Guide

How voice agents help you hire better people.

For as long as hiring has existed, most applicants have been assessed on one page they wrote about themselves. A voice agent can now have a real conversation with every one of them, about the actual job, and bring all of it to the people who decide.

This page sets out the whole case: what changes, why it couldn't happen before, what the evidence says and who gains.

Every number carries its source and strength. A: strong research. B: good, narrower. C: weak or survey. Measured: Kit's own data. Estimate: worked out, not measured.

400,000 words
of contextual information per hire: ~2,000 words per Kit conversation vs ~500 per CV, for a role with 200 applicants
Estimate
22.9 min
median Kit call
Measured
783 in 1,000
people who start a Kit call finish it
Measured
960 in 1,000
people are happy or delighted at the end of a Kit call
Measured
276 in 1,000
Kit calls happen between 7pm and 7am
Measured

The claim

Hiring goes wrong mostly because the people deciding don't have enough information about enough of the applicants. Voice agents fix the information problem, at the scale of the whole applicant pool, for the first time.

A CV is a thin, self-written sales page. It says where someone has been. It rarely says whether they would thrive in this team, with this manager, on this problem.
A good recruiter's conversation is the best information there is, and almost nobody gets one. There has never been time to call 200 people for one job.
A voice agent can hold that conversation with every applicant, any time of day, carrying what the hiring manager knows about the job into each one.
So the people most likely to have real impact become visible before anyone is picked for interview, including people whose CVs would have put them at the bottom of the pile.
The decision stays human. Kit recommends. The employer decides. Kit never rejects anyone.

The idea, in one sentence

A voice agent that understands the vacancy, the environment, its challenges and its goals can bring that understanding to a conversation with every applicant, which produces far more useful information than a CV, so the people likely to have high impact can be seen before interviews start.

The separate ideas inside it

Kit understands the job first. Not the job ad: the manager, the team, what is hard about the role, what success looks like.
It brings that to every applicant, from every source, assessed the same way.
The result is far more information than a CV. About 2,000 words from each person, about 400,000 words for a pool of 200.
High-impact people become visible before interview, so interview time goes on them.
This only became possible recently, because voice technology became good enough for people to want to speak to it, and cheap enough to use with everyone.

Who it matters to, and what they get

Applicants get heard instead of filtered. Recruiters get their hours back for the people worth calling. Hiring managers interview fewer people and better ones. Heads of talent get one consistent standard across every source. Finance spends less on hires that fail and on advertising that brings the wrong people. The board gets evidence about hiring quality it has never had.

“When I saw voice, it was like an epiphany moment... This gives us the contextual information that we need to make those predictions on scale.”

Dan Halstead, founder of Speak to Kit, 23 October 2026

The information gap

Every hiring decision is a prediction: will this person do well here? The quality of a prediction depends on the quality and amount of information behind it. Today, for most of the applicant pool, that information is a CV and sometimes a few form answers.

~500 words

CV, one page

Estimate

~2,000 words

Kit conversation

Estimate

Per applicant. The CV figure is a typical one-page CV and is an assumption. The 2,000 words is the applicant's own speech in a typical Kit call, consistent with the measured median call length below.

Across a whole pool

Take a role with 200 applicants. If each speaks to Kit, that's 200 × 2,000 = about 400,000 words about the people applying, every word gathered against the same picture of the job. That's roughly the length of five full-length novels, for one vacancy. Today the same 200 people produce 200 CVs, most of which get a quick scan.

400,000 words

of contextual information per hire: ~2,000 words per Kit conversation vs ~500 per CV, for a role with 200 applicants

Estimate

22.9 min

median Kit call

Measured

64.5 h

of conversation for one customer, Dot Collective: 293 people across 7 jobs

Measured

~500,000

spoken words in those 64.5 hours, at about 130 words a minute

Estimate

"400,000 words" is an illustration for a pool of 200, not a measured count for a named job. The 130 words a minute is a normal speaking rate, assumed, not measured from Kit's transcripts.

What the conversation holds that a CV can't

Dan Halstead spent his agency career doing this by hand. His method: understand the manager, the culture, the goals and what is genuinely difficult about the job (a demanding stakeholder, long hours, a tight budget, a lot of autonomy, complicated relationships). Then find people who had already done those things, and enjoyed them.

“All of that contextual understanding of the environment, the fit, why someone works out... that TAs can't get and recruiters don't ask for.”

