An AI agent can make the work look automated while leaving a person responsible for everything that goes wrong. The missing line in the business case is often the cost of human rescue.
A cheaper model call does not necessarily produce a cheaper completed task. Someone may still have to inspect the result, find the missing context, repair an integration, reconcile a payment or finish the work manually. Those minutes belong in the cost of the outcome.
DN’s thesis: measure the human effort required to deliver an accepted result, then distinguish planned oversight from unplanned rescue. Automation creates economic value when the combined system improves cost, quality or capacity under comparable conditions.
What Matters
- Count accepted outcomes. Attempts, tool calls and reassuring completion messages are different units.
- Separate oversight from rescue. A planned approval is not automatically a failure. Unplanned diagnosis and repair are additional work.
- Count failures in the numerator. Abandoned tasks still consume compute and labour.
- Compare equivalent quality. A cheaper result that misses the deadline or violates the acceptance criteria is not the same product.
- Protect scarce expertise. A small exception rate can consume expensive specialists if each exception is difficult.
Evidence Block
Source review: 13 October 2026. This article combines published research, engineering documentation and DN’s proposed accounting framework. DN has not run a cross-vendor human-rescue benchmark.
The 2023 NBER working paper Generative AI at Work reported roughly 14% higher productivity, measured as issues resolved per hour, among customer-support workers using an AI assistant. Benefits differed across experience levels. This is evidence from that deployment, not an autonomous-agent rescue-cost estimate.[1]
METR’s July 2025 study found that 16 experienced open-source developers took 19% longer on the studied tasks when allowed to use early-2025 AI tools. Its February 2026 follow-up reported that selection effects and time-measurement problems made the later experiment an unreliable estimate of the current productivity effect. The original result should not be presented as a universal verdict on today’s tools.[2] [3]
Anthropic’s evaluation guidance distinguishes the agent’s transcript from the actual outcome in its environment. That distinction matters economically: a claim that a booking or payment succeeded is not the same as a verified completed action.[4]
What remains unknown: there is no representative rescue-cost estimate established by these sources across agentic finance workflows. The calculator below uses your inputs; its defaults are illustrative.
The Labour Bill Hidden Inside AI Automation
Imagine a finance team that receives a polished reconciliation report every morning. Most entries are correct. A few contain inconsistent identifiers, stale exchange rates or a transaction whose settlement status is unclear.
The agent has saved drafting time. The team still needs someone who understands the underlying systems to investigate exceptions. If that person spends an hour reconstructing a transaction history, the workflow has transferred work into a more specialised queue.
This does not make the agent useless. It changes the question. How much human effort did the system remove, how much did it introduce, and what quality of service did the combined workflow deliver?
The problem becomes harder to see when rescue happens outside the product: a Slack message to an engineer, a manual spreadsheet adjustment, an unrecorded customer call or a founder fixing errors after hours. A dashboard can show excellent agent completion while missing the labour that made those completions possible.
A DN Taxonomy of Human Work
| Category | Example | Accounting treatment |
|---|---|---|
| Planned oversight | A reviewer approves a payment above a defined limit. | Count the time; do not label it rescue solely because a human participates. |
| Unplanned rescue | An operator investigates an ambiguous transfer and repairs the workflow. | Count diagnosis, coordination, correction and incremental retesting. |
| Manual fallback | The agent cannot complete the task; a person finishes it. | Count the human completion effort and the preceding agent costs. |
| Operational overhead | Monitoring, exception triage, maintenance and incident review. | Allocate relevant labour to the measurement period. |
| Business loss | Refunds, duplicate-payment loss or a contractual penalty. | Record separately from labour; include only when quantified without double counting. |
| Waiting and opportunity cost | A task waits for an authorised specialist. | Track elapsed delay separately. Monetise only with an explicit defensible method. |
A task may receive routine review and then require rescue. Both can be counted, provided the time intervals do not overlap. Several people working for ten minutes represent more labour than one person working for ten minutes.
Retries are not new business outcomes. Group attempts under a stable task identifier so that repeated runs do not inflate the apparent completion rate.
The Unit That Changes the Business Case
Cost per accepted outcome = all included workflow costs ÷ accepted outcomes.
The numerator includes work spent on rejected and abandoned tasks. The denominator includes only outcomes that meet the predefined requirements within the measurement window. Unresolved tasks should remain visible rather than silently disappearing from the report.
