AI Readiness Scorecard

Data sits across disparate docs, PDFs, emails, or has not been collected yet.

Clean Postgres/MySQL databases and standard SaaS APIs, but no dedicated vector/search index.

Structured tables + searchable knowledge base with clean schemas and metadata.

Automated ingestion, versioned datasets, and access control policies in place.

Using public AI web interfaces or default API endpoints without data retention agreements.

Using verified enterprise API endpoints (Anthropic / OpenAI / Gemini zero-retention).

PII stripping before prompt generation, per-tenant vector partitioning, and audit logging.

Testing by typing prompts by hand and judging if the response looks decent.

A curated list of 20-50 real questions/inputs tested before major releases.

Continuous evaluation on accuracy, latency, toxicity, and hallucinations on every build.

Single prompts sent directly to a single provider with minimal retry or fallback logic.

Vector search embeddings piped into an LLM context window.

Deterministically orchestrated agents, hybrid semantic+keyword search, and fallback routers.

No tracing for failed prompts, token usage per user, or latency spikes.

Logging prompt errors and tracking general API consumption in server logs.

Full trace graphs, per-feature token breakdown, latency heatmaps, and user feedback loops.

Engineers use ChatGPT for code snippets, but product architecture is 100% traditional.

Developers have built working AI features but lack formal guardrails, evals, and agent design patterns.

Team uses coding agents, rigorous prompt versioning, automated testing, and human-in-the-loop review.

Outputs go straight to customers without validation or safety guardrails.

AI generates drafts or suggestions; humans verify before execution or sending.

Deterministic schema verification, confidence gating, automated fallbacks, and human review queue.

Adding AI because everyone else is doing it, without a proven customer pain point.

Automating high-frequency tasks where users already spend hours doing manual work.

Unlocking unstructured data synthesis, autonomous workflows, or real-time intelligent agents.