Production Economics and Portfolio Choice

C17.1 — Frame the decision before valuing the options

An economics record begins with the decision, not the preferred initiative.

Record:

• the decision owner and decision date;
• the outcome or obligation being pursued;
• the scarce constraint being allocated;
• the affected Production Lines, services, users, and operators;
• the horizon over which consequences will be considered;
• the next review date or evidence trigger; and
• the choices that are actually available.

“Should we fund the platform?” is incomplete. It assumes a solution, conceals the displaced work, and leaves “fund” undefined. A governable question might be:

For the next two quarters, how should the organization use twelve engineering-months currently constrained by identity specialists: continue separate service integrations, build a bounded shared capability, buy and migrate to a managed service, run a discovery to resolve adoption uncertainty, or stop lower-priority demand?

The alternatives must include business-as-usual. Continuing the current arrangement has cost, delay, exposure, and option consequences. It is not a zero-cost baseline. Stopping should also be considered where viable. If an option is excluded because it cannot meet an obligation, violates a risk boundary, or exceeds the available capacity, record the reason.

The Green Book distinguishes early longlist appraisal from more detailed shortlist appraisal and permits indicative estimates and ranges while options are still being narrowed.C17-R05 The underlying lesson is useful beyond government: do not demand expensive precision before an option is viable, and do not use early roughness as permission to conceal assumptions.

The result of framing is not a business case for one initiative. It is a visible choice set.

C17.2 — Treat delay as a local consequence

Delay matters because value, obligation, exposure, and learning change with time. But “cost of delay” is not one universal number.

For a safety treatment, delay may extend exposure to a known hazard. For an accessibility repair, it may prolong exclusion. For a migration, it may increase dual-running cost or approach a supplier deadline. For a market experiment, it may reduce the remaining window in which learning can affect a decision. For a shared capability, delay may force several consumers to build local substitutes.

A local delay estimate should state:

• what consequence changes with time;
• who experiences it;
• the period and relevant deadline;
• the evidence source and observation window;
• whether the effect is linear, stepped, perishable, or unknown;
• the range and confidence; and
• what would make the estimate invalid.

Queueing evidence helps leaders reason about congestion. When work in progress rises relative to completions, elapsed time rises under the conditions of the model.C17-A03 That mechanism supports work-in-progress discipline. It does not prove the monetary value of one feature, justify maximum utilization, or provide a common exchange rate between safety, public value, learning, and revenue.

False precision creates a political technology: whoever controls the estimate controls the ranking. If a team assigns an unsupported currency value to every day, a spreadsheet can make contested judgement look settled. Use ranges, scenarios, and switching conditions instead. Ask, “At what plausible delay consequence would the preferred option change?” The answer shows which assumption deserves stronger evidence.

Some consequences should remain non-monetized. Record them quantitatively where possible—people affected, days of exposure, inaccessible journeys, manual interventions—or qualitatively where measurement would be invalid or disproportionate. Absence of a currency value does not mean absence of value.

C17.3 — Compare lifecycle cost, not the purchase

Software choices create costs across time and across owners. Initial build or licence cost is only one part.

A proportional lifecycle view may include:

• discovery and option appraisal;
• acquisition, build, integration, and data migration;
• assurance, security, accessibility, and compliance;
• infrastructure, suppliers, and licences;
• operation, support, incidents, and on-call burden;
• change, compatibility, and consumer migration;
• coordination, training, and documentation;
• parallel running and transition;
• retirement, archival, and exit; and
• displaced work at the active constraint.

GAO's Cost Estimating and Assessment Guide requires purpose, scope, a technical baseline, work breakdown, assumptions, data, sensitivity and risk analysis, documentation, and updates with actual costs. It expressly covers software systems while noting that its practices were developed largely for major acquisitions.C17-R06 NASA's Cost Estimating Handbook adds a first-party technical treatment of lifecycle estimating and cost risk in a space-program context.C17-R07 Their authorized role here is disciplined estimation, not a claim that heavyweight acquisition control should govern every software decision.

