Requirements Gathering for AI

The requirements gathering for AI turns "we want AI" into the functional and the non-functional requirements (L359) — the functional (the features: what the system does — the chat L348, the RAG L349, the agents L200, L359), the non-functional (the qualities: the latency L333, the scale L358, the compliance L371, the cost L368, L359), and the AI's (the model L148, the data L313, the evals L341, L359) — the enterprise loop's (L380) start (L359).

The one-sentence interview answer: "The requirements gathering turns the wish into the requirements (L359). The functional (L359): the features (L359) — what the system does (L359): the chat (L348), the RAG (L349), the agents (L200) — the user stories (L359). The non-functional (L359): the qualities (L359) — the latency (L333): the TTFT (L145) under 2s (L359); the scale (L358): the requests per day (L359); the compliance (L371): the data's residency (L261) and the privacy (L313); and the cost (L368): the budget (L368). The AI's (L359): the model (L148) — the task's (L340) fit (L359); the data (L313) — the sources (L359) and the quality (L359); and the evals (L341) — how the success (L341) is measured (L359). The elicitation (L359): the questions (L359) — "what does the user do?", "what's the latency?", "what's the data?", "what's the quality?" (L359). The AI shape (L173): the "we want AI" (L359) → the requirements (L359) — the functional (L359), the non-functional (L359), and the AI's (L359) — the ADR's (L361) input (L359)."

Think of the requirements as the architect's brief for the client's dream house. The client (the business, L359) says "we want a house" (L359). The architect (the solutions architect, L359) extracts the brief (the requirements, L359): the rooms (the functional, L359) — the kitchen (the chat, L348), the library (the RAG, L349), the workshop (the agents, L200); the qualities (the non-functional, L359) — the light (the latency, L333), the size (the scale, L358), the codes (the compliance, L371); and the materials (the AI's, L359) — the contractor (the model, L148), the lumber (the data, L313), the inspections (the evals, L341). The client signs the brief (L359) before the blueprint (L361). The house works because the brief is complete, and the client signed it (L359).

   the brief (the requirements, L359)
   ┌────────────────────────────────────────────────────────┐
   │ the rooms (the functional, L359) — the features (L359) │
   │ the qualities (the non-functional, L359) — the         │
   │ latency (L333), the scale (L358), the compliance (L371)│
   │ the materials (the AI's, L359) — the model (L148), the │
   │ data (L313), the evals (L341)                          │
   └────────────────────────────────────────────────────────┘

The mental model is the brief: the rooms, the qualities, and the materials (L359).

   "we want AI" (L359)
        │  the elicitation (L359)
        ▼
   ┌────────────────────── THE FUNCTIONAL (L359) ───────────────────────┐
   │  the features (L359): the chat (L348), the RAG (L349), the        │
   │  agents (L200) — the user stories (L359)                          │
   └──────────────────────────┬───────────────────────────────────────┘
                              ▼
   ┌────────────────────── THE NON-FUNCTIONAL (L359) ───────────────────┐
   │  the latency (L333): the TTFT (L145) · the scale (L358)           │
   │  the compliance (L371) · the cost (L368)                          │
   └──────────────────────────┬───────────────────────────────────────┘
                              ▼
   ┌────────────────────── THE AI'S (L359) ─────────────────────────────┐
   │  the model (L148) · the data (L313) · the evals (L341)            │
   └──────────────────────────┬───────────────────────────────────────┘
                              ▼
   ┌────────────────────── THE ADR (L361) ──────────────────────────────┐
   │  the requirements (L359) → the architecture decision (L361)       │
   └──────────────────────────────────────────────────────────────────┘

The flow is the loop's start: wish → requirements → ADR (L359).

• The functional (L359). The features (L359): the chat (L348), the RAG (L349), the agents (L200) — the user stories (L359).
• The non-functional (L359). The qualities (L359): the latency (L333), the scale (L358), the compliance (L371), the cost (L368).
• The AI's (L359). The model (L148), the data (L313), the evals (L341).
• The elicitation (L359). The questions (L359): "what does the user do?", "what's the latency?", "what's the data?", "what's the quality?" (L359).

