Two resumes, same tool, same author, nearly double the callback rate. In a two-month field test of 204 genuine applications to live postings, a raw AI-generated resume sent as drafted drew responses on 10% of applications. The same AI output, tailored to each specific posting, drew 18% (Kickresume, 2026). Nothing about the candidate changed between those two arms. What changed was how much work went into the submission.
That result should reorganise how you think about your own funnel. Most hiring teams in 2026 are braced for the wrong problem. The fear is that AI-written applications are too polished to screen โ that the machine-written resume sails through because it reads better than the human one. The data points the other way. The raw AI draft was the weakest performer in the entire run. What your resume screening rewards is not authorship and not capability. It is tailoring effort โ and tailoring effort is the one input that generative AI has made almost free.
What 204 Live Applications Actually Measured
The design matters, so here it is plainly. Over roughly April to June 2026, 204 real applications went out to live postings in the New York market, concentrated in customer service and sales-representative roles โ deliberately high-volume categories. Four resume versions were rotated, 51 applications each: a human-written single-column resume, a human-written two-column resume, a raw AI draft, and an AI resume tailored per posting (Kickresume, 2026).
Across the whole run, 36 responses came back positive โ about 18%, roughly one in six โ of which 22 were direct interview requests and the rest phone screens, video screens, positive emails and assessment invitations. Most replies arrived inside three days.
The spread between arms is the finding:
- Human-written, clean single-column: 31%
- AI tailored to each posting: 18%
- Human-written two-column: 12%
- Raw AI draft, sent as generated: 10%
Two things sit inside those numbers that are easy to miss. First, the two-column human resume underperformed the tailored AI resume โ layout parsing is still costing real candidates real interviews, which is a quiet indictment of the parsing layer rather than of the applicants. Second, 30 of the 36 positive responses came from genuinely well-matched roles, and customer service and support alone accounted for 18 of 36. Match quality dominated everything else in the data.
One honest caveat, which the authors flag themselves: the versions went out in phases rather than head-to-head on identical postings, so the human resumes met the best-matched openings first, before the local pool of well-matched roles was exhausted โ something the authors note took about a week in the country's highest-volume job categories. That confound inflates the human-single-column number by an unknown amount. Treat the 31% as directional. The comparison that survives the confound is the one inside the AI pair, where the tool and the author were held constant and only tailoring varied: 10% to 18%.
Your Resume Screening Is Reading Effort, and Effort Just Got Cheap
Here is the mechanism. When a screen โ human or automated โ rewards a resume for using the posting's vocabulary, mirroring its stated requirements and foregrounding the relevant third of a career history, it is not measuring whether the candidate can do the job. It is measuring whether the candidate spent twenty minutes on you rather than hitting submit on a generic file.
For most of hiring history that was a defensible proxy. Effort correlated with motivation, motivation correlated with fit, and the cost of tailoring was high enough to be informative. A candidate who rewrote their resume for your posting had revealed something real about their preference ordering.
Generative AI broke the price of that signal. Tailoring a resume to a job description is now a single prompt, performed in seconds, at zero marginal cost, at unlimited volume. The signal did not get noisier โ it got inverted. A funnel that rewards tailoring now preferentially advances the candidates with the best application tooling, which is uncorrelated with the candidates who will perform in the seat.
The employer-side data already shows the strain. Only about one-third of hiring leaders say they are very confident a resume reflects a candidate's true skills, yet roughly two-thirds still use a resume screen as the first step in the funnel; organisations where the resume is the primary decision driver are 35% more likely to report a bad hire (Criteria and Lighthouse Research & Advisory, 2026, survey of 998 hiring leaders). Separately, 54% of senior talent leaders say AI-generated resumes and applications have made it harder to assess true capability, while more than 90% of companies run AI across sourcing and screening and fewer than 5% call the results transformational (ManpowerGroup and Everest Group, 2026). And 67% of US HR leaders report that reviewing AI-generated applications has slowed their hiring, with 20% seeing delays beyond two weeks (Robert Half, 2026).
