A Skills-Inference Engine Reads Resumes Against Posting Language

Does the engine read the Skills section or the whole resume?

The whole resume, going by Workday's documentation, and only as text. A skills-inference engine takes the file a candidate uploaded, extracts what it can into fields, and matches the extracted phrases against a taxonomy of skill names. The administrator guide says Workday's machine learning reads twelve regions of the uploaded resume rather than the Skills section alone ¹. A phrase anywhere in the file can become a suggested skill.

Parsing and inference are two layers, and they fail differently.

Parsing comes first in what an AI resume screening tool does to applications before you see them, and it loses whatever the format hid. The inference layer then reasons over what survived.

Workday concedes the parsing half on its own page. Parsing results can vary based on resume format and order of words, the guide states, and for best results resumes should not contain images or image-based styles ².

So the input to the inference is a transcription, made from a document whose author had no view of the parser. A candidate who set a heading in an image-based style made a data-quality decision without knowing there was one. The engine inherits the result and infers from it.

Whose words became the skill names?

From text that employers and workers wrote, which is the part a demo skips. Workday's same page says the suggestions come from a machine learning algorithm and an integration with Skills Cloud, working from similar worker profiles and job requisition data ¹. What the suggestion starts from is other people's profiles and the requisitions employers wrote.

Nobody in that chain watched anyone work.

A published taxonomy shows the shape. Indeed's GenAI Skill Transformation Index rated almost 2,900 work skills found in US job postings, then mapped them onto more than 53.5 million postings between May 2024 and April 2025 ³.

Indeed rated those skills with large language models, fifteen passes each. The report calls the result a baseline index of transformation potential ³.

But the vocabulary underneath all of it is still moving. Indeed's August 2026 snapshot puts AI-related postings at 6.3 percent of US postings on the site, past a prior peak of 3.3 percent in 2022 ⁴. The figure counts postings that mention AI, which makes it a measure of employer language, and language is exactly what a taxonomy learns from.

A 2023 resume can describe the same work in words the current list has no entry for.

Tool names age fastest of anything on the page, and the judgment transfers across tools while the tool habits do not. An engine matching a stack name is measuring which product a former employer bought.

So an engine built this way tells you which candidates write like the postings. Phrasing is a real fact about a person. The requisition was written to find a different one.

Separate extraction accuracy from predictive validity

Two different claims travel under the word accuracy. Extraction accuracy asks whether the engine pulled the right phrase off the page, and a vendor can answer that from its own test set. Predictive validity asks whether the inferred skill relates to performance in the roles you are hiring for, which is a much harder claim to make.

Ask which one the demo is showing.

Bias testing is a separate claim with the same shape. The questions for a vendor who says they're bias tested apply here unchanged, because a skills engine is one of the tools they were written for.

The label on the product does not decide the question. Any measure, combination of measures, or procedure used as a basis for any employment decision is a selection procedure under the Uniform Guidelines ⁵.

The same regulation reaches the full range of assessment, from paper and pencil tests through informal or casual interviews and unscored application forms. The Guidelines date from 1978 and name no software, but read on their own terms, an inferred skill used as a basis for setting a resume aside fits that definition.

The validation burden lands where adverse impact appears, not on purchase, so the useful question is narrower than it sounds.

What has the vendor validated, against which outcome, on which roles, and would the study survive somebody asking to read it?

Does an inferred skill trigger Local Law 144?

The answer turns on what the output does in your process, not on what the vendor calls it. New York City's Local Law 144 bars using an automated employment decision tool on a candidate in the city unless a bias audit was done within the prior year and a summary of the results is posted publicly. The candidate also has to get notice at least 10 business days before the tool is used ⁶.

Scope is where the argument happens. The law reaches a tool that substantially assists or replaces discretionary decision-making, so an employer that says its engine merely informs a human reviewer takes itself out of scope.

The city does not second-guess that call up front. Whoever writes it down owns it later.

Outside the city, definitions can reach software nobody bought as AI. Knockout questions in an applicant tracking system may already be regulated under at least one state's definition, and a skills-inference engine does far more with a resume than a knockout question does.

A regulator has already audited tools in the wider category. Between August 2023 and May 2024 the UK Information Commissioner's Office audited developers and providers of AI sourcing, screening and selection tools. Some of the tools, the office reported in November 2024, estimated or inferred gender, ethnicity and other characteristics from an application or from a name alone ⁷.

The office concluded that the inferred information is not accurate enough to monitor bias effectively, and was often processed with no lawful basis and without the candidate's knowledge ⁷.

Those audits were consensual and the participants self-selected, so the finding sizes nothing. Read it as a regulator's account of inferring who a candidate is from application text, the same raw material a skills engine reads for what they can do.

Ask for the corpus in writing before the pilot

Put the four questions in writing before the pilot, because a written answer is the only version anyone can hold up eighteen months from now. A demo answers none of them and is not meant to. Send them to the account team, ask for a reply in the same document, and read what arrives rather than what was shown on the call.

1. The corpus. What text the taxonomy was built from, and when it was last refreshed. 2. The input. What the model takes as input for AI capability specifically, field by field. 3. The evidence. What ties the inferred skill to performance in roles like the ones you are filling. 4. The use. Whether the output orders, filters or excludes anybody, and who wrote that answer down.

Add a fifth that no demo covers. Ask for the false-negative behaviour on resumes written before the current vocabulary existed, and ask how it was measured.

A vendor that has tested it can say so in a sentence.

The same inference runs earlier in the funnel under a different name. Sourcing from companies known to be AI-forward infers capability from where someone worked, and it breaks on a related fact: the signal describes the company.

A recruiter reading four hundred resumes for AI evidence is reading words too, at speed, and with no record of which ones moved them. The screen the engine replaced ran on vocabulary as well, which is the objection with the most force behind it.

Ask whether anyone has measured the two against the same pile.

Workday's setup page is a public document, and it answers the input question, and part of the corpus question, for one product in about forty words ¹. Any vendor could publish the equivalent.

One date may already be on the calendar. If the tool is in scope, ten business days before it touches a New York City candidate, somebody has to write the notice saying it exists ⁶. The notice is easier to write with the corpus sentence in hand.

Get it before the pilot.

Read the evidence >