling time points towards routing capacity. Long handling time signals that the review packet lacks the evidence required for judgement.
AI approval agents need decision memory around authority and exception handling. Unit economics memory adds labour cost and downstream result, allowing the company to judge whether the checkpoint protects more value than it consumes.
Failed runs carry rework and recovery cost
A failed agent run can consume tokens and still leave no useful output. A more consequential failure can send a wrong message, update a customer record, issue a refund or trigger work in another system.
The economic record should distinguish:
- abandoned execution with no external action
- retry completed without human intervention
- correction requiring operator time
- downstream reversal across connected systems
- customer, revenue or compliance impact requiring case management
Recovery cost includes investigation, correction, communication, compensation, data repair and senior attention. Preserve the original trace, failed step, affected actions, recovery owner, elapsed time and final disposition.
OpenAI’s Agents SDK documents traces and spans for agent runs, generations, tool calls, hand-offs and guardrails. A trace establishes execution evidence. The operating layer must connect that evidence to the ticket, order, account or ledger item where the consequence appears.
Recovery data also defines a safe authority boundary. A workflow with cheap, reversible errors can tolerate a different review threshold from one where a wrong action is expensive to unwind. Authority should expand according to observed loss and recovery performance within a named case segment.
Quality grades need commercial consequence
Trace evaluation tests whether an agent followed instructions, used the right tools and produced an acceptable path. OpenAI’s trace-grading guidance supports scoring decisions and tool calls across an execution trace.
Those grades become more useful when joined to operating results. A high-quality trace can still produce weak economics if the case waits in a queue or duplicates work already completed elsewhere. A low-cost path can damage margin when it applies the wrong concession. An apparently successful interaction can create repeat contact because the underlying issue stayed unresolved.
Build the measurement chain in four layers:
- Execution: latency, model usage, tools, retries and failures
- Judgement: evaluation score, confidence, review result and override reason
- Operations: handling time, queue time, exception rate, rework and recovery
- Outcome: accepted opportunity, resolved case, posted payment, protected margin or another verified business state
Copilot Studio agents need evaluation memory so test sets, rubrics, transcripts and fixes survive release cycles. Unit economics memory carries that quality record into production cost and realised value.
Compare marginal value as authority expands
A successful pilot handles the cleanest cases first. Expansion introduces weaker evidence, rarer exceptions and actions with greater consequence. Average pilot performance therefore provides a poor basis for wider authority.
Segment economics by case type, risk band, source quality, customer group, action, agent version and review route. Then compare the marginal cases entering the workflow. The next 10 per cent of volume may demand more context, stronger reasoning and specialist review while producing less operational value.
A practical expansion gate asks:
- Did accepted outcome value exceed full variable cost?
- Which exceptions consumed human attention?
- How many actions required correction or reversal?
- Did queue time move into another team?
- Which source, tool or model change shifted the result?
- Can the company reproduce the calculation from stored evidence?
The final question protects management quality. A dashboard number without traceable inputs becomes a negotiation between teams. Governed memory preserves the basis for the decision and allows the calculation to change when pricing, models, workflow rules or commercial priorities move.
Build one economic receipt per completed case
Start with one workflow where the outcome already appears in a trusted system. Define the unit, completion rule, value basis, permitted actions and recovery boundary. Instrument model and tool usage, then add reviewer effort, exceptions, corrections and the final operating state.
The receipt for each case should contain:
- workflow and case identifier
- agent, model, prompt and tool versions
- sources and permissions used
- execution cost and elapsed time
- evaluation result and confidence
- human review and active handling time
- actions proposed, accepted and reversed
- recovery work and owner
- verified outcome, timestamp and source system
Review receipts by segment every week during rollout. Fix the largest source of economic drag before widening scope. That may be an oversized context window, a missing system integration, a review packet that forces reconstruction or an action whose recovery path remains too expensive.
NIST’s Generative AI Profile frames AI risk work through governance, mapping, measurement and management. Unit economics memory turns those functions into an operating discipline: named ownership, mapped consequences, measurable performance and controlled intervention.
Model Operator’s Agentic Company Brain and AI Initiative Consulting packages connect workflow memory, permissions, review paths, measurement events and outcome write-back. The result is a defensible view of where agent authority creates commercial leverage and where hidden operating cost still consumes it.
Start a Model Operator build conversation or email alexander@modeloperator.io.