Value under error costs and adoption

Why automated volume is not value The first opportunity lesson established a digital support flow of 1,000 requests per week, with 720 correct first handoffs and 280 transfers. A product manager can promise fewer transfers, but an investment decision needs more than multiplying requests by a minute rate. Some transfers cost little. Others expose a sensitive request to an inappropriate team. Users may change behavior when the intake interface changes. Build a small value model that shows these assumptions separately, then test the assumptions with the most influence. The model supports a decision under uncertainty; it is not an accounting record of future results. Define the economic unit before calculating: one request entering the digital channel during a week. The baseline contains 280 transfers per 1,000 requests. Suppose one transfer uses staff time and operations worth an estimated 4.50 monetary units. This is a fictional case assumption, not a market price or a fact about a particular center. Also assume 10 severe misroutes, each with an additional remediation cost of 20 units above ordinary transfer cost. Estimated ordinary transfer cost is 280×4.50=1,260. Severity adds 10×20=200, producing a weekly total of 1,460. Do not silently count customer waiting time as lost revenue. Record it as an experience measure until its financial effect is established, or treat it as a separate non-negotiable guardrail.

Compare alternatives without double counting Option A repairs the intake field and routing rules. In this hypothetical forecast it reaches 800 correct first handoffs and 200 transfers, while severe misroutes stay at 10. Its weekly cost, after spreading build effort across expected use, is 80 units. Savings against baseline are (280−200)×4.50=360. Subtract 80 and net operational value is 280 units a week. Do not add the value of “80 newly correct handoffs” again; their effect is already in reduced transfers. Counting the same benefit twice is a common way for a weak project to look attractive. Option B is a full classifier. It is forecast to reach 830 correct first handoffs, leaving 170 transfers. Its assumed weekly costs are 80 for model operation, 100 for human review of low-confidence cases and 120 for data maintenance and evaluation, or 300 in total. There is a warning signal: severe misroutes may increase from 10 to 20, ten additional cases. Ordinary savings are (280−170)×4.50=495. Subtract 300 in operation and an incremental 10×20=200 in severity handling, yielding −5 units per week. First-handoff correctness improves from 72% to 83%, yet modeled value does not. More seriously, a severity guardrail may prohibit launch even if the financial value turns positive. Harm to a sensitive group cannot simply be canceled by convenience gains for routine cases. Option C is hybrid: rules handle clear cases, a classifier addresses the remainder, and sensitive cases receive review. Its forecast is 850 correct first handoffs, 150 transfers and eight severe misroutes, at an all-in weekly cost of 450. Ordinary savings are 130×4.50=585. Two fewer severe incidents save a further 40 in modeled remediation. Subtract 450 to obtain net value of 175. That is below Option A's 280 despite higher routing accuracy because complexity costs money. The result does not prove hybrid designs are generally poor. The hybrid might shorten resolution time or help other channels, but those benefits have not yet been measured. The present choice favors testing A first and keeping C as an option if a tractable residual problem remains.

Separate observation from assumption The figures 800, 830 and 850 are forecasts, not observed trial results. The 4.50 cost is an average that may hide important differences. On the estimate card, record each variable's source and a plausible range. Derive transfer cost from sampled staff time, the count of transfers from weekly logs, severity from case review rather than a category name alone, and adoption from a limited field test. State how a fixed build cost is spread over expected weeks of use; a shorter life raises its weekly share. Where evidence is absent, present a range instead of a precise-looking fiction. Nor should a weekly benefit be multiplied by 52 until seasonality, volume changes, adoption and recurring costs have been considered. Suppose Option A reaches only 60% of requests because the new field is used by that share during the initial release. If a reduction of 80 transfers was forecast at full adoption, a first approximation is 0.60×80=48 fewer transfers. Adjusted value becomes 48×4.50−80=136 per week, provided early adopters resemble the rest of the channel. They may not: the first users might have the clearest issues. Measure adoption and outcomes by segment rather than applying one rate indiscriminately. If actual transfer reduction at current adoption is between 25 and 50, modeled value ranges from 25×4.50−80=32.50 to 50×4.50−80=145. Even a positive low bound within this narrow financial model does not authorize launch if resolution time or severe routing worsens.

Break-even and what money alone cannot say Show management the weekly estimate together with the request volume behind it. Fixed maintenance does not halve automatically when volume halves; recalculate when the case mix changes. Reverse the calculation. Instead of asking only “How much might we save?”, ask “At what value would this test cease to pay for itself?” With 80 fewer transfers and a weekly cost of 80, Option A breaks even when a transfer costs one unit: 80/80=1. Below that cost, its narrow financial value is negative despite a better routing measure. Option B must cover 300 of operation and 200 of additional severity handling through 110 fewer transfers. Its break-even ordinary transfer cost is 500/110, about 4.55. The assumed 4.50 lies just below it. A change of 0.10 in that estimate could reverse the sign of the financial result, so presenting −5 as a decisive fact would be poor judgment. The possible increase in severe misroutes remains a separate concern even if the financial sign changes. Alternatives cannot have their benefits added mechanically. If A launches first and B is later applied to the same requests, B cannot claim its full 110 saved transfers against the original baseline again. Measure B's incremental effect above A and charge its incremental cost, or compare both against a common control group. If B serves a segment that A never reaches, identify that segment and its volume before adding benefits. AI product investments often compete for the same value or depend on each other. A dependency map prevents an investment committee from paying twice for one outcome. Nor does a severity cost become a license to purchase harm. The example's 20 units represent remediation effort; they may omit loss of trust, privacy impact or delay borne by a requester. Add a separate operating constraint, such as “sensitive misroutes must not exceed baseline,” and inspect each incident before widening exposure. When the effect cannot be estimated credibly, a staged decision with human review may still be justified even if a spreadsheet shows positive net value. A sound value estimate says what is known, what remains outside the calculation and which bounded test will reduce the uncertainty most likely to change the choice.