All work

Résume for Hire

Resume for Hire

All work

Résume for Hire

Resume for Hire

Indeed · Recruit Holdings (TYO: 6098)

Design & Research

Reducing AI Hiring Bottlenecks for 3.5M+ Employers on Indeed

Impact

Indeed's UX team named Role Calibration and SmartScreen a reference point for future AI explorations. Presented case to Indeed, Google and Amazon teams.

Overview

Indeed's network and two-sided data advantage positions it as the biggest hiring platform in the world by an order of magnitude. However, candidate quality was the biggest complaint and the top driver of negative NPS on Indeed, yet about 60% of applicants met the posted qualifications. Owning design end to end, I worked with another designer to lead an AI screening workflow that lets employers increase hiring velocity and quality.

RESPONSIBILITIES

Design

Prototyping

Research

Timeline

6 months

Jan-Jun 2025

Team

Design (2)
Research (2)
Indeed Design (2)

Watch the walkthrough

Back

HIGHLIGHTS

Problem | 01

60% of applicants met qualifications found in the job post, yet employers (EMPs) struggle to land high quality applicants

Indeed marketplace provides 500M+ job seekers with free access to 30M+ jobs from thousands of company websites and job boards, enabling companies of all sizes to hire talented employees. But if employers can’t secure strong candidates consistently on Indeed's marketplace, they’re more likely to churn in the long run.

Indeed provides 500M+ job seekers with free access to 30M+ jobs from thousands of company websites and job boards, enabling companies of all sizes to hire smart. But if EMPs can’t secure strong candidates consistently and quickly, they’re more likely to churn in the long run.

Indeed gives 500M+ job seekers free access to 30M+ jobs from thousands of company websites and job boards, helping companies of all sizes hire smart. But with one-click applications, employers (EMPs) are flooded with candidates, and when they can’t find strong hires quickly, they’re more likely to churn.

Indeed's marketplace

Job seekers

525M+profiles

~150job seekers for every employer

~17job seekers for every open job

Job listings30M+

Employers3.5M+

Squares are drawn to scale by area. Source: Indeed.

Problem | 01

Low candidate quality is the biggest NPS driver

Low candidate quality is the top driver of negative NPS

Low candidate quality is the top driver of NPS

41%

of negative employer feedback flagged quality

4,000+

complaints of low quality

Problem | 02

AI was built to help with speed and quality but EMPs found its output unreliable

When both sides of the marketplace uses AI

CandidatesTailor resumes with AI

AI writes

Strong match

The pileResumes written by AI, ranked by AI

EmployersScreen large applicant pools with AI

AI ranks

ResultMore AI on both sides made resumes look stronger, but not easier to trust.

Teams felt responsible for good hires that AI "auto" reject

AI summaries were meant to speed up decisions but fewer than 20% of EMPs use them. People stop using automation they don’t trust, even when it could help. Hiring teams didn’t want AI to reject candidates for them because they were accountable for every person they passed on.

Measured against Microsoft’s Guidelines for Human-AI Interaction (Amershi, CHI 2019), global "summaries" missed 3 criteria that matter most for trust: they didn’t show why the AI decided (G11), didn’t let people correct it (G9) and didn’t let teams choose how much it did (G17).

<20%

of employers engaged with Indeed's AI summaries.

Jobseekers use AI too. Indeed’s algorithm tells them about the fit and rewrites their resume in that job post’s language. When an EMP’s AI then summarizes each applicant, it’s summarizing resumes that all echo the same post and every summary sounds alike. This is why EMPs can’t differentiate who has actually done the work. They go to resume or schedule interviews, looking for proof.

Job seekers use AI too. Indeed's AI tells them why they fit a job and rewrites their resume in that job post's language. So when an EMP's AI summarizes each applicant, it's summarizing resumes that all echo the same post and every summary sounds alike. The words match but we don't know who has actually done the work. That's when employers stop trusting the summary and go back to the resume itself for more truth.

Project goals

Become the Trusted & Efficient Hiring Partner for Hiring Managers & Recruiters

Improve candidate quality

Leverage GenAI to surface stronger, more diverse talent without introducing new bias.

Build calibrated trust

Make AI-assisted tools reliable and transparent so teams feel confident adopting them. Build trust that matches how reliable AI actually is.

Project goals

Become the Trusted & Efficient Hiring Partner for Hiring Managers & Recruiters

Improve candidate quality

Leverage GenAI to surface stronger, more diverse talent without introducing new bias.

Build calibrated trust

Make AI-assisted tools reliable and transparent so teams feel confident adopting them. Build trust that matches how reliable AI actually is.

