Fixing what the parser got wrong before a person sees it
Reconciling parsed employment history against the source document and flagging the fields that disagree, rather than presenting parser output as fact.
The parser reads the resume, then somebody corrects the employment dates it merged, which is most of the time it was supposed to save.
Resumes arrive in every format that has ever existed, including scans with no text layer and designed templates in two columns that get read in the wrong order. Parsing is nearly always a third-party component, and its characteristic failure is employment history: dates written as ranges with words in them produce merged employers and wrong tenure.
Where a client mandates a vendor management portal, entry into it is manual by contract rather than by technical accident. The agency holds the candidate in its own system and types them again into the portal, then types timesheets into both, every week, for the length of the assignment.
Submission records are legally significant here in a way outsiders miss. When two agencies present the same candidate, the acknowledgement of representation and the timestamps decide the fee, which makes the integrity of that record more important than its convenience.
Ranges written with abbreviations or the word current produce wrong tenure, merged roles, dropped positions or a title filed under the wrong employer. Any ranking that trusts parsed tenure inherits the error and presents it with confidence.
Entry into a vendor management system is a contractual obligation with no interface offered to the agency, so the duplicate typing cannot be engineered away from the agency side.
The same person arrives from a job board and from a referral, and again as a direct application, and matching on name or address is unreliable enough that the duplicates survive.
Job board interfaces are rate limited and restricted by terms, and scraping the large professional network violates its terms and ends accounts. This is a hard boundary on the obvious idea.
Reconciling parsed employment history against the source document and flagging the fields that disagree, rather than presenting parser output as fact.
Chasing the client for feedback on candidates already submitted and keeping the record in one place, given that the record decides fee disputes.
Weekly and per contractor, without end.
A requisition posted in SAP Fieldglass reaches a supplier as a notification, and the distribution tier the agency sits in decides how long it holds the posting before a second tier is released into it.
Fieldglass refuses a job seeker outright once a supplier has used the submission slots allowed on that posting, and the error names the limit rather than the fix.
The same person arrives as a board application, then as a referral, then again when a second recruiter uploads a document that came in by email.
Free busy lookup stops at the boundary of the tenant, so the calendar of a hiring manager is invisible to a supplier and availability comes back as prose in an email.
A compliance file for a clinical assignment is a pile of scans.
A pre adverse notice is a packet rather than a letter.
Checkr fires an invitation expired event when a candidate never finishes the form, and most integrations listen for a completed report and nothing else.
A start packet is built from parts that come from different places.
Reverification lives on Supplement B of Form I-9, which replaced the old third section, and it is triggered by a date on a document rather than by an event anyone announces.
A work order confirms a rate, a start date, an approver and a rule for overtime.
Where automatic invoicing is switched on against the supplier record, the portal raises the invoice and the agency is left matching a document it did not write against hours it has already paid out.
The end date on a placement record and the end date on a work order are two different facts, and only one of them moves when a client extends.
An invoice goes out when the person starts, and the agreement usually hangs two conditions on it.
Ranking or scoring candidates brings a file with it.
This is the sector where the boundary is sharpest, and it is worth stating plainly: rejection is the regulated act, not enrichment. Parsing and deduplication carry little exposure, and so do formatting, scheduling and the chasing of timesheets. Ranking, scoring, filtering or automatically rejecting candidates engages selection-procedure rules on adverse impact, and several jurisdictions now impose specific duties on automated employment decision tools, including bias auditing with published results and advance notice to candidates in one city. Another state requires notice and prohibits using postal code as a proxy for a protected class, and recordkeeping obligations covering the system and the selection criteria in another. Adjudicating a background report separately triggers a required sequence of pre-adverse and final notices with a copy of the report. Those steps need a named human decision maker and a retained record.
Related: Accounting and bookkeeping Law firms. Or start from the four shapes of work instead of from an industry.