For comparisons, use the same task mix, acceptance standard, deadline and cost boundary. If one approach completes fewer tasks, compare unit costs and separately disclose the difference in coverage. Equal unit cost does not imply equal capacity.
Loaded hourly labour cost can include salary, employer costs and an appropriate overhead allocation. It is not necessarily the hourly amount invoiced by a contractor. Use actual role-specific rates where possible; the calculator uses a blended rate for simplicity.
DN Human Rescue Cost Lab
Model one measurement period in a single currency. Defaults are a hypothetical example, not market averages or vendor results. Planned-review minutes are averaged across every attempted business task. Rescue minutes include all rescue episodes per rescued task and exclude planned-review time.
Inputs stay in this tool’s page memory; the script does not transmit or save them. Ordinary site analytics operate separately. No external scripts, API calls, wallets or credentials are required. If a publishing editor removes scripts, the worked example and formulas remain readable.
Worked Example: When Compute Is the Small Line Item
In the default hypothetical scenario, 1,000 business tasks produce 900 accepted outcomes. Technology costs 500 currency units. Planned review takes 500 minutes; rescue takes another 1,800 minutes. At 60 units per hour, those activities cost 500 and 1,800 respectively. Five operational hours add 300.
| Included cost | Illustrative amount |
|---|---|
| Technology | 500 |
| Planned review | 500 |
| Unplanned rescue | 1,800 |
| Other operational labour | 300 |
| Total included workflow cost | 3,100 |
| Cost per accepted outcome | 3.44 |
Human labour represents approximately 84% of the included cost. The workflow still costs less per accepted outcome than the assumed manual benchmark of 8. A large human share can coexist with useful automation.
Under these assumptions, the unit-cost parity threshold is approximately 39.3 additional rescue minutes per rescued task. Above it, the modelled workflow costs more per accepted outcome than the manual benchmark. This is arithmetic at a fixed volume and acceptance count, not a forecast: in practice, changing rescue effort may change acceptance too.
Why Finance Exceptions Can Be Expensive
Financial workflows often require resolving state, ownership and authority before acting. A timeout may leave a transfer pending rather than failed. A quoted price may expire. A position can appear closed in one interface while another system still shows exposure.
A responsible operator may need to check authoritative records, establish whether an action already occurred and decide who can authorise the next step. Repeating an action without resolving that ambiguity can increase the problem.
These are illustrative failure modes, not measured incident rates. Their significance depends on the workflow. A read-only research assistant and an agent authorised to transfer assets have different consequences and should not share one undifferentiated rescue benchmark.
Human approval can be a valuable control. LangGraph’s documentation, for example, describes interrupts that pause execution and allow human input before resuming.[5] The economic goal is appropriate oversight with a manageable exception burden, supported by evidence about actual outcomes.
The Capacity Problem Behind the Average
An average of twelve rescue minutes can hide a long tail. Many exceptions may take one minute; a few may occupy a specialist for hours. Report the median, a high percentile and the maximum alongside total labour hours.
Also measure queue delay. Twenty exceptions arriving together can be harder to service than twenty spread across a week. A workflow may look economical on average while missing deadlines during concentrated incidents.
Use actual staffing schedules and escalation paths before promising continuous autonomous operation. Include standby costs where relevant, without counting the same paid time again as both overhead and task-level labour.
The Signal
Watch for rescue minutes per accepted outcome falling while quality, coverage and authority boundaries remain stable. That combination provides stronger evidence of operational improvement than a rising count of agent runs.
DN Alpha Thesis
As some forms of inference become cheaper, exception handling may become a more important source of differentiation. Systems that preserve useful traces, make ambiguous state legible and support efficient recovery could improve workflow economics even without winning a model benchmark.
This is a conditional commercial thesis, not an investment recommendation or a forecast of vendor winners. It weakens if model improvements eliminate most rescue, if recovery features commoditise quickly, or if customers cannot capture the labour savings.
A Measurement Protocol Teams Can Actually Use
- Define the cohort. Specify task type, risk level, deadlines and system version before collecting results.
- Set acceptance criteria. Include correctness, required evidence and any relevant authorisation constraints.
- Assign one task ID. Attach retries, tool calls, interventions and the final outcome to it.
- Record human time by category. Log planned review, rescue and manual fallback without overlapping intervals.
- Keep unresolved work visible. Report its age and the costs already incurred. Define how later completions will update the cohort.
- Compare a credible baseline. Where feasible, randomise comparable tasks or use a matched cohort with clearly disclosed limitations.