Proportionality matters. A reversible two-week investigation does not need a 476-page estimate. A multi-year, hard-to-exit identity migration needs more than a one-line effort guess. The record should be strong enough for the consequence and uncertainty of the decision.

Lifecycle cost also prevents local savings from masquerading as system savings. A central team may reduce its workload by shifting configuration, support, or failure recovery to every Production Line. A procurement may lower build cost while increasing migration and exit exposure. An automated control may save reviewer time while adding remediation burden to producers.

Costs should therefore retain their owner, period, unit, and confidence. Totals can be useful, but the distribution must remain inspectable.

C17.4 — Value discovery and reversibility explicitly

Some work creates an operated outcome. Some work buys information. These are different option structures.

A discovery option is defensible when uncertainty can change a named future decision. It should declare:

• the uncertainty being reduced;
• the evidence to be produced;
• the smallest credible method;
• the decision that will use the result;
• the option's cost and expiry;
• the stop condition; and
• what will happen if the result is ambiguous.

“Run a pilot” is not enough. A pilot without a decision, comparator, evidence threshold, or stopping rule can become a quiet commitment. Chapter 20 will examine pilot design in detail. Here the economics boundary is simple: learning has option value only if it preserves or improves a consequential choice.

Reversibility changes what evidence is proportionate. A small, isolated, rapidly reversible change can be staged with a short review horizon. A choice that creates data lock-in, shared contracts, irreversible migration, safety exposure, or widespread consumer dependency requires stronger evidence and exit design.

Real-options methods can help describe the value of waiting, staging, expanding, or abandoning under uncertainty. They can also introduce spurious accuracy when probabilities and payoffs are weak. HM Treasury's current guidance explicitly notes that risk techniques requiring scenario probabilities can create such precision.C17-R05 The chapter therefore uses option language as a decision prompt, not a required financial model.

The practical questions are:

• What choice remains available after this step?
• Which choices disappear?
• What is the price of preserving the option?
• When does the option expire?
• Which evidence justifies expansion, revision, or abandonment?

C17.5 — Treat Factory Assets as two-sided investments

A Factory Asset is reusable production capability: a paved path, component, test harness, deployment mechanism, policy implementation, data product, observability package, or knowledge asset used across Work Centers or Production Lines.

An asset competes with immediate demand. Calling it “strategic” does not remove that opportunity cost. Nor does reuse guarantee return. Reuse research has reported potential quality, productivity, and economic effects while also identifying the costs and conditions required to realize them.C17-A15

Evaluate the producer side:

• creation and product-management capacity;
• assurance, documentation, support, and operation;
• compatibility and evolution;
• demand discovery and consumer enablement;
• retirement and migration obligations; and
• work displaced at the producing constraint.

Evaluate the consumer side:

• integration and migration effort;
• learning and local adaptation;
• fit to the consumer's operating conditions;
• dependency, waiting, and exception burden;
• avoided local creation or maintenance; and
• exit cost if the asset ceases to fit.

Then compare alternatives: local implementations, an external service, a shared asset, a temporary bridge, or stopping the demand. Record expected consumers and adoption evidence, but do not turn forecast adoption into realized return.

There is no authorized universal break-even point in this chapter. A local analysis may estimate one when costs, consumers, horizons, and alternatives are credible. The estimate must remain a forecast with ranges and switching conditions. Later outcomes should be compared with it.

Separate records, separate roles

The evidence cases answer different questions.

GAO's 2025 agile portfolio report synthesizes practices across selected leading companies, including recurring reassessment, iterative business cases, user evidence, and staged allocation.C17-R01 It supports the legitimacy of revisiting investment. It does not expose a complete comparable portfolio, prescribe a six-month cadence for every organization, or prove a causal performance effect.

The UK National Audit Office's programme-reset synthesis shows that continuing, resetting, replacing, or stopping should be reconsidered when affordability, benefits, risks, delivery confidence, and wider priorities change.C17-R02 It concerns major public programmes, not continuous software capacity allocation.

GDS's decision to decommission GOV.UK Platform as a Service is a first-party platform-sunset record. GDS described a changed market, alternative capabilities, and the opportunity cost of continued team and financial investment, then announced a transition window.C17-R03 The record does not contain a full counterfactual net present value, complete lifecycle cost, forecast error, or independent causal evaluation.