Dan Halstead, 16 October 2026

Specific knowledge nobody lists. On a Formula 1 role, about 20 applicants knew how F1 sponsorship works. "No way you would find that on someone's CV." (Dan Halstead, 29 October 2026)
What someone actually ran. One Rhaelm applicant's CV read like a medical transcriber and data cleanser. In conversation it emerged she had run work for a Lloyd's of London business. Kit rated her around 90%. She got the job.
Why they want this job. Desire for the specific job is one of the stronger predictors of who stays (Barrick and Zimmerman). A CV can't show it. A conversation can.
Whether they know what they're walking into. Unmet expectations drive early quitting far more than ability does (Griffeth and others, 2000). A conversation that explains the real job is a realistic job preview, which reduces early turnover (Earnest and others, 2011).

Dan Halstead puts the larger point this way: today's process, built on two documents (the job description and the CV), wouldn't exist "if it had been invented in an AI world". Kit can "bring tens of thousands of words of context into every decision".

High-impact people become visible before interview

The usual process narrows the pool on the thinnest information, at the start, and only gathers rich information (interviews) on the few who survive. That's backwards. The people cut at the CV stage never get the conversation that would have shown what they could do.

Kit turns it the right way round. Everyone gets the conversation first. Then the shortlist is built on the rich information, and interviews go to the people most likely to have impact. Dan Halstead called this the "sleight of hand" of his best agency years: only put people likely to have real impact in front of the manager.

Why the top of the pool is worth finding

Performance isn't a bell curve. Across hundreds of samples of people at work, a normal bell curve fitted only 4.8% of them. Most show a long tail: a small number of people produce a large share of results (Joo, Aguinis and Bradley, 2017). Strength A.
The gap is widest in complex jobs. In high-complexity jobs, the spread of output between people (one standard deviation) is about 48% of average output, against 19% in simple jobs (Hunter, Schmidt and Judiesch, 1990). Strength A.

So the cost of missing the strongest applicant isn't small. In professional roles, the difference between a good hire and a great one is large, and the great one is often not the best-looking CV.

“You cannot talk to everyone, so you start screening CVs, and quite often you just miss out on good candidates whose CV didn't stand out.”

Svetlana, Dot Collective, speaking on a public podcast

Why it wasn't possible before

Everyone in hiring already knows a real conversation beats a CV. The reason it hasn't been done for every applicant is that four things stood in the way. Each has now moved.

1. Recruiter time

200 applicants at the median Kit call length of 22.9 minutes is about 76 hours of talking, before notes and scheduling. That's two working weeks for one vacancy. No recruiter has that, so the CV filter exists to protect their time, not because it predicts well.

Dan Halstead puts the daily reality plainly: recruiters lose two to four hours a day to inbound noise, most of it about people who will never be hired, and they aren't making more placements than a decade ago. He left agency recruitment because "I couldn't really personally impact more than a handful of transactions in any kind of month."

2. Voice good enough that people want to speak to it

Until recently, automated phone and chat screening felt like a form read aloud. People dropped out or gave short answers. That changed with the latest generation of voice models. Dan Halstead's own first call with a voice agent began with "I don't want to talk to AI" and ended with "so good".

783 in 1,000

people who start a Kit call finish it

Measured

22.9 min

median call. Longest at Dot Collective: 56 minutes

Measured

78%

of applicants chose the AI interviewer when offered the choice, in a 70,000-person trial

“People end up talking to it for much longer than... time limited talent person.”

Svetlana, Dot Collective, public podcast

3. When people can speak

The best applicants are often in jobs already. They can't take a recruiter's call at 11am. Kit is there whenever they are.

276 in 1,000

Kit calls happen between 7pm and 7am

Measured

361 in 1,000

Kit calls happen outside working hours (Monday to Friday, 8am to 6pm), weekends included

Measured

3 weeks

the wait Svetlana described for "a recruiter to find like 20 minutes", which Kit removes

4. Assessing everyone the same way

Even a recruiter with unlimited hours can't hold 200 conversations and score them consistently. Fatigue, order and mood all creep in. Research going back decades shows that applying the same scoring rule every time beats human judgement by more than 50% (Kuncel and others, 2013). Strength A. Kit asks every applicant about the same job, from every source, and scores every one the same way. As Dan Halstead puts it: you "can't have a screening function for one half of the candidate pool and not for the other half".

