Construct a benefits case

Saying “the transformation saved 80,000” is inadequate without knowing what changed, relative to what, for whom, and whether the saving is cash or staff time that can be redeployed. A **benefits case** joins outcome definition, baseline, mechanism, measurement, costs, assumptions and decision owner. Not every benefit is monetary: user time or decision accuracy can improve. Combining unlike benefits into one unexplained figure makes honest review impossible. The GOV.UK Service Manual distinguishes cashable and other benefits, and calls for a baseline and testing alternative solutions before treating benefits as established. We use that reasoning for a fictional organization, not the UK government's spending obligations.

Write a short causal chain: accurate status message → customer understands the request's position → unnecessary enquiry calls fall → staff time is released or an additional cost is avoided. Each arrow needs a check. A message may be opened but not understood; it may be understood but outdated; calls may decline because the phone channel failed. “Messages sent” is therefore not a final benefit. Measure enquiry calls per request, message accuracy, decision time and complaints from people who do not use the digital channel. These guardrails stop an attractive single metric from masking a worse service.

This is the change-comparison logic described in the World Bank's impact evaluation handbook for groups with and without an intervention. The arithmetic alone does not establish causality. At a minimum, assume the regions would have followed similar trends in follow-up rates without the new messages, and that no important additional change occurred only in A. Examine several earlier quarters if available, request-type mix, staffing, eligibility rules and movement of customers between regions. If B's decline reflects a general campaign that also reached A, comparison may help. If A received undocumented special training, the effect attributed to messages may be overstated. State the uncertainty rather than stamping the table “proven impact.”

Convert the result into a conditional financial case. In the example, a follow-up call takes 15 minutes and a loaded hour of staff capacity is valued at 40 monetary units. For A's volume, 400 additional avoided calls per quarter represent 100 hours, or 4,000 units of capacity value per quarter, and 16,000 a year if demand and effect persist. That value is not automatically cash removed from the budget. Staff may be able to serve other work, which is a separate operational outcome to measure. If messages expand to a service with 50,000 requests per quarter and the same effect holds across its categories, the modeled benefit is 2,000 calls and 500 hours, or 20,000 units a quarter and 80,000 a year. This extrapolates from a bounded trial; it is not an observed result.

Suppose initial build cost is 120,000 and maintenance is 20,000 a year. Under the optimistic extrapolation, arithmetic annual capacity value net of maintenance is 80,000 − 20,000 = 60,000, giving a simple initial payback of 120,000 / 60,000 = two years. If only half the effect holds in the wider service, capacity value is 40,000, net value is 20,000 and payback becomes six years. If staff time cannot be used productively or an incremental hire avoided, neither case should be presented as a cash saving. Incorrect messages or support for users who cannot receive notifications may add costs; do not assign them zero merely because measurement is incomplete. The expansion assumption changes the decision even though the arithmetic is sound.

A reviewable benefits register entry would read: “reduce unnecessary enquiry calls per 100 completed eligible requests relative to comparable requests, without increasing wrong status messages or time to decision.” Attach the definition and counting window of a follow-up call, data source, measurement owner, pre-trial rate, post-trial rate, comparison group, capacity valuation method and review date. Keep forecast expansion and observed trial results in separate columns. If the service mechanism changes, update the definition or create a new baseline. Comparing unlike denominators cannot demonstrate success.

The decision is to expand only to request types resembling the trial, monitor accuracy and call rate weekly, and defer a whole-service financial claim until more varied groups are observed. A stop condition is wrong-status messages above a threshold set by the service owner before launch, or rising delay complaints despite fewer calls. If evidence is insufficient to set a safe threshold, develop one with service and risk staff before expansion. A two-year payback calculation does not outrank decision correctness or a user's ability to understand the status of a request. Financial value is one element of operating a service, not a substitute for quality monitoring.

Test sensitivity to the conflict rate. If it reaches 30% instead of 15%, review work becomes 30,000 minutes, postchange total 110,000, and released capacity falls to about 166.7 hours. If new entry takes twelve rather than eight minutes, another 16,000 minutes is needed, nearly erasing the benefit when conflict is also high. A funding gate should therefore identify assumptions able to reverse the decision: observed entry time from a sample, conflict share and postreview decision quality. The repair may still make an answer clearer to users, but measure that separately from staff hours. A single average must not hide that conflicts concentrate among people needing longer assistance.

Measurement and interpretation are different tasks. Before launch, collect at least a suitable spread of cases across channels and compensation types and fix when work time begins and ends. After launch, keep the denominator and include cases passed into manual review. If half the complex claims move to another team, the first team's apparent speed can rise while agency effort does not fall. Finance and operations should review transfer, training, support and reopening cost. If the new flow adds a document request for users, count their effort before declaring experience improved. The useful board decision is not “the metric is green”; it is “expand a limited slice because a conservative benefit range remains positive without reducing correctness,” or “stop until the conflicts are understood.”

Audit the case beyond conflicts already detected by the new system. Sample records it classifies as “no conflict” to find false negatives: two people might have matching names but different requests, or an effective date might shift without an alert. If ten errors appear in two hundred apparently clean cases, do not automatically generalize a five-percent rate to every claim without a suitable sample design. Treat it as a stop signal for any claim that correctness has been proved. Include correction work and user contact in a conservative benefit range. Name who may accept residual error and which error is unacceptable at any frequency. If the change speeds a financial entitlement, one wrong correction can cost more than many entry minutes. Funding therefore needs a harm scenario alongside average time; the justified next step may be a data-quality test before expansion rather than a binary choice between old and new forms.