The Economics of Verified Work

A workflow is not economical merely because the model is inexpensive.

It is economical when the organization can repeatedly reach the correct final state, with an acceptable level of risk, at a total cost below the value created.

Attempt cost is not outcome cost

TheAgentCompany provides a useful illustration.

In the NeurIPS 2025 benchmark, Gemini 2.5 Pro operating through OpenHands fully completed 30.3% of 175 workplace tasks. It received a partial-completion score of 39.3%, used an average of 27.2 model steps, and incurred approximately $4.20 in measured API cost per attempted task. The reported cost excluded enterprise integration, human review, exception handling, rework, monitoring, and the consequences of an incorrect action.

The $4.20 figure is a cost per attempt. It is not a cost per completed business outcome.

A purely mechanical calculation would divide $4.20 by the 30.3% full-completion rate:

That ratio is useful as a warning, but it is not a benchmark-reported cost-per-completion figure.

It assumes that:

• every attempt has the same cost;
• every completed task has the same value;
• failed tasks produce no additional expense;
• tasks are independent;
• no human review is required; and
• no incorrect action causes downstream loss.

In a real business workflow, the actual cost per verified outcome would usually be higher.

Six outcomes that should not be mixed

Before calculating unit economics, the organization should distinguish what happened to each case.

These categories have different economic meanings.

A verified automated completion creates an operational outcome. An appropriate handoff is a correct control decision, but it still requires human work before the original outcome is achieved. A false completion may look successful in a dashboard while creating investigation, correction, customer, financial, or compliance costs.

Appropriate handoff is not verified completion. False completion is not productivity.

A working economic model

The following is a Nova Epitome working model, not an established accounting or industry standard.

The total cost of a workflow over a defined period can be represented as:

The complete workflow metric is then:

This denominator includes only cases in which the target business state was eventually reached and verified. Appropriate handoffs that remain unresolved are not counted as completed outcomes.

A second metric can isolate the automated path:

The organization should not use one metric in place of the other. The first evaluates the entire operating workflow. The second evaluates the economics of the automated path itself.

The hidden cost of false completion

A visible failure is often less dangerous than a false completion.

When a system says that it failed, the organization knows that work remains. When it says that the task is complete incorrectly, the error may move into another system, reach a customer, trigger a payment, affect a report, or remain undiscovered until much later.

The expected financial effect can be represented as:

Possible components include employee time spent detecting and investigating the problem; correction and reprocessing; customer remediation; duplicate or incorrect payment; delayed cash collection; service-level penalties; legal or compliance response; security incident handling; lost revenue or opportunity; and reputational impact where it can be estimated responsibly.

The 2026 agent-reliability study found that action costs became less predictable on more complex tasks and that some models used substantially more actions when tasks became difficult. It also noted that even infrequent high-severity failures, including unauthorized data exposure or incorrect financial transactions, can carry disproportionate cost.

A low average error rate may therefore be economically unacceptable if the tail risk is severe.

The economics of an appropriate handoff

An appropriate handoff should not be valued as though the business process had already been completed. But it may prevent a much larger loss.

Consider an agent that identifies conflicting payment information and pauses before releasing funds.

Its economic contribution may include preventing an incorrect payment; reducing the reviewer's investigation time; preserving the evidence; keeping the process in a safe state; and allowing the person to resume from a defined point.

A cautious economic estimate is:

This value depends on the counterfactual. The organization must ask: What would probably have happened if the system had continued?

If the answer is "the existing manual process would also have caught the error," the incremental value of the handoff may be limited. If the answer is "the payment would probably have been issued incorrectly," the avoided loss may be substantial.

The value of handoff should therefore be measured separately from the value of completion.

An illustrative workflow scorecard

The following example is fictional and is included only to demonstrate the calculation structure.

Assume an invoice workflow processes 1,000 cases in one month.

The 220 human-resolved cases are a subset of the 250 handoffs. The final number of verified business outcomes is therefore 650 + 220 = 870.

The apparent model-and-tool cost is:

The complete program cost is:

The full workflow cost per verified business outcome is:

The $1.30 figure and the $8.62 figure describe very different things. The first is a narrow technology expense. The second is closer to the economics of the actual business process.

Neither number is meaningful until it is compared with the measured baseline.

Compare with the real baseline—not with zero

The correct comparison is not: AI costs $8.62 and doing nothing costs nothing.

The comparison is:

What did the previous process cost, and what value did it produce?

The value created by the new workflow may come from several sources: reduced human processing time; increased capacity without proportional hiring; faster cycle times; fewer errors; fewer severe incidents; improved service levels; faster billing or cash collection; and better evidence for audit and dispute resolution.