These records are not pooled. Together they show that reassessment, staging, reset, and sunset are legitimate parts of portfolio choice. They do not show that one method produces an optimal portfolio.

Evidence Status — Bounded. Current evidence supports explicit options, proportional appraisal, lifecycle estimation, uncertainty treatment, and the legitimacy of stop/reset/sunset decisions. It does not establish a universal formula, allocation cadence, Factory Asset return, or causal improvement from the proposed canvas.

C17.6 — Preserve the forecast and learn from error

An allocation decision produces a forecast whether or not it is written down. Teams expect cost, time, adoption, risk reduction, learning, or operational consequence. If the forecast disappears after approval, the organization cannot distinguish a surprising outcome from a rewritten memory.

Preserve:

• the options considered and excluded;
• the selected option and decision owner;
• baseline conditions and business-as-usual;
• estimates, ranges, sources, and confidence;
• critical assumptions and dependencies;
• expected monetized and non-monetized effects;
• affected groups and unresolved distributional questions;
• switching conditions and review triggers; and
• dissent, challenge, and evidence gaps.

At review, compare the forecast with observed outcomes. Record differences in cost, time, scope, adoption, service effect, risk, burden, and displaced work where evidence permits. Do not classify every difference as failure. Conditions may change; new information may justify a different path. The purpose is to improve the reference class and expose systematic optimism, not to punish uncertainty.

HM Treasury's optimism-bias guidance asks public bodies to use historical forecast error where available and to adjust estimates when costs, duration, or benefits have been systematically optimistic.C17-R05 GAO likewise calls for estimates to be updated with actual costs.C17-R06 These controls support organizational memory. They do not prove that recording error eliminates bias.

A forecast-error register belongs to the portfolio, not only to individual projects. It can reveal recurrent omissions: migration, consumer enablement, assurance, coordination, retirement, operational burden, or adoption. Chapter 16's measurement discipline applies: definitions, units, windows, distributions, and authorized uses must remain explicit.

The production economics canvas

For a consequential allocation decision, complete the following record:

1. What is the decision, owner, constraint, horizon, and review date?
2. Which outcome, obligation, or exposure motivates it?
3. What are the viable alternatives, including business-as-usual and stopping?
4. What will each option delay or displace at the constraint?
5. What are the lifecycle costs, owners, periods, ranges, and confidence?
6. What risks and uncertainties remain, and which assumptions can switch the choice?
7. What discovery is being purchased, which decision will use it, and when does the option expire?
8. Which choices remain reversible, and which create lock-in or transition obligations?
9. What reusable capability is created, and what producer and consumer burdens accompany it?
10. Which effects are monetized, quantified but not monetized, or qualitative?
11. Who receives benefits, bears costs or risks, and remains absent from the evidence?
12. Will the decision be choose, stage, defer, stop, or investigate?
13. Which forecast and rationale will be preserved for outcome review?

The canvas is a prompt sequence, not a scoring model. Options do not need to “win” every dimension. Some obligations are constraints, not benefits to be traded away. Some evidence remains uncertain. The decision record should show where authority and judgement entered.

Open figure

Figure F17.1 production specification: Production Economics and Capacity-Allocation Canvas

Figure F17.1 — A named allocation decision moves through explicit alternatives, distinct economic and public-value prompts, uncertainty and switching conditions, an authorized choice, and later forecast–outcome review. The canvas is not an additive score or valuation formula.

What to remember

Scarce capacity makes every accepted item an economic choice.

Begin with the decision, constraint, and alternatives—not the favored initiative.

Include business-as-usual and stopping where viable.

Estimate delay and lifecycle cost locally, with ranges and confidence.

Treat discovery as an expiring option tied to a future decision.

Evaluate Factory Assets from both producer and consumer perspectives.

Keep monetized, non-monetized, and distributional evidence distinguishable.

Preserve forecasts and compare them with outcomes.

Use economics to expose judgement—not to replace it with a spreadsheet.