Can an AI agent hire you? What changes for your CV

What actually happened when an AI agent started hiring people?

An AI agent is already paying humans to work — not filling salaried jobs, but hiring for tasks. On 28 September 2026, San Francisco startup Wajo opened its personal agent, Fo, to the public after a month of invite-only testing. Fo has its own phone number, its own email address and single-use virtual payment cards, so it can book, buy and call without borrowing your credentials. When it hits a wall — a phone call to a barbershop or a plumber with no API to talk to — it hands the task to a human executive assistant on Wajo's staff, who receives only what that one task requires and not the user's chat history.

A few days later, Vinod Khosla made the case publicly that Wajo will win the agent market on trust rather than raw capability, pointing at the single-use cards, the separation of identities and the human oversight baked into the design. Wajo's founder, who led engineering at Google DeepMind, framed the human layer as deliberately contrarian: every other personal assistant pretends software can do the whole job, while this one employs people for the parts it can't. On the company's own benchmarks — tasks deliberately seeded with setbacks like sold-out slots and silent vendors — Fo finished 71% without help and stayed inside user-approved limits on 94% of runs.

Treat those figures as a vendor's own scorecard rather than independent testing, because that's what they are. The interesting number is the gap: close to three tasks in ten still need a person. That gap is the whole story for anyone job-hunting. Somewhere in that pipeline, software decides a human is required, picks which human, writes the brief and releases the money — with nobody senior reviewing the choice. Once that pattern spreads beyond errands and into freelance and contract work, the question stops being whether a recruiter likes your CV and becomes whether a machine can read it accurately enough to pick you.

Can an AI agent really hire you without a recruiter involved?

For a salaried role, not yet. For paid work, already. What exists today is an agent that allocates a single task to a person and pays for it, alongside recruiting software that sources, ranks, messages and schedules candidates before any human opens a profile. In Fo's case the assistants are staff, matched to a task by software: the agent decides a person is needed, selects one, briefs them with the minimum information, and closes the loop. Nobody interviews anyone for that phone call. That's hiring in the economic sense — work assigned, work paid — even if it never touches an HR system or a contract.

The corporate side is moving the same direction from a different angle. Surveys of talent leaders heading into 2026 found the overwhelming majority planning to use AI in recruiting, and roughly half intending to add autonomous agents to their teams outright. In practice, most platforms sold as agentic are semi-agentic: they run long multi-step workflows but keep people at the checkpoints. One startup that billed its product as the first fully autonomous recruiting agent abandoned the category entirely earlier this year and became a different company. Full autonomy in hiring is still more pitch deck than production, and anyone telling you otherwise is selling something.

Where this lands for you is less dramatic and more immediate than the headlines suggest. The realistic 2026 setup is a hard filter that eliminates on parsed fields, a semantic ranker that scores you against the job description, and a human who reviews the middle band. A common threshold pattern auto-advances the top fifteen to twenty per cent, auto-rejects the bottom half, and sends the remainder to a person. You can be strong on paper and still never reach that person, because the elimination happened upstream — in extraction, not in judgement.

What signals does an AI agent screening your resume actually weigh?

Agents weigh what they can extract into structured fields: job titles, employer names, date ranges, canonical skill names, seniority, and outcomes attached to numbers. Everything else is decoration. A parser converts your document into rows in a database, then normalises what it found — collapsing "React", "ReactJS" and "React.js" into one skill identifier, mapping your title to a standard level, calculating tenure from your dates. Scoring runs on that normalised table, not on your prose. If the extraction step mangles a title or drops two years of dates, every score downstream is computed from wrong data, and no amount of eloquent writing in the bullets repairs it.

Verifiable outcomes carry disproportionate weight because they're the one thing a semantic ranker can distinguish from boilerplate. "Responsible for improving performance" and "cut p95 API latency from 820ms to 190ms across 14 services" occupy the same line count and score nothing alike. The second gives the model a tool, a scope, a metric and a direction. The same logic applies to scale: team size, budget, ticket volume, headcount supported, revenue closed. Agents reward specificity not because they appreciate rigour, but because concrete nouns and numbers survive extraction and map cleanly onto the requirements pulled out of the job description.

Extraction accuracy is measurable and the spread is brutal. Clean, single-column CVs with standard section headings tend to extract at roughly 90–95% accuracy; infographic layouts, table-heavy designs and multi-column templates often fall to 70–80%. Losing a fifth of your content at random is not a cosmetic problem. Enterprise systems soften this with OCR fallbacks and multiple parsing passes, but you have no idea which system you're facing when you click apply. If you want to know what a parser sees before an agent decides, run the file through a free ATS resume check and look at what comes back.

How do you make your CV legible to an agent, not just a human?

Build the document so a parser can lift every field without guessing. One column, top to bottom. Standard headings — Work Experience, Education, Skills, Certifications — not "What I Bring" or "My Journey". Contact details in the body of the page, never in the header or footer layer, which parsers routinely skip entirely. Dates in one consistent format, MM/YYYY throughout. Body text at 10–12pt in a font that's been around for decades. No tables, no sidebars, no text boxes, no icons standing in for words, and no skill-rating bars — write "Python (advanced)" as plain text instead of drawing four filled circles.

File format is a real decision, not a preference. A DOCX exported from a normal word processor extracts about as reliably as anything, which is why it's the safer bet with older enterprise systems. A text-based PDF parses well with modern applicant tracking systems as long as it's genuinely single-column with no floating elements. An image-based or scanned PDF is the worst outcome available to you: it forces OCR, loses structure and degrades every field. Default to PDF, switch to DOCX when the posting asks for it, and spell acronyms out alongside the full term — "Search Engine Optimization (SEO)" — so either query finds you.

The cheapest test takes thirty seconds: copy the whole CV and paste it into a plain text editor. If the reading order scrambles, if your job titles end up beside the wrong dates, if whole blocks vanish — that's roughly what the machine ingests. Then go further and have the file scored properly. HRLens's free CV analysis returns an overall score out of 100, five category scores including ATS compatibility, and a visual layout analysis showing how the page reads structurally. Paste in the job description too, and the analysis names the skills you're missing against that specific ad.

Where does human judgment still decide whether you get hired?

Humans still decide everything that follows the shortlist — and, quietly, most of what matters. Agents are good at parsing, matching, ranking and recommending; they're poor at assessing whether you can actually do the work. That's why serious employers pair screening with a short skills exercise and keep a person on every adverse decision. The honest summary: machines control who gets considered, people control who gets hired. Optimising only for the machine wins you a seat in a room where a human then judges your reasoning, your examples and how you handle being pushed back on. Both halves need to be true.

Here's my contrarian position: the people losing most to agentic screening aren't juniors with thin CVs — they're senior professionals with gorgeous two-column documents built in a design tool. They've had fifteen good years, they're being filtered out at extraction, and the feedback loop never tells them. A plain, slightly boring single-column CV with concrete numbers beats a beautiful one every time now, and that will feel like a downgrade when you look at it. Make peace with it. The reward for a visually restrained document is that every claim you made actually reaches the ranking stage intact.

The Wajo story also puts a quieter obligation on the table. Reporters testing Fo found the disclosure murky when it rang someone on a user's behalf — and if agents are going to allocate paid work to people, candidates deserve to know when software, not a person, is doing the choosing. Push for that. In the meantime, control what you can: the file. If you're rebuilding from scratch, HRLens's AI CV builder works by chat — paste an old CV or describe the work, refine it in conversation, and export from six templates in PDF or Word.

Frequently asked questions