Hiring Will Be Rebuilt for AI

The modern hiring funnel was designed for a world in which writing an application and reading it were expensive human activities. Generative AI has made both cheap.

Candidates can find roles, tailor documents and answer application questions at far greater speed. Employers can summarise, rank and reject applications at similar scale. Each side is responding rationally. Together, they create a race to the bottom: more applications create more screening, which lowers the expected value of any one application and encourages still more volume.

The pressure is visible across several independent datasets. Greenhouse reports that applications per job rose from 116 in 2022 to 244 in 2025. The Institute of Student Employers reports 86 graduate applications per vacancy in 2022/23 and 140 in both 2023/24 and 2024/25. Gem's benchmark shows applications handled per recruiter rising from 925 in 2021 to 2,479 in 2024.

Candidate-side research points in the same direction, but is harder to compare cleanly. NACE found that graduating students averaged 17 applications in 2023 and 22 in 2024, while the share receiving an offer before graduation fell from 56% to 45%. Its following survey reported medians rather than averages: 6 applications for the class of 2024 and 10 for the class of 2025. Those pairs show increasing effort within their own definitions; they should not be joined into one line.

Ashby saw application volume rise too: across roughly 14 million applications, applications per business role increased 207% between January 2021 and January 2024. Huntr users recorded 602,260 applications and 29,676 interviews from July 2025 to July 2026, a recorded interview rate of about 4.9%. In Q2 2026, users logging 100 or more applications recorded interviews against 2.86% of applications.

The important point is not that AI alone caused every increase. It is that AI destroys the old process's remaining constraint. A document exchange cannot produce reliable signal when both document production and document filtering approach zero marginal cost.

The current process still moves documents through a pipeline: a candidate uses AI to write an application; an applicant-tracking system stores it; another model screens it; a recruiter decides what survives. The technology at each end is new, but the interface between them is not.

An AI-native process will connect two persistent knowledge systems instead. A candidate agent will understand the person's history, goals, constraints and evidence. An employer agent will understand the work to be done, the team, the environment and the evidence that predicts success.

Their first task will not be to produce and judge another cover letter. It will be to decide whether a conversation is worth creating.

This changes the optimisation target. The current funnel maximises completed applications and filtered applicants. The future system should maximise mutually wanted, evidence-backed conversations.

The agents can discover, compare, explain and recommend. They should not make irreversible commitments. A person still controls identity disclosure, interviews, compensation decisions and offers; an employer remains accountable for its criteria and hiring decision.

The CV remains useful. It is concise, familiar and good at communicating a career history to another person. It is also incomplete, lossy and quickly outdated.

A candidate's primary record will become a richer career knowledge base: part master CV, part story bank and part verified portfolio. It will be designed for an AI system to reason over while remaining easy for the person to inspect, correct and delete.

It should contain roles, dates and quantified outcomes; projects and the decisions behind them; work samples; stories that demonstrate judgment and working style; goals, motivations, constraints and non-negotiables; and the source, date and permission attached to important facts.

The agent can build this record from documents, connected tools and conversation. A relaxed, agent-led interview can draw out what people rarely write down: why they made a decision, where they struggled, what they learned and the conditions in which they do their best work.

Provenance matters. The system should distinguish a verified outcome from a candidate statement and an agent inference. It should expose contradictions rather than quietly smoothing them away. The owner must be able to see what will be shared before anything leaves the system.

This is not a longer CV. It is a living model of the candidate that can generate a CV, profile or introduction when a human needs one.

A job search is currently treated as a temporary campaign. The person decides to move, reconstructs several years of work, searches listings and starts applying. Most of the useful context disappears when the campaign ends.

A career agent can search continuously. With permission, it can speak to employer agents and ask a simple question: is there a genuinely better opportunity for this person?

The first comparison should be anonymous. An employer does not initially need the candidate's name, photograph or current company. The candidate agent can share only the evidence and constraints needed to establish possible fit. Each side receives an explanation of why the match may be worth considering. Identity is revealed only when both parties express interest and the candidate approves it.

This resembles a trusted introduction more than an application. It protects people who are quietly exploring, reduces irrelevant inbound volume for employers and makes consent part of the system rather than a privacy toggle buried inside it.

Continuous discovery should not mean continuous interruption. The agent needs a high bar for surfacing an opportunity, a clear explanation of what changed, and settings that reflect whether its owner is actively looking, passively open or unavailable.

Interviews select partly for the ability to interview. They reward recall, confidence, polish and storytelling. Those qualities may matter, but they are not a complete proxy for performance in the job. Generative AI and invisible interview assistants will make the gap harder to ignore.

Interviews will survive, but their purpose should become narrower and clearer: understanding motivation, communication, values, working relationships and whether both sides want to work together.

Evidence of ability should come from the work itself. For many knowledge roles, candidates can enter a transparent, time-boxed assessment containing the tools and information they would normally have at work. They solve a realistic problem. The employer assesses the output, the reasoning, the trade-offs and how the candidate used the available tools.

AI use should not automatically count as cheating. If AI is part of the real job, the assessment should show whether the person can use it well. The relevant question is whether the candidate can exercise judgment, verify outputs and produce good work.

These assessments must be bounded, comparable and clearly separated from unpaid productive work for the company. Recording and analysis must be transparent and proportionate. The goal is evidence, not surveillance.

Getting hired is one moment in a longer relationship.

After someone starts, their career agent can help capture achievements while the context is fresh. It can keep their story bank and work evidence current, prepare them for performance reviews, benchmark compensation, identify missing skills and suggest projects or relationships that open future paths.

It can also help them judge the employer. Public reviews, professional networks and forums contain useful information mixed with selection bias and noise. The agent can gather that evidence, explain its limitations and relate it to what the person values.

When a better internal or external opportunity appears, the agent already understands the candidate. There is no frantic reconstruction and no cold start. The next move begins with current context.

This changes the product's incentive. A tool paid only when someone applies or gets placed is encouraged to create transactions. A trusted career agent should be useful when the right decision is to stay, develop and wait.

CVs, job boards and applicant-tracking systems will not disappear overnight. Legacy interfaces survive while a replacement earns trust, fills coverage gaps and integrates with the old one.

CVs will become generated views of the underlying career knowledge base. Job boards remain useful where employer-agent coverage is incomplete. ATSs continue to handle compliance and workflow, but gradually lose their position as the primary interface between a person and an opportunity.

The long-term shift is more important than the transition period: the system of record moves from documents and application rows to persistent, permissioned models of candidates, roles and outcomes.

The best hiring system will not produce the most applications, the fastest rejections or even the most introductions. It will find the person most likely to perform well, give both sides credible evidence and avoid making either endure a horrible process to get there.

The best career agent will not just help someone get hired. It will help them get what they want from a career: more money, fulfilment, flexibility, better work or some combination of them, with less of the repetitive labour that currently surrounds every move.

That future will not come from adding more AI to the existing funnel. It will come from replacing the funnel with a system built for AI from the beginning.