These benefits must not be double-counted.

For example, an organization should not count both "hours saved" and "salary eliminated" unless the saved capacity actually changes staffing cost. Likewise, reduced rework and avoided failure loss may overlap if they describe the same corrected incident.

Why measurable workflows should come first

Stanford's 2026 AI Index reports that 88% of surveyed organizations used AI in at least one business function in 2025, while scaled AI-agent use remained in the single digits across nearly all functions. The report also summarizes evidence that productivity gains tend to be strongest in structured, measurable work where outputs are comparatively easy to monitor.

This helps explain why the first pilot should not be the organization's most politically important or ambiguous process.

A better candidate generally has:

• a clear unit of work;
• a measurable baseline;
• a defined target state;
• authoritative evidence;
• manageable exception classes;
• bounded consequences; and
• enough volume to evaluate performance statistically and operationally.

The goal of a pilot is not to prove that AI can do something impressive. It is to establish whether the workflow has repeatable, defensible unit economics.

Four stages of economic evaluation

1. Baseline

Measure the current human or conventional automated process before introducing the agent. Record volume, labor time, cycle time, error rate, rework, exception cost, and downstream losses.

2. Shadow evaluation

Allow the AI workflow to process the same cases without changing production systems. Measure its proposed completions, handoffs, false completions, resource use, and review time.

3. Approval-gated pilot

Permit low-risk actions while requiring approval for consequential actions. Measure actual tool cost, human intervention, state recovery, correction work, and verified outcomes.

4. Exception-based operation

After performance is stable, allow the system to complete validated routine cases and send exceptions to people. Continue to monitor drift, model changes, interface changes, policy changes, and new case types.

Continuous evaluation applies at every stage.

The pilot scorecard

A useful scorecard separates outcome, control, cost, and value.

A workflow should not be scaled merely because the model price is low; the demonstration is fast; employees like the interface; the agent produces many intermediate outputs; or the autonomous-completion percentage looks high.

It should be scaled when:

• the final outcome is verifiable;
• false completion is within an acceptable threshold;
• human review is predictable;
• exception handling is operationally manageable;
• the cost per verified outcome compares favorably with the baseline; and
• remaining risks fit the organization's tolerance and responsibilities.

Pricing follows unit economics

The separate Nova Epitome article on seats, usage, and outcome-linked pricing addresses how commercial terms may be structured.

This article addresses the question that comes first.

Before selecting a seat fee, usage fee, managed-service fee, or outcome-linked component, the parties need to understand the cost of an attempt; the cost of verification; the cost of human review; the cost of exceptions; the expected cost of failure; and the value of a verified result.

Commercial terms should follow the measured behavior of the workflow. They should not substitute for measurement.

The goal is not the cheapest model

The central economic question is not: Which model has the lowest token price?

It is:

Which combination of model, tools, controls, verification, and human review produces the required business outcome at the lowest defensible total cost?

A more expensive model may be cheaper overall if it requires fewer retries; produces fewer false completions; creates better handoff packages; reduces review time; or avoids costly errors.

A cheaper model may be the better choice when the task is simple; verification is inexpensive; the action is reversible; a stronger model does not materially improve outcomes; and routing difficult cases elsewhere is reliable.

The answer will often be a portfolio: inexpensive models for classification and extraction; stronger models for difficult reasoning; deterministic rules for validation; external systems for authoritative verification; and people for exceptions and accountable decisions.

The most economical AI workflow is rarely the one with the lowest cost at every individual step. It is the one that minimizes the cost of reaching the correct final state.

Verified Work is an economic discipline

Verified Work is not only a safety or governance idea. It is also a way to prevent incomplete, incorrect, or unverified activity from being misreported as economic value.

A trustworthy business case distinguishes attempts from completions; partial progress from final outcomes; handoffs from completed work; visible output from verified state change; model cost from total workflow cost; and average errors from tail-risk losses.

AI value should be measured at the point where the business outcome is verified—not at the point where the model stops generating.

That is the economic foundation of Verified Work.

References

1. TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks — NeurIPS 2025, Datasets and Benchmarks Track. Reports a 30.3% full-completion rate, 27.2 average steps, and approximately $4.20 in measured API cost per attempted task for its highest-performing evaluated configuration.
2. Towards a Science of AI Agent Reliability — arXiv:2602.16666, version 3, June 2026. Reports that resource consistency degrades on more complex tasks and that infrequent high-severity failures may carry outsized costs.
3. Stanford AI Index Report 2026 — Economy — Reports 88% organizational AI adoption in 2025, while scaled agent use remained in the single digits across nearly all business functions; it also summarizes stronger productivity evidence in structured, measurable work.