You cannot search for what was never structured
Your database search works on fields. The records your team adds today probably have them: a discipline, a level, a sector, picked from a list. The records from five years ago have a job title somebody typed, "Snr FC (interim)" or "Head of Ops, EMEA", and nothing else. A search for senior finance people in manufacturing walks straight past them. They are in the database and invisible to it.
That is why this play sits in front of the others as a gate. Sourcing from your own ATS, scoring your candidates against a live role, deciding which segment to re-engage: all of them filter on structure first. Until the data is structured, they are accurate only for the part of the database that was tagged properly on the way in.
What you get back
A review list, as a file: record, field, current value, proposed value, confidence, and the words in the record the proposal rests on. Unclear records are listed separately with the reason. After you approve, the tags are in your CRM in the fields you named, and you get a change log by record id. Run a search you know was missing people and compare the count.
Variations worth knowing
File only. If your CRM cannot take the write, or you would rather import it yourself, the play stops at the approved file.
One dimension first. Discipline alone is the biggest gain and the easiest to check. Add seniority and sector once you agree with how it reads your records.
A clean title alongside the tags. The agent can propose a normalised job title in a field of its own. It never replaces the title the person gave you.
Refresh before you classify. If the segment is old, making the database searchable again brings titles and employers up to date first, and the tags describe who people are today.
Then match. Tagged records are what finding the candidates already in your ATS filters on before it scores anyone.
Where this goes wrong
A taxonomy that overlaps. If both Finance and Accounting are on your list, the agent has to guess which one you meant, and so does every consultant who tags by hand. Fix the list before the run.
Thin records. A title with no employer and no other text cannot be placed honestly. Expect a share of unclear, and treat it as information about the record.
A write that says it worked. JobAdder will accept a custom-field value in the wrong shape, report success and save nothing. The read-back on the first record exists to catch that before it happens across the whole segment.
Tagging people, not work. The taxonomy is about discipline, level and sector. The agent does not classify on age, gender, family status or any other protected attribute, even where a field exposes one.