• An enterprise engagement (L379). The requirements (L359) — the case study's (L379) start (L359).
• A greenfield AI product (L173). The features (L359) — the chat (L348), the RAG (L349).
• A regulated enterprise (L371). The compliance (L371) — the data's residency (L261) — the non-functional (L359).
• A model rollout (L365). The evals (L341) — the success metric (L359) — the AI's (L359).
• Anything enterprise (L380). The brief (L359) — the loop's (L380) start (L359).

The through-line: the brief is the loop's start — the functional, the non-functional, and the AI's (L359).

Say it in four moves:

1. The functional. "The features — the chat, the RAG, the agents (L359)."
2. The non-functional. "The latency, the scale, the compliance (L359)."
3. The AI's. "The model, the data, the evals (L359)."
4. The elicitation. "The questions — the wish → the brief (L359)."

• The requirements are the design's half (L359). The brief (L359) — the wrong brief (L359) is the wrong architecture (L358) — the elicitation (L359) first (L359).
• The latency is the non-functional's first (L333). The TTFT (L145) — the L151 budget (L151) — the design's (L358) constraint (L359).
• The eval is the requirement (L341). The success metric (L341) — the groundedness (L337), the task success (L340) — defined up front (L359) — the L346 suite (L346) gating (L359).
• The compliance is the boundary (L371). The data's residency (L261) and the privacy (L313) — the L371 frameworks (L371) — the non-functional (L359).
• The cost is the budget's (L368). The tokens (L332) and the infra (L368) — the L368 estimate (L368) — the business (L360) signs (L359).

• The wish accepted (L359). "We want AI" (L359) without the brief (L359) — the requirements (L359) are the design's (L358) input (L359).
• The functional-only (L359). The features (L359) without the latency (L333) and the scale (L358) — the architecture (L358) unconstrained (L359).
• The eval-less (L341). The success (L341) undefined (L359) — the quality (L341) un-measurable (L359).
• The compliance after (L371). The residency (L261) bolted on (L371) — the requirement (L359) up front (L359).
• The un-signed brief (L359). The requirements (L359) un-agreed (L360) — the scope (L359) creeps (L359).

• Elicit the three (L359) — the functional, the non-functional, the AI's (L359).
• Define the evals (L341) — the success metric (L359) up front (L359).
• Bound the latency (L333) — the TTFT (L145) and the budget (L151).
• Ask the compliance (L371) — the residency (L261), the privacy (L313).
• Sign the brief (L360) — the stakeholders (L360) agree (L359).

Q: Walk me through the requirements gathering for AI.

A: The wish → the requirements (L359). The functional — the features: the chat, the RAG, the agents (L359). The non-functional — the latency, the scale, the compliance (L359). And the AI's — the model, the data, the evals (L359).

Q: What are the AI-specific requirements?

A: Three (L359): the model (L148) — which task and which fit (L359); the data (L313) — the sources (L359) and the quality (L359); and the evals (L341) — the success metric (L341): the groundedness (L337), the task success (L340) — defined up front (L359). The eval (L341) is a requirement (L359), not an afterthought (L359).

Q: How do you turn "we want AI" into the requirements?

A: The questions (L359): "what does the user do?" — the functional (L359); "what's the latency and the scale?" — the non-functional (L333, L358); "what's the data and the compliance?" — the AI's and the boundary (L313, L371); and "what's the success metric?" — the eval (L341). The answers (L359) are the brief (L359).

Q: Why do the requirements matter?

A: The design's input (L359): the wrong requirements (L359) are the wrong architecture (L358) — the rework (L359) and the cost (L368). The brief (L359) — signed by the stakeholders (L360) — is the ADR's (L361) input (L359).

• What's the functional (L359)?
• What's the non-functional (L359)?
• What are the AI's requirements (L359)?
• How do you elicit (L359)?
• What's the eval's role (L341)?

