Cost per decision

Ask a team what a cognitive feature costs and you usually get a monthly total from a billing dashboard. A real number attached to nothing in particular. Ask which part accounts for most of that total and the shrug becomes universal, because the question has no answer when parts do not have names.

Once they do, arithmetic becomes available. This page applies two budgets and the Cognitive Units to per-edge economics. Evaluation as engineering supplies the evidence wrappers must move. Guards from deterministic guards before cognition change volume before price attaches.

Helpful context: Feature cost versus decision cost holds monthly and per-edge lenses together. Paying for reliability covers deliberate wrapper spend on consequential edges.

What a call costs

A named cognitive units is one call, or a declared composition of calls with the same contract shape. Its cost can be computed rather than only observed after the fact. Model tiers differ by orders of magnitude. Exact prices move with providers. The shape does not: small, mid, and large sit roughly a decade apart, and the token footprint is knowable from template and inputs once someone documents them.

Wrappers multiply the base: sampling, critics, cascades, and retries draw on both budgets from two budgets. Latency, human review, and the cost of failure belong in the same account when the decision is consequential. A cheap call that fails open on allergy may cost more in support and trust than a double-run judge on household fit, because the failure externalizes cost that the provider bill never shows.

Cost component	Attaches to
Model call	Named cognitive units edge
Wrappers	Declared reliability stack on that edge
Human review	Residual cases after guards
Failure	Consequence of a wrong judgment

A spreadsheet you can replicate

For the Meal Companion on one busy week (illustrative, not prescriptive):

• GenerateCandidates times 7 evenings times about $0.02 = about $0.14
• AssessMealPracticality times 7 times about $0.008 = about $0.056
• ComposeWeek times 2 loop iterations times about $0.03 = about $0.06
• CritiquePlan times 2 times about $0.015 = about $0.03

Total per plan: roughly $0.29 before wrappers. Triple-sample JudgeHouseholdFit on a high-consequence path adds about $0.045 per invocation. Two loop iterations because critique failed once doubles the compose and critique lines.

The arithmetic is simple. The discipline is naming edges so the spreadsheet exists at all. Without names, optimisation becomes taste, prestige, or guesswork. With names, every conversation about spend has a referent that engineering, product, and finance can point to.

Full populationGuardsOpen judgmentsWrappers where justified

Cost follows volume after guards, not prestige alone.

Cost follows volume

In a feature assembled from several cognitive units, not every edge runs on every case. Deterministic guards filter the population before expensive judgment runs. Halving a token rate on a rare path may save little. Raising autonomy so fewer cases need human review may save orders of magnitude more.

The cognitive unit everyone demos (JudgeHouseholdFit) may be a tenth of the bill. GenerateCandidates running on every evening every loop iteration may dominate. Cost follows volume, and volume is highest at the front of a funnel where work is comparatively mechanical. This is why guard placement and closed-work extraction, which operate on high-volume edges, typically dwarf per-call optimisation on prestige edges when measured in monthly totals.

One call means one price, and prices sum. That is what makes design decisions arguable.

The cognitive unit, Ch. 12

Wrappers change the line item

Tuning by wrapping adds combinators that preserve contract while changing spend. Sampling three times on AssessMealPracticality triples that line. The spreadsheet shows wrapper policy per edge so finance can see reliability spend without reading templates.

Remeasuring suite movement when wrappers change is what separates deliberate reliability investment from theatre. Paying for reliability without measured return means the wrapper exists for comfort rather than for the outcome it was supposed to produce. Every wrapper that lacks before-and-after evidence on the same regression set is a candidate for retirement in the next quarterly review.

Failure cost belongs in the account

Wrong allergy enforcement is not a $0.008 mistake. It is trust collapse, support load, and possible harm. Consequential edges deserve higher wrapper spend and lower autonomy until eval coverage justifies expansion. The spreadsheet is incomplete if it lists tokens but omits failure cost on the same row, because the row then misleads optimisation toward cheapness when the true cost of failure would justify higher spend.

This matters for the invoice intake parallel as well. A field interpretation cognitive unit that misreads a vendor amount on a high-value invoice carries consequence far beyond its token cost. The account must reflect that consequence or the architecture will optimise the wrong line.

Sensitivity analysis

Changing one variable at a time on the spreadsheet reveals which knob moves the monthly total most: loop iterations, evenings per plan, wrapper multiplier, guard clearance rate. Often loop depth and population before guards dominate. Token rate on a rear-edge judge rarely moves the feature total enough to justify weeks of prompt tuning.

Sharing sensitivity results in product review means "use the bigger model" arguments attach to line items rather than to comfort. Without that discipline, prestige models spread because they sound safer, not because arithmetic supports them. A sensitivity table that shows loop depth moving the monthly total ten times more than a model tier upgrade on a rare edge changes the conversation from opinion to evidence.

Composition and declared stacks

When one cognitive unit composes several internal calls, cost per decision still attaches to the contract edge the caller invokes. Internal composition should be documented so feature cost spreadsheets do not undercount. A single ComposeWeek call that internally cascades three models is one price only if the contract declares that stack. Hidden internal calls are how bills surprise finance, because the spreadsheet says one call while the provider bill says three.

Recording cost per decision in the architecture artifact means each named edge on the decision map carries a default model tier, wrapper stack, last measured volume after guards, and illustrative cost per call. Updates happen when substitution or wrap changes, so the artifact becomes the spreadsheet source of truth for feature reviews. New engineers reading cost rows before proposing "just add another critique pass" can see whether another pass is arithmetic or negligence.

Shadow pricing before cutover

During adoption of cognitive units in existing code, shadow mode lets teams price a new edge against legacy volume before cutover. Comparing call counts and per-call prices on the same traffic shape turns cutover decisions into arithmetic rather than hope. The shadow period typically reveals whether the new cognitive unit's per-call cost multiplied by actual volume produces a feature total the team can defend in review.

Portfolio roll-up

Rolling up per-edge spreadsheets into a portfolio view matters when multiple features share cognitive units. Meal Companion and grocery copilot may share AskTargetedQuestion from the organisational library. Shared edge spend appears once in portfolio roll-up with attribution tags per product line. Duplicate counting hides reuse wins that finance should celebrate, because library reuse means the eval investment pays back across products rather than staying siloed in one feature's budget.

Decision cost in autonomy debates

Autonomy debates need per-decision cost, not monthly totals alone. A cheap monthly bill may hide expensive per-case judges on exception paths that dominate clerk time. Decision cost including wrapper retries, human review triggers, and eval runs triggered by abstention reveals the true price of the current autonomy level.

Comparing decision cost before and after guard placement shows whether effective guards reduced volume on expensive edges without hiding structural need for the cognitive unit. The comparison is how teams argue for or against autonomy expansion with numbers rather than with intuition about readiness.

What to do next

Building one spreadsheet for a single feature with named edges and volumes is the entry point. Identifying the highest-volume line item and asking whether guards or closed-work extraction should shrink it usually reveals the largest cost lever. Attaching wrapper policy to consequential edges with suite thresholds, and sharing the monthly feature total alongside the per-edge breakdown in the same review, creates the alignment between engineering, product, and finance that makes spend decisions arguable.

See deterministic guards before cognition, paying for reliability, and tuning by wrapping.

Read next: Feature cost versus decision cost.