What the science says

Kit's design isn't a guess. It follows the most replicated findings in selection research.

Structured conversations predict performance much better than unstructured ones

Structured interviews (same questions, same scoring) predict job performance at about .42. Unstructured interviews manage about .19 (Sackett and others, 2022). Strength A. Earlier work found validity rising from about .20 to .57 as interviews become more structured (Huffcutt and Arthur, 1994). Strength A. Every Kit call is structured around the same scorecard for the role.

The things a CV shows best predict least

Years of experience barely predict performance (.06) and don't predict who stays at all (.00). Grades are weak too (.21 for performance, slightly negative for staying) (Van Iddekinge and others, 2019). Strength A. The CV filter leans hardest on exactly these.

CV filters miss good people, and the misses aren't random

In a large field study, applicants that a machine model picked and human screeners rejected were 14% more likely to get an offer after interview, and 0.2 to 0.4 standard deviations more productive once hired. The gains came mostly from people with non-traditional CVs (Cowgill, 2020). Strength B.

Overriding a consistent method makes hiring worse

When firms introduced a consistent test, new hires stayed about 25% longer. Managers who overrode the test's advice hired people who left 6 to 7% sooner (Hoffman, Kahn and Li, 2018). Strength A. This is an argument for giving managers better information, then letting them decide with it in front of them.

The first controlled trial of AI voice interviews

In 2025, Jabarian and Henkel ran a randomised trial with about 70,000 applicants for entry-level customer service jobs in the Philippines. Applicants were interviewed by an AI voice agent or by a human recruiter. Humans made every hiring decision in both groups. Strength B.

Outcome	AI voice interview vs human
Job offers	+12%
People starting the job	+18%
Still in the job after 30 days	+17%
Applicants who chose the AI when given the choice	78%
Applicants reporting gender discrimination	3.30% vs 5.98%
Applicants who dropped out rather than speak to the AI	5%

Recruiters in the study predicted the opposite result. Limits: one country, entry-level roles, one firm. It's strong evidence that the method works, not proof for professional roles in the UK.

Who stays

Ability barely predicts quitting (correlation about .01). Unmet expectations do (Griffeth and others, 2000). Wanting the specific job, knowing people there and a history of staying in jobs predict early leaving (Barrick and Zimmerman). Realistic previews of the job reduce turnover (Earnest and others, 2011). Referred hires quit less (Burks and others, 2015). All of these are things a conversation can find out and a CV can't.

What has happened with Kit

Rhaelm: the CV said 100th. Kit said first. She got the job.

About 200 applicants. The applicant tracking system marked one applicant as missing two of the five criteria. Rhaelm's hiring lead guessed she was "about the 100th CV down the list". Her CV read: "medical transcriber, data cleanser... how's that relevant". Kit rated her around 90%. In conversation she described running the show for a Lloyd's of London business. She was hired.

The hiring lead also met fewer people: "I probably had to at least speak to five or six, whereas actually I've spoken to three." Applicants "felt that it was in their corner." 66 of 70 applicants who started a call finished it (Measured).

Rhaelm's hiring lead on call and by email; platform record.

A global pharmaceutical manufacturer hiring clinical trials staff: 85% of Kit's picks were right

The hiring lead's feedback: 85% of Kit's picks for the role were right. He also said of the decision: "I am the loop."

One role only, so this is a single result, not an average.

The customer's hiring lead on call. Name withheld.

Omnicom: an offer to someone Kit spoke to

Omnicom's hiring team reported an offer made to a person Kit had spoken to: "really hard to ignore that". Asked if it saved time: "absolutely". They also said they wouldn't rely totally on the score, which is exactly how Kit is meant to be used.

Platform record for the Octagon role at Omnicom: 213 invited, 46 spoke to Kit, 41 finished, 8 recommended (Measured). About 20 applicants showed real knowledge of F1 sponsorship.

Omnicom's hiring team on call; platform record.

Dot Collective: everyone gets a chance at the start

556 applications across 7 jobs. Kit spoke to 293 people, 64.5 hours of conversation, the longest 56 minutes (Measured). Svetlana, on a public podcast: "at that entry point everyone has a chance". She described the process as CV blind, with 55 data points per applicant.

Platform record; Svetlana, public podcast.

Rocket Yard and one of our customers: from hundreds to a handful

Rocket Yard: 180 applied, 15 spoke to Kit, 2 met, 1 hired (Measured).