Project goals

Become the Trusted & Efficient Hiring Partner for Hiring Managers & Recruiters

Improve candidate quality

Leverage GenAI to surface stronger, more diverse talent without introducing new bias.

Build calibrated trust

Make AI-assisted tools reliable and transparent so teams feel confident adopting them. Build trust that matches how reliable AI actually is.

Solution

A proactive AI-assisted workflow that validates teams' priorities before automated screening

JD + notes

Priorities set

JD + notes

Priorities set

01 · Role Calibration

A quick intent check that turns hiring goals into structured criteria for AI matching.

Filter

Screen

Compare

AI planAdjust

Filter

Screen

Compare

AI planAdjust

02 · SmartScreen

AI-powered widgets to correct matching, verify resumes and compare tradeoffs.

ClaimLine 3

03 · Candidate Resume View

Every AI claim points back to the resume line behind it.

SOLUTION · 01

Role Calibrator

WHY

Capture better, more relevant context from EMPs

AI can match candidates to a prompt, but it can't decide which priorities matter most to a hiring team.

Role Calibration connects Indeed's job creation workflow to priorities that live outside the job post: team notes, market data and hiring preferences. These information increases relevancy. Recruiters set baseline minimums, soft skills, team fit and risks. This captures constantly changing operational context that job materials leave out, bringing more specificity to future model training data.

Confirm & correct what AI has interpreted

AI turns attachment into bulleted criteria, each traceable to the files. Before matching begins, EMPs can confirm at a high level. "Am I happy if AI said 3-5 things about this role?" and fix anything AI interpreted incorrectly.

Adjust how much each hiring signal counts for matching algorithm

EMPs edit weightings on how each signal adds to matching percentage % score, like experience over technical skills.

Shifting the weights brings different strengths to the top, giving visibility for EMPs to compare who rises in each version of the pool. Later, they have documentation to see how their past selection shaped what AI prioritized. This allow EMPs steer the model directly instead of only reacting to AI output (Lee & See 2004). Trust in Automation: Designing for Appropriate Reliance.

EMPs edit weightings on how each signal adds to matching percentage % score, like experience over technical skills.

Name your "fit" before AI assumes it

Technical skills tends to be where AI-polished resumes look most alike, echoing keywords in job post. What separates candidates is how they work and that rarely makes it into a job post. Only the hiring team knows the soft skills it needs.

For Indeed, team fit preferences are data no job post holds, and they’re specific to each business unit. Each calibrated role sharpens matching, improving one-shot results.

Because the shortlist reflects criteria the team set themselves, I expected AI shortlist use to climb from <20%. I’d measure shortlist-to-interview rate, resumes reread per hire and repeat job posts.

SOLUTION · 02

Smart Screen

Filter

Screen

Compare

AI planAdjust

Why

Trust can be proved through evidence that EMPs can validate in seconds

SmartScreen exposes the logic of AI matching. EMPs see the evidence and open it in a click: the resume, LinkedIn or other data.

SmartScreen exposes the logic of AI matching for the purpose of control. EMPs see the evidence.

What

Let the evidence prove trust

SmartScreen shows which signals AI uses, how each filter applies and how many hires every AI step removes. EMPs compare match strategies and verify claims against external data signals.

Customize shortlist through softskills or impact among AI-polished resumes

Provide global controls. Let teams decide how much AI does.

SMARTSCREEN

SmartScreen turns Role Calibration inputs into a prefilled prompt. EMPs refine it with suggestions like “Spot impact not pedigree,” and the shortlist reranks instantly with Indeed's bias guardrails on. Most resumes in the market now echo the job post because of AI but each candidate’s mix of experience is unique. Prompting for likely skills surfaces that mix where a generic summary would flatten it.

Trace candidate cutoff to prove "quality" in process

Make clear why the system did what it did.

Trace candidate cutoff to show quality in process

In manufacturing, traceability protects quality: when a defect shows up, you must trace the step that caused it and fix it. SmartScreen brings that to screening. It shows the signals it used, the filters it applied and how many candidates each one removed.

When a cut looks wrong, EMPs can find the step, fix it and bring good candidates back. Each fix also becomes training data for Indeed’s matching model.

In manufacturing, traceability protects quality: when a defect shows up, you can find the step that caused it and fix it. SmartScreen brings that to screening. It shows the signals it used, the filters it applied and how many candidates each one removed.

When a cut looks wrong, EMPs can find the step, fix it and bring good candidates back. Each fix also becomes training data for Indeed’s matching model.

