In 2023, fewer than two in ten job seekers rated their hiring experience as excellent. In 2026, after three years of accelerating investment in hiring technology, that number is 11% (PeopleScout, 2026). The candidate experience did not hold flat while the tooling improved. It fell by roughly seven points.
The obvious suspect is the AI you added to the funnel. The data does not support the obvious suspect, and that is the more expensive finding — because it means the fix is not a procurement decision you can defer to next year's budget.
What Three Years of Hiring-Tech Investment Actually Bought
PeopleScout surveyed more than 1,000 job seekers across 10 markets and separately audited employer hiring journeys across 20 sectors, scoring both against its 2023 research (PeopleScout, 2026).
| Measure | 2026 | 2023 |
|---|---|---|
| Job seekers rating their experience very positive | 11% | 18% |
| Candidates who say they were never ghosted | fewer than 1 in 10 | — |
| Job seekers who used AI in their search in the last six months | 52% | — |
PeopleScout's own verdict on the decline is a single sentence: "This is not a technology story. It's a communication story."
That is a staffing vendor's framing and I will come back to whether it holds. But take the structure of the claim seriously first. The thing that got worse is not the part of the process a screening tool touches. It is the part no screening tool was bought to do.
The Candidate Experience You Are Actually Being Scored On
Read the ghosting number again: fewer than one in ten candidates say they were never ghosted. That is not a measure of whether your scoring model is fair, explainable, or fast. It is a measure of whether anyone ever told them where they stood.
This matters operationally because the two numbers come apart. You can run a defensible, audited, well-calibrated screen and still produce a terrible candidate experience, because the experience is made of status updates, reasons, and the presence of a human at the point where the answer is no. The behavioural evidence points the same way: 38% of US candidates have withdrawn from a hiring process at least once specifically because it included an AI interview, and the most-cited trigger was a pre-recorded video interview scored by AI with no human present (Greenhouse, 2026).
Note what that trigger is made of. Not the AI. The empty chair behind it.
Silence Also Corrupts Your Own Data
There is a self-interested reason to care, separate from employer brand. A candidate who is never told anything cannot distinguish a slow process from a dead one, so they keep applying elsewhere and accept the first offer that arrives. Your acceptance rate, your offer-to-start conversion and your time-to-fill all then carry a term you are not measuring: how many of your finalists had already mentally left because nobody confirmed they were still in play. Ghosting does not only cost goodwill. It adds noise to the three funnel metrics you report upward.
The Volume Side Is Getting Worse While the Response Side Thins
Here is why the communication deficit compounds rather than holding steady.
More than half of job seekers — 52% — used AI in their search in the last six months, mostly to write applications, optimise CVs and prepare for interviews (PeopleScout, 2026). On the employer side, 92% of recruitment leaders now say AI-generated résumés are commonplace (Criteria and Lighthouse Research, 2026). And the added load is landing on throughput: 67% of US HR leaders report that reviewing AI-generated applications has slowed their hiring, with 20% reporting delays of more than two weeks (Robert Half, 2026).
Applications go up, review time per hire goes up, and the first thing a stretched team stops doing is the thing with no deadline attached: telling the 94% who did not get it. The ghosting rate is not a culture failure. It is a capacity arithmetic that nobody re-ran after application volume changed shape.
The asymmetry is worth naming precisely. Every other failure in the funnel has an owner and a deadline — an unfilled requisition has a hiring manager asking about it, a stalled offer has a start date attached. An unsent rejection has neither. It is the only output of the process whose absence generates no internal complaint, which is why it is the first thing to go under load and the last thing to be noticed when it does.
Your Screen Is Reading Effort, Not Capability
If the document itself were carrying good signal, the volume problem would at least be buying you accuracy. A field test suggests it is not.
Kickresume sent 204 genuine applications to live postings over two months in 2026, rotating four résumé versions of 51 applications each: human single-column, human two-column, a raw AI draft, and an AI résumé tailored to each posting. Thirty-six came back positive, about 18%, of which 22 were direct interview requests (Kickresume, 2026).
The spread is the finding:
| Résumé version | Positive response rate |
|---|---|
| Human-written, single-column | 31% |
| AI draft tailored to each posting | 18% |
| Human-written, two-column | 12% |
| Raw AI draft, sent as generated | 10% |
Same tool and same author separated by 8 points, purely on whether the output had been adapted to the specific posting. And 30 of the 36 positive responses came from genuinely well-matched roles.
So the screen is reading two things: fit, and how much time the candidate spent on you. Fit is worth reading. Effort was a decent proxy for motivation back when tailoring a résumé cost an hour. It now costs a prompt, which makes it a signal with a collapsing half-life — and one that systematically favours whoever has the better tooling rather than the better capability.
Employers appear to know this and gate on it anyway. Only about a third of employers say they are very confident résumés reflect a candidate's true skills, while two-thirds still use a résumé screen as the first step — and companies where the résumé is the primary decision driver are 35% more likely to report a bad hire (Criteria and Lighthouse Research, 2026).
The Honest Counter
Four limits, stated plainly.
Three of the five datasets here are vendor research. PeopleScout is a staffing firm, Criteria sells psychometric assessment, Kickresume sells résumé building. Each has a commercial interest in the direction its finding points. Robert Half is also a staffing firm. Treat the convergence as corroboration, not independence.
The PeopleScout sample is thin per market. Just over 1,000 respondents spread across 10 markets is roughly a hundred per country. The global aggregate is usable; a per-market read is not.
A decline across two survey waves is not a causal claim about AI. The 2023 and 2026 waves are different respondents in a labour market that changed substantially in between. The honest reading is that candidate experience deteriorated during a period of heavy hiring-tech investment — which rules out "the tooling fixed it," not much more.
The Kickresume design is phased, not head-to-head. Versions went out in sequence rather than against identical jobs, so the human résumés met the best-matched roles first. The vendor flags this itself. Treat the 31/18/10 spread as a strong signal, not a controlled result.
What survives all four caveats is the part that is hardest to argue with: the measure that fell is the one no screening tool was ever instrumented to move.
What to Change This Quarter
Four moves. None of them is a purchase.
- Put a response service level on every stage, and measure it like you measure time-to-hire. Most replies in the Kickresume run arrived within three days, so three days is the operating tempo candidates already experience from your competitors.
- Write the rejection into the workflow as a required step, not a courtesy. If fewer than one in ten candidates have never been ghosted, the default path through your ATS ends in silence. Change the default; the message does not need to be long, it needs to exist.
- Keep a named human at the point where the answer is no. The withdrawal data is specific — what candidates abandon is the scored process with nobody behind it, not the use of AI itself.
- Stop treating tailoring as signal. Decide explicitly which stage measures capability rather than document craft, and move that stage earlier than the résumé screen.
The fourth is the one with a structural fix rather than a process fix. If the document is now a measure of effort and tooling, the alternative is a capability signal that does not depend on the document at all — verified skills, cognitive and personality measures, and a structured interview carried by the candidate rather than reconstructed by each employer. Scovai's Talent Passport is one worked example of what that profile contains.
The deeper point is about what a candidate experience score measures. The 11% is not a verdict on your AI. It is a verdict on your silence — and silence is the only part of the funnel that costs nothing to produce and nothing to monitor, which is exactly why it grew.
Pull your own number this quarter: of everyone who applied last month, what share received an answer, and how many days did it take? If you cannot produce that figure by Friday, you already have the finding.
