10 things you must know before working with AI

“Add AI” is not a requirement. Name the person, the decision or task, the current cost, and the acceptable outcome. A narrow workflow with a measurable result is easier to evaluate than a general assistant expected to know everything.

Different tasks need different interfaces, models, retrieval systems, or deterministic software. Classification may need a small model; a high-stakes workflow may need structured rules and human approval. Do not pay for maximum capability when consistency is the real requirement.

Know whether prompts may contain personal, confidential, regulated, copyrighted, or customer-owned information. Review provider retention and training terms, minimize what you send, and keep secrets out of prompts. Data governance applies even when the interface feels conversational.

Models can produce fluent statements that are unsupported or wrong. Design for that reality with retrieval from trusted sources, citations, validation rules, confidence thresholds, and escalation. Never confuse polished language with verified truth.

Human review is not a decorative approval button. The reviewer needs the source context, authority to reject, and enough time to notice a problem. Increase oversight when output affects health, employment, benefits, safety, legal rights, or significant money.

An AI system that reads external content or calls tools can be manipulated through prompt injection. Limit permissions, isolate untrusted content, validate tool arguments, require confirmation for consequential actions, and log what happened. Follow the principle of least privilege.

Create representative test cases, expected characteristics, unacceptable failures, and a scoring process. Include ordinary cases, edge cases, adversarial inputs, and examples from real users. Re-run evaluations when prompts, models, data, or tools change.

Token cost is only one line item. Include retrieval, storage, observability, retries, human review, support, evaluation, and engineering maintenance. Set usage limits and track cost per successful outcome rather than cost per request alone.

Document who controls prompts, evaluations, source code, customer data, embeddings, domain, model accounts, and workflow history. Use portable formats and clear export paths where continuity matters. A provider can be replaceable only if your essential knowledge is not trapped inside it.

AI behavior can shift when models, policies, source data, or user behavior change. Assign an owner, monitor failures, gather feedback, review access, update evaluations, and maintain a fallback. Launch begins the operating phase; it does not end the project.

Before funding a build, require a one-page brief covering the outcome, users, data classification, risk level, evaluation, human review, permissions, cost ceiling, ownership, and operating owner. If the team cannot answer those questions, the next step is discovery—not development.