The NFR Budget

Overview

You now have five families of requirement, each with numbers. Written separately, they look satisfiable. Put side by side, several of them are in direct competition, and every one of them is bounded by constraints that were never negotiable.

This topic assembles them into one artifact and makes you choose a corner.

Why a budget, not a list

A list of non-functional requirements implies that all of them can hold at once. A budget implies a finite resource being allocated — which is the accurate picture, and it changes the conversation from “is this requirement reasonable?” to “what are you taking it from?”

The resource being allocated is not money alone. It is the single interaction: how much context you can afford to retrieve, how many checks can sit in the path, how large a model can answer, how long the user will wait. Every requirement spends the same interaction.

The three-way pull

Accuracy, cost and latency form a triangle, and the mechanism of each tension is worth stating because it tells you which lever to reach for.

You do not get three corners. You pick an operating point, and the dashed lines are the distances you are accepting — the short one to accuracy is the choice this budget makes.

Reading the triangle in both directions:

• Toward accuracy — retrieve more context, use a larger model, add a verification pass, retry on low confidence. Each raises cost per interaction and adds latency.
• Toward cost — route to a cheap model first, trim context, cache aggressively. Each gives up accuracy on some slice of requests, and caching gives up freshness.
• Toward latency — fewer checks, smaller context, no second pass. Accuracy pays for all three.

Safety sits outside the triangle but spends from the same interaction: every guardrail call adds latency and cost, and a stricter filter trades usability for assurance.

Constraints bound the triangle before you start

A constraint is not a point in this space; it shrinks the space. Data residency may remove the models you would otherwise have chosen, which lowers the accuracy ceiling and raises the cost floor simultaneously — before you have made a single trade-off.

This is why the three kinds matter. Treating a constraint as a requirement invites someone to negotiate it, and the negotiation is always with the wrong person. The budget records constraints as the boundary of the page, not as rows to be traded.

The column that makes a budget a budget

A normal NFR table has target, measure and owner. Add a fourth column: the give.

For every line, what yields when the system cannot satisfy everything at once? Which requirement degrades, by how much, and who was told in advance?

This is uncomfortable to write and it is the whole point. Under pressure, something always gives. The only question is whether it was chosen deliberately in month one or by an engineer at 23:00 in month nine, in favour of whichever requirement their ticket happened to mention.

Then state the ordering once, in a sentence: when these collide, this one wins. An ordering everyone has seen is worth more than a set of individually reasonable numbers nobody has ranked.

The budget, with the ordering made explicit

Requirement	Target	Owner	The give
Correct on golden set	≥ 90% of 300 questions	Maya Ortiz	Does not give. Below this the project repeats 2023
Confidently wrong answers	≤ 1% of answered	Maya Ortiz	Does not give
Unnecessary handoffs	≤ 15%	Grace Adeyemi	Gives first — rises to 25% before accuracy is touched
p95 time to first token	≤ 1.5 s	Maya Ortiz	Gives to 3 s if the alternative is a weaker check
p95 answer completion	≤ 8 s	Maya Ortiz	Gives to 12 s
Cost per deflected interaction	≤ $1.54 p50, ≤ $2.50 p95	Dev Raghunathan	p95 gives before p50; monthly ceiling never gives
Monthly inference ceiling	$8,000, degrade at 80%	Dev Raghunathan	Does not give — degrade to a cheaper tier instead
Refusal rate on legitimate requests	≤ 5%	Grace Adeyemi	Gives to 8% during the first month of tuning

Constraints bounding all of the above — not negotiable, not rows in this table: no regulated customer data leaves its region; SOC 2 Type II controls apply; personal data is redacted before embedding; the paid Copilot tier has a two-quarter commitment that changes scope rather than any number here.

The ordering, stated once: accuracy first, cost second, latency third. When they collide, deflection volume and response speed give way before answer quality does.

That ordering is a judgement, and it is arguable — Priya Shah would rank the deadline first, and she is not wrong about her own driver. What makes it defensible is the 2023 routing feature: an organisation that has already watched one quietly-wrong system get switched off will not fund a second. The cheapest way to lose the whole initiative is to hit every number except correctness. Note also that the first two rows and the monthly ceiling are the lines held absolutely — the budget is only credible because most of the other rows do give.

What this costs

Three real costs, in order of how much they sting.

Naming the give pre-authorises a sacrifice. Someone will later point at the row and take the degradation you wrote down. That is the system working, and it will not feel like it.

The ordering makes someone lose in advance. Whoever holds the requirement at the bottom now knows they lose. Better that they know now, from you, than discover it in a release meeting.

The budget will be wrong. The numbers rest on assumptions no one has measured yet. Date it, name the assumptions, and revisit after Module 3 picks a structure and again after the first real measurements. A budget revised with evidence builds credibility; a budget silently abandoned destroys it.

Feeds the workbook

This is the NFR budget artifact, and it is the one Module 3 scores its style decision against. Include the give column and the ordering sentence — without them you have written a list.

Check your understanding

1. Write the give column for your own budget. Which row did you least want to write, and why is that the most important one?
2. Your ordering puts cost above latency. Which stakeholder loses, and what do you say to them before they find out?
3. A constraint appears in your table as a row with a target. What goes wrong, and where should it have gone instead?

---

→ Ten topics done. The requirements are testable and the conflicts are named. — Module 3 — Choosing a Structure

Optional workbook: Exercise 2 turns this module into artifacts. The exercises build on each other, so work them in order or not at all.