- Report distributions. Publish acceptance, rescue incidence, labour minutes, queue delay and cost per accepted outcome.
- Recheck after changes. Model, prompt, tool, permission and task-mix changes can alter the economics.
Time released from a salaried employee is a capacity gain until the organisation uses it productively or reduces an actual cost. The calculator’s difference from the manual benchmark is not automatically cash savings, profit or return on investment.
DN Rescue Ledger: A Proposed Data Structure
Use this editorial template to build your own observation dataset. It contains no empirical vendor results. Keep credentials, personal information and financial identifiers out of public exports.
| Field group | Record |
|---|---|
| Identity and version | Task ID, cohort, date, model/harness/tool versions. |
| Outcome | Accepted, rejected, abandoned or unresolved; evidence and deadline. |
| Human effort | Role, loaded rate, review minutes, rescue minutes and fallback minutes. |
| Failure and recovery | Observed symptom, contributing factors, actions and recovery evidence. |
| Cost boundary | Technology spend, allocated overhead, quantified loss and exclusions. |
| Service quality | Elapsed time, queue delay, later reopening and downstream corrections. |
Methodology, Limits and What Would Falsify the Thesis
The DN Human Rescue Cost Lab is a deterministic accounting model. It assumes a single period, one currency, a blended labour rate and stable input values. It does not simulate correlated failures, changing acceptance rates, uncertain losses, taxes or future reliability.
For mixed roles, calculate labour as the sum of each role’s hours multiplied by its loaded rate. For uncertain business losses, show separate scenarios rather than treating an invented expected loss as an observed cost.
A fair test of the article’s thesis would compare equivalent workflows and record all human intervention. If rescue is consistently negligible, or if it adds little information beyond existing total-cost measures, a separate rescue metric would offer limited value. If better recovery features do not reduce labour or improve accepted outcomes under controlled comparison, the proposed recovery advantage would lack support.
Reducing review while leaving more errors undetected does not establish improvement. Include delayed corrections and downstream harm within a declared follow-up window.
Build the Full DN Agentic Finance Toolkit
- AI Agent Cost Calculator: compare broader build, buy and agency ownership costs.
- DN Agent Trust Dataset: examine evidence about authority and controls.
- Agentic Finance Guide: connect delegation, execution and accountability.
- DN Dictionary: understand harnesses, human oversight and recovery concepts.
Sources
- Brynjolfsson, Li and Raymond — Generative AI at Work, NBER Working Paper 31161. The historical working-paper result is used here; it is not a current agentic-finance estimate.
- METR — Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. Published 10 July 2025.
- METR — We are Changing our Developer Productivity Experiment Design. Published 24 February 2026; important limits on interpreting the follow-up.
- Anthropic — Demystifying evals for AI agents. Published 9 January 2026.
- LangChain — LangGraph interrupts documentation. Feature documentation, not independent performance evidence.
Frequently Asked Questions
What is the cost of human rescue?
It is the labour cost of unplanned human diagnosis, correction and completion work required after an AI workflow encounters a problem. DN tracks it separately from planned oversight.
Is every human approval a rescue?
No. A scheduled approval is planned oversight. It still has a cost, but it should not automatically be classified as a failure.
How do you calculate cost per accepted outcome?
Divide all included workflow costs, including spending on unsuccessful tasks, by the number of outcomes meeting the predefined acceptance criteria.
Can a workflow with high rescue costs still be worthwhile?
Yes. It can still improve cost, quality or capacity compared with an equivalent baseline. Evaluate the combined system rather than the human share alone.
Does the calculator estimate real vendor performance?
No. It calculates a scenario from your inputs. Default values are hypothetical and are not measured vendor results.
Are modelled savings the same as cash savings?
No. Reduced labour time can create capacity without reducing payroll. Cash savings depend on costs actually avoided or reduced.
What happens when there are no accepted outcomes?
Cost per accepted outcome and unit-cost comparisons are undefined. The tool still reports included spending and labour.
Does this article prove that AI reduces productivity?
No. The cited studies cover different settings and periods. The article argues for measuring the actual workflow, including human effort and quality.
Does the tool send my inputs anywhere?
The calculator script does not transmit or persist inputs. Ordinary site analytics operate separately.
Disclosure: DN’s framework, taxonomy and calculator are editorial work. Examples are hypothetical. This article provides operational education, not financial, legal or investment advice. No vendor paid ranking or safety certification is presented.
About the Author