The senior read: the AI's column is the difference — the model, the data, the evals (L359).

// The requirements (L359) — the brief for the ADR (L361).
// 1 · THE FUNCTIONAL (L359) — the features (L359).
const functional = [
  { id: 'F1', story: 'As a user, I can chat with the assistant' },      // L348
  { id: 'F2', story: 'As a user, I can ask about the docs' },           // L349
  { id: 'F3', story: 'As an admin, I can review the refunds' },         // L350
];

// 2 · THE NON-FUNCTIONAL (L359) — the qualities (L359).
const nonFunctional = {
  latency: { ttft: '< 2s', p95: '< 4s' },         // L333, L145
  scale:   { requestsPerDay: 1_000_000 },          // L358
  compliance: { residency: 'eu', privacy: 'gdpr' },// L371, L261
  cost:    { monthlyBudget: 50_000 },              // L368
};

// 3 · THE AI'S (L359) — the model, the data, the evals (L359).
const aiRequirements = {
  model: { task: 'chat + grounded answers', candidates: ['claude', 'gpt'] },  // L148
  data:  { sources: ['help-center'], quality: 'reviewed' },   // L313, L349
  evals: { groundedness: '>= 0.9', taskSuccess: '>= 0.85' },  // L337, L341
};

// 4 · THE ADR (L361) — the brief → the decision (L361).

What the reader must SEE — the brief, written:

  F1-F3: the user stories     → the functional (L359)
  ttft < 2s, 1M/day, gdpr    → the non-functional (L333, L358, L371)
  the model candidates        → the AI's (L148, L359)
  groundedness >= 0.9         → the evals (L341, L337)
  → the ADR (L361)            → the loop's next (L361)

  The functional, the non-functional, and the AI's (L359).

• The brief is the design's speed (L359). The right requirements (L359) — the rework (L359) avoided (L359).
• The latency is the design's constraint (L333). The TTFT (L145) — the L151 budget (L151) — the architecture (L358) follows (L359).
• The scale is the capacity's (L358). The requests (L358) — the L369 capacity (L369) sized (L359).
• The eval is the quality's gate (L341). The thresholds (L341) — the L346 suite (L346) gating (L359).

• The requirements gathering = the loop's start (L359): the functional, the non-functional, the AI's.
• The functional: the features (L359) — the chat (L348), the RAG (L349), the agents (L200).
• The non-functional: the latency (L333), the scale (L358), the compliance (L371), the cost (L368).
• The AI's: the model (L148), the data (L313), the evals (L341).
• The elicitation: the questions (L359) — the wish → the brief (L359).

REQUIREMENTS GATHERING FOR AI = the wish → the brief

THE FUNCTIONAL (L359)
  the features (L359): the chat (L348), the RAG (L349),
  the agents (L200) — the user stories (L359)

THE NON-FUNCTIONAL (L359)
  the latency (L333) — the TTFT (L145) under the budget (L151)
  the scale (L358) — the requests per day (L359)
  the compliance (L371) — the residency (L261), the privacy (L313)
  the cost (L368) — the budget (L368)

THE AI'S (L359)
  the model (L148) — the task's (L340) fit (L359)
  the data (L313) — the sources (L359), the quality (L359)
  the evals (L341) — the success metric (L359):
  the groundedness (L337), the task success (L340)

THE ELICITATION (L359)
  "what does the user do?" · "what's the latency?"
  "what's the data?" · "what's the quality?"
  → the ADR (L361)

INTERVIEW, 4 MOVES
  1 functional "the features (L359)"
  2 non-functional "the latency, the scale, the compliance (L359)"
  3 the AI's  "the model, the data, the evals (L359)"
  4 elicitation "the questions (L359)"

You now hold the brief: the functional, the non-functional, and the AI's — with the evals defined and the compliance bounded. The client's dream is a signed brief — and the blueprint comes next (L359).

Next: the executives, the engineers, and the risk of AI theatre — Stakeholder Communication (L360).