Würk builds workforce management and HR technology for highly regulated industries, including cannabis, and an AI workflow we created is called the "Job Title Classifier". Titles like "Full-time Budtender" or "Bud Tender I" must map to one canonical title before they can appear in our benchmarking reports. Titles that didn't match were simply missing from what clients saw. This pattern works for any business with a product database, a taxonomy, and a normalization problem.

Here is how it runs:

1. Each night, a SQL query compares active employee job titles against our approved taxonomy and lists every unmatched title, sorted by how many employees share it, so the highest-impact titles get reviewed first.
2. Before any AI call, the system checks whether the title was already approved in a previous session. If it was, the existing mapping is reused instantly.
3. For titles with no prior approval, an AI agent receives the full taxonomy and returns a structured suggestion: canonical title, industry vertical, job code, seniority level, job category, and a confidence score with a plain-language reason.
4. Suggestions appear in a review queue, where the product owner approves, edits, or skips each one. Nothing writes to the database without human sign-off.
5. Every approval becomes a cache hit for future sessions. The same title from another client resolves instantly, so the AI gets called less over time.

We previously handled 30 to 40 titles per session in spreadsheets and AI chat tools, with no audit trail. Now we process hundreds per session, log every decision with its method, confidence score, and timestamp, and clients see more compensation data points.