Guidelines for Human-AI Interaction (Amershi et al., Microsoft, CHI 2019)

Bulk compare candidate pool to uncover your decision tradeoffs

Comparing candidates one at a time gets slow at high volume scale. SmartScreen lets EMPs compare whole pools at once and see who rises, who drops and who appears in only one.

Clicks and applies are signals every job board collects. A hiring manager choosing between 2 weighted pools shows an expert trade-off, like “leadership over pedigree.” Each choice also gives Indeed preference data edge no competitor has.

Validate resume through external data signals

EMPs verification tested very well because it helps EMPs catch exaggerated skills or false identities early. This concept works by initiating it from the demand side: jobseekers-permissioned sharing.

LinkedIn launched Verified on LinkedIn in April 2025 to improve online trust and authenticity.

Members can carry their LinkedIn-verified identity, workplace and education to partner platforms. Current partners are Adobe, G2, TrustRadius and UserTesting. None are hiring platforms yet, which makes this a true engineering partnership opportunity for Indeed.

Collect granular feedback behind rejections

Resumes are auto-highlighted to show the experience AI used to rank each candidate. EMPs check the source of truth without leaving Indeed, and recruiters finalize decisions without rereading resume top down.

We also recommend collecting granular feedback from EMPs to teach the model what "quality" means. Over time, that feedback data improves Indeed’s AI loop and makes first-pass rankings more accurate.

Results

Role Calibration and SmartScreen became a reference point for Indeed's AI work. Indeed designer mentioned how they will include the feedback loop for AI since some of the design components already existed.

Role Calibration and SmartScreen became a reference point for Indeed's AI work. Indeed designers mentioned they will adapt our design that pulls rejection feedback loop for its training model.

Research

How do businesses of all sizes hire?

Employers carry a mental model of the ideal candidate.

But it's hard to articulate into a job post and templates miss the nuance. Over 90% of users reuse old JDs that may be outdated/misaligned.

Matching isn't about perfection.

EMPs want help spotting potential, breadth and transferable skills. Exact keyword matches miss these.

"I've had a lot of success hiring people who aren't 100% slam dunk fits. More of that 75%, 80% fit, who might not look close on paper but thrive in the actual position."

Recruiter at a big tech company (P7)

Employers need to calibrate their trust in AI. Trust in AI has to be earned and calibrated.

Without AI, EMPs judge by candidate quality and time to hire. With AI, the system also has to show why someone surfaced and let EMPs override it."Calibrated trust" means trusting AI exactly as much as it deserves: more when it’s usually right, less when it’s often wrong.

What changes with size is who decides and how fast.

At larger companies, recruiters screen and hiring managers decide, so context gets lost in handoffs. Smaller and hourly teams hire at higher volume with fewer people, so speed matters more than detail.

WHAT EMPLOYERS NEED

What do teams need to hire well?

Define what a quality hire looks like

Find strong fits beyond keyword matches

Verified

See why AI picked someone, and correct it

+AI

Final pick: you

Move fast alongside teams without losing control or context

interviews with recruiters and hiring managers, from big tech to pharma and aviation

1,000+

Reddit posts from job seekers and hiring teams

20+

studies on bias and fairness in AI hiring

products compared: LinkedIn, ZipRecruiter, Rippling, Tezi

Our 2 researchers led the literature studies and the Reddit analysis.
We worked alongside them to scope what design could take on.

See full research documents

RESEARCH Synthesis

Why do teams struggle to find high-quality talent: is it the process or the algorithm? Both.

We sorted every pain point by where it starts. Many begin before the algorithm runs, in how the role gets written and assigned to teams.

Cause tree: EMPs struggle to find standouts in a large candidate pool, broken down by JD writing, search and filtering, AI evaluation, and verification.

3 bottlenecks kept showing up in high-volume hiring

3 bottlenecks kept showing up in high-headcount hiring:

3 bottlenecks kept showing up in high-volume hiring: lack of relevant information for AI, lack of visibility & control

The hiring journey

Step 1

Employer intent

The ideal candidate in their head

Step 1

Employer intent

The ideal candidate in their head

Step 2

AI matching

Ranks every applicant

Step 2

AI matching

Ranks every applicant

Lack of understanding

AI can't capture the candidate employers have in mind.

Step 3

Shortlist

Match scores and summaries

Step 3

Shortlist

Match scores and summaries

Lack of transparency

Employers don't know the reasoning behind AI results.

Step 4

Employer review

Decides who moves on

Step 4

Employer review

Decides who moves on

Lack of control

Employers have no way to correct AI results.

Employers still read every resume for screening.

