Calculate value across the whole workflow

A faster model response is not a business case. Value depends on how many eligible tasks use the new workflow, how much human effort remains, whether freed capacity is put to use, and what the service costs to build and operate.

Define the units before the formula

Use one time period and one currency. Here the period is a month and currency is SGD. An eligible task is a task the proposed service is intended and allowed to handle. An attempted task includes failures and abandonment; otherwise the model hides unsuccessful work. Review time is averaged over every attempted task. Fallback fraction is the share of attempts requiring additional fallback work.

attempts = eligible_tasks × adoption_fraction
assisted_minutes = interaction_minutes + review_minutes
                 + fallback_fraction × additional_fallback_minutes
freed_hours = attempts × (baseline_minutes − assisted_minutes) / 60
capacity_value = freed_hours × loaded_hourly_value × realization_fraction
operating_net = capacity_value + incremental_margin
              − variable_cost − expected_incremental_loss − fixed_monthly_cost
net_after_build_allocation = operating_net − build_cost / allocation_months

When freed hours are negative, charge the full added labor cost: the realization multiplier becomes one. Do not discount an operational burden because positive benefits would be hard to realize. Incremental margin and expected loss must be incremental to the same baseline; exclude any benefit or cost already captured in handling effort.

A fully specified base case

Twelve thousand attempted tasks require 4.8 human minutes each: two minutes interaction, two minutes review and 0.8 expected fallback minutes. Compared with 1,600 baseline hours for these attempts, assisted work requires 960 hours, freeing 640 hours. Valuing half of that capacity at SGD 60/hour gives SGD 19,200/month. Variable cost is SGD 3,600. Operating net is SGD 9,600/month; after SGD 2,000 monthly build allocation, net is SGD 7,600/month.

The 8,000 unattempted tasks still consume 1,066.67 baseline hours. Total human workload is therefore 2,026.67 hours rather than 2,666.67 hours. The 960-hour figure applies only to attempted tasks, not the whole operation.

Capacity value is not verified cash savings. It becomes economically useful through an explicit plan: serve a backlog, avoid approved hiring, improve turnaround or reduce paid external work. Record which mechanism is expected and who can realize it. A reduction in expense needs its own accounting evidence.

Test the assumptions that could reverse the decision

Other inputs remain fixed. These are scenario calculations, not confidence intervals or forecasts with assigned probabilities. Demand, staff availability and dependency quotas may limit the optimistic case.

Under base assumptions, net value per attempted task before fixed cost and build allocation is SGD 1.30. Break-even adoption is therefore 8,000 / (20,000 × 1.30) = 30.77%. At 60% adoption, the maximum average review time for break-even is 3.27 minutes. If review reaches that threshold, the value case needs revision even if output accuracy remains high.

Simple payback on the SGD 24,000 build cost is 24,000 / 9,600 = 2.5 months, conditional on the assumed steady operating net starting immediately. A real cash-flow plan also includes ramp-up, payment timing, working capital, taxes where applicable, discounting and replacement investment. Do not call the simplified ratio a realized payback.

Include costs that usually disappear from the slide

Count integration maintenance, evaluation labeling, human escalation, on-call coverage, observability, storage, data rights, provider charges, network egress, training runs, failed requests, retries and idle reserved capacity. Avoid double-counting support labor in both fixed costs and a separate labor line. Separate sunk costs from future avoidable costs when deciding whether to continue.

Use cost per verified outcome, not only cost per API call. If 12,000 attempts yield 10,800 verified completions, SGD 9,600 monthly variable-plus-fixed service cost becomes SGD 0.89 per verified completion before build allocation and human effort. The completion count here is a planning assumption; it does not change the labor model unless its failure/fallback work is also reflected.

Replace assumptions with evidence

Instrument start, completion, review duration, fallback, abandonment and final accepted business state. Compare comparable tasks, periods and operators. A randomized rollout can help isolate effects; where it is infeasible, document confounding and avoid causal certainty from before/after averages.

The benefit owner reviews assumptions monthly. Update the forecast with observed adoption and review effort, preserve prior versions, and show variance. Use the interactive workbench to explore the core labor-and-cost model, then add business-specific margin and risk effects in the engagement record.