Pillar detail

Workflow design

Good workflow writing covers prompt libraries, evaluation sets, human review gates, and how to measure quality—not just throughput. We care about failure modes: when a confident wrong answer costs more than a slow right one, and when “automation coverage” is inflated by cherry-picked pilot data. The hardest part is not writing the first prompt; it is keeping the workflow honest when product copy, pricing, and integrations change underneath you.

Evaluation sets that match production

A pilot built on tidy examples will outperform reality until the first messy ticket arrives. Stratify test cases by frequency and difficulty: include “ugly but common” inputs, multilingual fragments, and partial records. Tie each case to an expected outcome and a tolerance for human override. This discipline connects directly to when automation ROI is a mirage, where we separate vanity throughput from rework and escalation load.

Human review as a designed step—not an apology

Some categories should never be fully automated at your maturity level: high-stakes refunds, legal escalations, or anything involving regulated data. Design the handoff so reviewers see structured context, not a wall of model output. If reviewers spend time re-deriving facts the system already had, you have automated the wrong layer. Agencies often feel this first because client trust is part of the deliverable.

Ownership beats cleverness

The sustainable workflow is the one your team will maintain. Read documentation versus marginal accuracy and align incentives: who updates prompts after a product rename, who owns the evaluation set, and how you detect drift when model behavior shifts without a version bump. The seven-layer framework ends on team literacy for exactly this reason.

Versioning prompts like you version code

Informal prompt editing in shared chats is how production behavior drifts without a changelog. Treat prompt templates as artifacts: named versions, short rationale for changes, and a link to the evaluation cases that must still pass. You do not need heavyweight tooling on day one—you need a habit. When someone says “I tweaked the wording,” the follow-up question is: which cases did you re-run, and who approved the regression risk? That habit connects directly to operations ROI: silent quality drift shows up as rework long before it shows up on a leaderboard.

Cadence: weekly triage beats heroic quarterly audits

Evaluation sets rot because products and customers change. A lightweight weekly triage—five new failures pulled from real traffic, five old cases retired when they no longer reflect reality—keeps the set honest without turning your team into full-time benchmark curators. Rotate who leads triage so the workflow does not depend on one person’s taste. If you cannot spare thirty minutes a week, you are understaffed for automation, not “too busy for process.”

When to split workflows by customer segment

One global workflow is seductive until enterprise customers need stricter redaction, or SMBs need faster defaults. Splitting workflows is a product decision: different evaluation sets, different review gates, possibly different models or routes. The failure mode is invisible fork—two teams maintain slightly different prompts without naming the divergence. If you split, name the segments, document the boundaries, and measure quality separately. Otherwise you will blend metrics and misread regressions. Agencies often formalize splits early because client tiers are contractual, not cosmetic.

Where this pillar meets the rest

Workflow without security is reckless; workflow without stack coherence is redundant. Continue to procurement & security, operations ROI, and stack coherence when you move from design to operating reality.