Every recruiting product now says it uses AI. Some of that is real and useful, some is a rebranded keyword filter, and some is a demo that will not survive contact with a messy resume. We have spent the past two years deciding where to use it in Apex ATS and, just as importantly, where not to. This is a plain account of that thinking, what we shipped, and what we deliberately did not.
What AI is genuinely good at in this job
The useful cases share a pattern: unstructured input that a person would otherwise have to read and retype, or a first pass over a large volume that a person will then review. In both cases the model does the tedious part and the recruiter stays in charge of the decision.
- Parsing resumes into structured records. Contact details, employers with dates, education, skills. Modern models handle odd layouts, two column designs and PDF exports from phone apps far better than the template-based parsers of five years ago.
- Matching a resume against a job's requirements and explaining the match. Not a score by itself, but 'has 3 years of forklift experience, certified, no weekend availability listed', which is what a recruiter actually wants to know.
- Drafting a first version of a message or a job description that a person edits. The draft is rarely final. It is a faster starting point.
- Summarizing a long thread so a manager who is stepping in can catch up in twenty seconds.
What is mostly hype
Fully automated hiring decisions. Personality inference from writing style. Video interview analysis that claims to read confidence or honesty. Chatbots that promise to replace the phone screen. Some of these are unproven, some are actively risky under the employment laws now in force in a growing number of states and countries, and all of them share a problem: they put a model between a candidate and a decision without a person accountable for it. We think that is bad for candidates and bad for the employers who will have to explain those decisions later.
There is a quieter category of hype too: features that call themselves AI and are actually a keyword count. Keyword scanning is useful. We ship it. But it is not intelligence, and a vendor who dresses it up as such is telling you something about how they will describe everything else.
What we shipped
Resume parsing was the first and is still the one customers notice most. Upload a resume, or have a candidate upload one from a phone, and the record is populated: work history with dates, education, skills, contact details. Parsing accuracy on structured fields sits above 95 percent in our testing on real customer resumes, and the candidate or recruiter can correct the rest in a confirm step. The practical effect is application forms that are three fields long.
Requirement matching came next. For each job you define requirements: must-haves, nice-to-haves, and knockouts. Every applicant is compared against them and ranked, with a short plain-language explanation next to the rank. The explanation is the feature. A rank of 4 out of 40 with 'meets all must-haves, missing CDL Class A, 6 years relevant experience' lets a recruiter scan a pipeline in minutes and disagree where they should. Ranks are never used to auto-reject. A person moves the stage.
- 01Message drafting: a suggested first draft inside the messaging composer, in your saved template's tone, that you edit before sending.
- 02Job description drafts: from a title, location and a few bullet points, with pay range fields left blank on purpose so nobody publishes a placeholder.
- 03Thread summaries in the shared inbox for conversations longer than ten messages.
- 04Duplicate detection: flagging when a new applicant looks like an existing record, with the evidence shown so a person confirms the merge.
What we did not ship, and why
We do not auto-reject. We do not score candidates on anything other than the requirements you wrote for the job. We do not analyze video or voice. We do not infer protected characteristics, and we log every ranking with the version of the requirements it was based on so you can show your work if you are ever asked. Several customers have asked for a 'just send me the top five' button. We understand the appeal. We also know that top five by whose definition is the question a regulator or a plaintiff's lawyer will ask, and the answer needs to be a person's name.
AI in recruiting is at its best doing the reading and retyping that used to eat a recruiter's afternoon, and at its worst pretending to be the recruiter. We built for the first case. If you want to see how the parsing and ranking behave on your own resumes rather than a vendor's demo set, email management@apexsales.ai and we will run a batch with you.