One of our customers: 112 applied, 108 spoke to Kit, 2 recommended (Measured).

Platform records.

Apogee (an HP company): the same front door, however busy

“Kit keeps the front door consistent no matter how busy things get.”

Kelly, Head of Talent Acquisition, Apogee (an HP company)

Kelly's quote, cleared for use.

A leading global recruitment group: capacity is the problem

"Capacity is definitely the key issue." Their idea of the future: a "five-minute qualification call after they've spoken to Kit rather than... 20 minutes".

Call with a leading global recruitment and RPO group. Name withheld.

Each person's case

A purchase like this is decided by several people with different worries. Each needs their own reason.

The applicant

Today most applicants are assessed on a page and never hear back. With Kit, every applicant gets a real conversation about the job, at a time that suits them, and gets heard for what they can do, not only where they've been.

This matters most for the people CVs miss: a mother coming back to work after five years at home with her children, someone changing career, someone whose best work was never in their job title. Cowgill's study found the people human screeners wrongly rejected were mostly those with non-traditional CVs. Dot Collective's description: "at that entry point everyone has a chance." Rhaelm's applicants felt Kit "was in their corner".

The recruiter

Kit gives recruiters the conversation they would have with every applicant if they had the time. Their hours go to the people most worth calling, and their calls start further along: five minutes to confirm instead of twenty to discover. Kit recommends; the recruiter and the manager decide. One manufacturer's hiring lead put it simply: "I am the loop."

The hiring manager

Fewer interviews, with people more likely to succeed. Rhaelm's hiring lead met three people instead of five or six. Every shortlisted person comes with what they said about the real job, so the interview starts from understanding, not from the CV.

The head of talent

One standard for every applicant from every source: job sites, the careers page, referrals, agencies, sourcing. That means the team can finally see which sources bring the people most likely to succeed, not just the most applicants, and put advertising money there. Only 3 in 10 UK HR teams measure how new hires perform in their first year (CIPD, 2026, 1,014 HR professionals). Strength B. Kit gives them a quality measure at the point of hire.

Finance

Bad hires are expensive, and early leavers more so. Replacing someone costs about £30,614 and takes 28 weeks to reach full productivity (Oxford Economics, 2014). A bad hire can cost up to three times salary; the REC's worked example puts a £42,000 hire at £132,015 (REC, 2017). 41% of UK employers saw new hires quit within 12 weeks (CIPD, 2024). Kit also points advertising spend at the sources that bring the strongest people, so less goes on volume that never hires.

The board

People are most companies' biggest cost and biggest lever, yet boards rarely see any evidence about hiring quality. Kit gives a consistent, recorded basis for every hiring decision: who applied, what they said, why they were recommended. That's a quality measure the board can follow over time, and a fairer process it can defend.

The hardest objections, answered

"People won't speak to a machine."

The strongest version: Good applicants will walk away rather than be interviewed by AI.

The people who speak to Kit like it. 960 in 1,000 people are happy or delighted at the end of a Kit call; 11 in 1,000 aren't (Measured). 783 in 1,000 people who start a Kit call finish it, and the median call runs 22.9 minutes (Measured). In the 70,000-person trial, 78% chose the AI interviewer when given the choice and 5% dropped out rather than speak to it. People speak to Kit for longer than a recruiter could give them.

"AI is biased."

The strongest version: An AI will repeat the prejudices in its training and do it at scale.

The comparison is with today's process, not with perfection. Today a CV filter decides who gets an interview, assessing people on proxies: name, schools, gaps, job titles, years served. Only a few ever get one. Kit replaces that selection by CV. It doesn't replace the interview.

Kit goes deeper than any of those proxies. It starts from the real environment of the job: the manager, the team, the pace, what is genuinely hard about the role and what success looks like. Then it explores, with every applicant, whether they have the capabilities a person needs to thrive in that environment, and the examples that show it. People are assessed on merit against the real job, not on where they went to school or how their career looks on paper.

Everyone gets the same conversation about the same job and is scored against the same scorecard, from every source. In the trial, applicants reported gender discrimination at 3.30% with the AI against 5.98% with humans. Kit recommends and a person decides. Kit is under UK GDPR, and anyone can ask for their details to be deleted at any time.

"Our policy says AI must not make recommendations that affect candidates."

The strongest version: Some employers' policies say generative AI must not make recommendations affecting candidate outcomes.