More reliance on AI lead to users falling asleep at the wheel. When AI seems reliable, people stop checking it. We designed for calibrated trust.

When AI is perceived as highly reliable, users become complacent and less likely to examine its recommendations. By deliberately dialing back AI model accuracy, researchers found that AI that is good enough strikes balance between providing useful guidance and encouraging human oversight.

In a field experiment with recruiters, those given a near-perfect AI put in less effort and made worse calls than those given a good-enough one (Dell’Acqua).

Check everythingReread every resume. AI saves no time.

Calibrated trustSpot-check the evidence, override when needed.

Trust blindlyStop checking. Asleep at the wheel.

Source: Dell'Acqua, 2024

Imperfect AI keeps recruiters engaged... when AI seems reliable, people stop checking it.

Synthesized from interviews, Reddit and literature.

DESIGN ASSUMPTIONS

We assumed supply was healthy and matching was broken. Information that were relevant for matching were hidden.

Distrust comes from the interface, not the model’s accuracy.

THE PROBLEM TREE MAP

EMPs told us candidate quality was poor, yet nearly 60% of applicants met the job post’s qualifications. I pushed the team to read that discrepancy as a trust gap problem instead of focusing our efforts treating it as a candidate supply problem.

In marketplace terms, I assumed supply was healthy and the breakdown was in matching. With one-click applications and AI-polished resumes, applicants start to look alike. When people doubt how an automated system decides, they stop relying on it, whether or not it’s accurate (Parasuraman & Riley, 1997). Under 20% of EMPs engaged with AI summaries, which fits that pattern. Building on that assumption, we prioritized giving EMPs ways to shape AI over automating more.

Why I assumed it: candidate quality was poor, yet nearly 60% of applicants met the job post’s qualifications, and Indeed has about 150 job seekers for every employer. The behavior we saw pointed to distrust: under 20% of EMPs engaged with AI summaries, and most reread every resume. One-click applications and AI-polished resumes make applicants look alike, and creating doubt on automated systems, increasing the risk of abadonment.

That assumption set our scope. We didn’t try to bring in more applicants. We helped EMPs judge the resources they had. We believe the success of the designs will be reflected through metrics like resumes reread per hire, instead of applicant volume. If the pool turned out to be weak, these tools would surface that faster.

Velocity comes from delegating routine work to AI. Control comes from verifying and correcting its output. We designed for calibrated trust: reliance that matches what AI can actually do.

Design mission

How might we help hiring managers & recruiters screen faster and stay in control?

Velocity comes from delegating routine work to AI. Control comes from verifying and correcting its output. We designed for calibrated trust: reliance that matches what AI can actually do.

Velocity comes from delegating routine work to AI. Control comes from verifying and correcting its output. We designed for calibrated trust: reliance that matches what AI can actually do.

Design mission

How might we help hiring managers & recruiters screen faster and stay in control?

Ideation

AI compresses sourcing and candidate reachout, giving HM time on judgment: qualifying, interviewing and closing.

Where does hiring slow down most? At the top of the funnel. Our workflow map showed the essential tasks cluster in Find and Screen. Speeding up these steps gives HMs more time with candidates. Job seekers would hear back sooner. EMPs would have more time to judge fit.

Indeed's Hiring Journey Map | Opportunity Evaluation

For ideation, we focused on 2 questions:

DESIGN QUESTION 01

How might we help employers articulate their needs better?

#1 Candidate persona

#2 Outcome check

#3 JD edit suggestions

#4 Role calibration

DESIGN QUESTION 02

How might we give employers transparency and control over matches?

#4 Pool-based qualification

#5 Maybe candidates

#6 Summary with resume citation

#7 Customize AI screening on result

#8 External data enrichment

Opportunities · decisions · RISKS

Pressure-testing concepts for key revenue drivers: large scale hiring teams.

The matching model itself was outside our scope, which left one question: what can design do when the model’s accuracy isn’t in our hands? Our answer was to make AI’s reasoning visible and its results correctable. The biggest untapped revenue, a $675M segment, sits with satisfied HMs and proactive partners, therefore our primary users are hiring managers and the recruiters who work with them.

Growth opportunityEst. Indeed revenue

Struggling hiring manager

$675M

Satisfied hiring manager

Hiring managers

Reactive partner

$470M

Proactive partner

Tactical partner

Collaborative partners

Source: Indeed. Estimated revenue and growth opportunity by employer segment.