Read those four findings together and the picture is specific: the gate everyone still opens first is the one almost nobody trusts, it has become more expensive to operate, and the thing it now measures is purchasable.
The Two-Column Result Is a Separate Warning
Set the AI question aside for a moment and look at the two human arms. The same candidate, the same history, the same claims โ rendered single-column, 31%; rendered two-column, 12%. That gap has nothing to do with capability, motivation or fit. It is a rendering artifact.
Two-column layouts are where resume parsers routinely fail: text columns get interleaved into nonsense, skills sidebars get dropped, dates get detached from the roles they belong to. If your applicant tracking system is scoring or ranking on parsed fields, a layout choice made for aesthetic reasons silently moves a candidate down your list. In the field test's own numbers, a human-written resume in the wrong layout underperformed an AI draft that had merely been pointed at the posting (Kickresume, 2026).
This is worth an afternoon of diagnostic work regardless of what you conclude about tailoring. Take ten resumes you rejected at the screen in the last quarter, run them through your parser, and read what your system actually ingested versus what the candidate actually submitted. The failure mode is not that the parser throws an error โ it is that the parser succeeds, produces degraded fields, and the ranking downstream treats them as true.
The Counter: Isn't Tailoring Still a Real Signal?
The strongest objection deserves a fair hearing. If a candidate bothers to tailor โ even with AI โ they have still chosen your posting over the forty others in their tab. Doesn't that reveal preference?
Partly, and the field test supports it: 30 of 36 positive outcomes came from well-matched roles, which means fit was doing most of the work. But preference and fit are not the same variable, and tailoring only proxies the first. A candidate can tailor perfectly to a role they will leave in five months, and a strong-fit candidate with a plain single-column resume and no tailoring habit can be screened out before any human reads a word. Worse, the proxy degrades asymmetrically: the candidates most likely to be mass-applying with AI assistance are precisely the ones for whom tailoring costs nothing, so the correlation between tailoring and genuine interest weakens fastest exactly where application volume is highest.
The defensible version of the objection is narrower than it sounds. Tailoring is weak evidence of preference and almost no evidence of capability. It is fine as a tiebreaker. It is indefensible as a ranking input.
Measure the Capability, Not the Document Craft
Roughly eighty years of selection research points the same direction: work samples, structured interviews and cognitive-ability measures are among the strongest predictors of job performance, while resume-adjacent inputs like years of experience and educational credentials are comparatively weak (Schmidt and Hunter meta-analysis, 1998). That finding predates generative AI by a quarter of a century. AI did not invalidate the resume โ it removed the last practical excuse for gating on it first.
The operational distinction is between inferring capability from how a document was assembled and measuring it against what the role requires.
What to change this quarter
Pick one high-volume req โ the kind where application count has doubled and your recruiters are drowning in near-identical, well-tailored files. Then do four things:
- Demote the resume to a knock-out filter. It answers eligibility questions (work authorisation, hard credential, non-negotiable experience floor) and nothing else. It does not produce a rank.
- Put a structured, job-representative signal ahead of the ranking step โ a short work sample, a validated assessment, or a structured interview scored against the role's stated requirements rather than against the other applicants.
- Weight the requirements before you see a single candidate. A requirement that blocks the hire should not count the same as a nice-to-have, and deciding that ordering after applications arrive is how post-hoc justification enters a funnel. Explicit criticality weighting is the part of scoring mechanics most teams skip, and it is the part that makes a ranking auditable.
- A/B it against a comparable resume-first req and watch two numbers: quality-of-hire at the 90-day mark, and recruiter hours reclaimed from filtering.
If the tailored-AI arm of that field test tells you anything operationally, it is that your funnel will happily pay a premium for twenty minutes of prompt work. The question for this quarter is whether you would knowingly rank candidates on that. Until the capability signal sits ahead of the document, your resume screening is buying twenty minutes of prompt work and calling it evidence โ and that is the purchase you make again every time you open the pipeline.