This is a real constraint for some employers and the honest answer respects it. Kit's recommendation is information for a human, never a decision. Where a policy forbids any ranking, the conversation itself still gives the human far more to go on than a CV, for every applicant rather than a few. And a CV keyword filter already affects candidate outcomes, with much less information behind it.

"Applicants will use AI to game it."

The strongest version: A polished, rehearsed applicant will beat an honest one.

The things employers rely on today are the easiest to fake. A CV can be rewritten by AI in seconds to match any job ad. Cover letters and written application answers can be generated outright. Online skills tests are easy to fake too: someone else can sit them, or an AI can answer them in another window.

A live conversation is far harder to game. Kit asks about the real job, then follows up on what the applicant actually says: what happened, what they did, what they would do differently, how that compares with what this team needs. A rehearsed answer runs out after the first follow-up. And the applicant has to keep it up, consistently, for twenty minutes or more.

This is an argument from how the conversation works. Kit doesn't yet have data on how many rehearsed applicants it catches.

"A score can't capture a person."

Agreed, and Kit doesn't ask it to. Kit's job is to make sure the right people reach the room, with what they said attached. Omnicom's hiring team wouldn't rely totally on the score, and still found the result "really hard to ignore". The interview, the judgement and the decision stay with people.

"This replaces recruiters."

Kit takes the part no recruiter has time for, speaking to everyone, and hands recruiters a shortlist and the context for it. The recruitment group's view: capacity is the problem. Recruiters spend their time on the people worth calling.

"Experience and qualifications are good enough."

Years of experience predict performance at .06 and staying at .00 (Van Iddekinge and others, 2019). A filter built on them is mostly filtering on noise.

Evidence table

FindingSource and sampleStrength
AI voice interview: offers +12%, starts +18%, 30-day retention +17%; 78% chose AI; 5% dropped out
Jabarian & Henkel, 2025 · ~70,000 applicants, entry level, Philippines, randomised
B
Structured interview .42 vs unstructured .19
Sackett and others, 2022 · Meta-analysis
A
Validity rises .20 to .57 with structure
Huffcutt & Arthur, 1994 · Meta-analysis
A
Experience .06 performance, .00 staying; grades .21 / -.02
Van Iddekinge and others, 2019 · Meta-analysis
A
Consistent rule beats judgement by over 50% (.44 vs .28)
Kuncel and others, 2013 · Meta-analysis
A
Testing raised tenure ~25%; overrides cut it 6-7%
Hoffman, Kahn & Li, 2018 · Field data, 15 firms
A
Machine-picked, human-rejected: +14% offers, +0.2-0.4 SD productivity; mostly non-traditional CVs
Cowgill, 2020 · Field study, one large firm
B
Bell curve fits 4.8% of performance samples
Joo, Aguinis & Bradley, 2017 · Hundreds of samples
A
Output spread 48% of mean (complex jobs) vs 19% (simple)
Hunter, Schmidt & Judiesch, 1990 · Meta-analysis
A
Ability barely predicts quitting (.01); unmet expectations do
Griffeth and others, 2000 · Meta-analysis
A
Desire for job, knowing staff, prior tenure predict early leaving
Barrick & Zimmerman · Field studies
A/B
Realistic job previews cut turnover
Earnest and others, 2011 · Meta-analysis
A
Referred hires quit less
Burks and others, 2015 · Firm data, 9 firms
A
Replacement £30,614; 28 weeks to full productivity
Oxford Economics, 2014 · UK employers
A/B
Bad hire up to 3x salary; £42k hire = £132,015
REC, 2017 · UK modelled example
B
41% of employers saw new hires quit within 12 weeks
CIPD, 2024 · UK employer survey
B
3 in 10 measure first-year performance
CIPD, 2026 · 1,014 UK HR professionals
B
960 in 1,000 happy or delighted at the end of the call; 11 in 1,000 not
Kit, end-of-call rating · Kit calls
Measured
Per 1,000 calls: 783 completed; 276 between 7pm and 7am; 361 outside working hours (Mon to Fri, 8am to 6pm), weekends included. Median call 22.9 min
Kit, PostHog · 14 Jul to 25 Oct 2026
Measured
Dot Collective: 556 applications, 293 people spoken to, 64.5 hours
Kit platform · 7 jobs
Measured

Strength reflects sample, method and how close the setting is to UK professional hiring.

Speak to Kit.

What's important when hiring for your organisation?