The risk

RISKS

Over-Automation on the Wrong Stages

The riskiest move is letting AI interpret a job description with no human judgment. Many job descriptions start from a template and what HMs actually needs often isn't in the post. AI reads the words literally, ranks every applicant on them, and a shortlist built on the wrong criteria looks just as confident as a right one.

AI reads it alone

AI drafts, employer confirms

Job description

Data AnalystFrom a template

5+ years of experienceAdvanced SQLBachelor's degree required

CriteriaSet by AI

5+ yearsSQLBachelor's degree

Coaches junior analystsWhat the manager needs. Never in the post.

Shortlist248

Strong match

Confident, but ranked on the wrong signals

CriteriaConfirmed

5+ yearsSQLBachelor's degreeCoaches junior analysts

Employer edits once, before ranking

Shortlist12

Evidence

Every match points to a resume line

Design Principles to Avoid Over-Automation

HCI research suggests automation should vary by stage. Gathering and sorting information can be highly automated, while high-stakes decisions need people involved. Defining what a good hire looks like is one of those decisions. Principle should be automating repetitive stages and augmented the ones that need judgment.

Design Decisions and Our Trade-offs

Optimizing the top of the funnel

over the whole hiring funnel

When we rated each hiring task by how much AI could help, the high scores clustered in Find and Screen. We explored a little bit of offers and onboarding stage, but that would ignore the distrust problem caused by AI.

Visible automation

over full automation

We automated filtering and verification while giving employers flexibility to see and edit every step of algorithmic matching.

Checkable evidence

over a complete explanation

We couldn't change how accurate the Open AI x Indeed model is. So instead of explaining everything, each claim points to a resume line employers can check in seconds.

Fewer calibration steps

over finer control

Round 2 found the flow redundant. Employers now confirm the criteria once instead of tuning each one.

Testing

Recruiters and managers prefer plain language, visible sources and fewer steps.

Recruiters and hiring managers prefer plain language, visible sources and fewer steps.

We tested in three rounds with 14 participants: co-design (3), concept tests (5) and validation (6).

parsed signalsskills that match your notes

Their own words

"Parsed" and "signals" confused them. Use hiring language.

Where it came from

"Not sure where the summary came from." Show the source of every AI output.

92% match

Evidence ›

A quick read first

They still rank candidates. A match score up top, evidence one click away.

What we changed in round 3 sits with each feature, under User validation.

VISION

FUTURE VISION

MVP Milestone

Phase 1

Phase 2

Phase 3

Phase 4

Phase 1 · Q1Build the core flowRole Calibration, SmartScreen and Resume View with Indeed recruiters.

MVP metrics

AdoptionAI summary use above 20%

TrustShortlist kept on first try · More thumbs up on AI

QualityFewer quality complaints in NPS

OutcomeMore successful hires

Limit: 14 participants, mostly from larger companies. High-volume hourly hiring is untested.

Reflection

Indeed design suggests we think more about…

Indeed Design suggests we think more about…

Helping one hiring team decide is different from automating thousands of workflows. HCI researcher Ben Shneiderman (2022) argues the best AI systems aim for high automation and high human control at the same time. Our concepts built the control first partly because during the research phase, we focused on explainable AI, not the repercussions of automation. Scaling automation is the next step.

Quality is the North Star, not AI.

I kept wondering if AI could take screening off employers' plates entirely. What grounded me was candidate quality. With or without AI agents, good hires bring EMPs back and keep Indeed's flywheel moving.

Designing for Trust Between People and Machines

Indeed's OpenAI model sends up to 20 million personalized job messages a day, each explaining why a role fits. Started applications rose 20%. Our work brings that explainability side for 3.5M+ employers. I think we were grounded in the belief that hiring isn’t just a transaction. Hiring affects people’s sense of self-esteem because it’s the primary way people improve their finances. Access to jobs affect how people are treated in systems from healthcare to housing to long-term security access. This project is one with the most controversial impact I dealt with because I am directly impacting someone’s long-term economic access/opportunities.

Simplify & Scale Automation

Indeed encouraged us to think about automation in the future. Due to time, we assumed automating more on top of distrust, we assumed, would push EMPs further toward rereading every resume. We have a very strong distrust evidence.

Google PAIR recommends increasing automation gradually as trust grows. Our design works by consent: AI proposes and EMPs approve.

PRODUCT VIDEO

Thank you Mark and Meghna at Indeed. To Hongdi, Hongjin and Jessamine for your feedback 🫶🏻

Thank you to Meghna and Mark at Indeed for your feedbac, and to Hongdi, Hongjin and Jessamine.

More work…

See all work

TESTING

Nordstrom

2025

VISUAL DESIGN

Clever

2021