Autonomous job application

Debates about whether AI should apply to jobs for you go in circles because both sides mean different things. Borrowing the levels framing from vehicle autonomy dissolves most of the disagreement.

Each level is a real product that exists today. They differ in one respect that matters more than any feature: who is responsible when it is wrong.

The framing is useful precisely because it stops the argument being about capability. Nobody disputes that a system can fill a form; the disagreement is about which level is appropriate for something carrying your name, and that is answerable.

• Level 0 — you do everything; tools only search
• Level 1 — the system suggests roles; you decide and apply
• Level 2 — it drafts a tailored application; you review and submit
• Level 3 — it submits, but pauses for approval on anything unusual
• Level 4 — it runs unattended and reports afterwards

Almost every deployed product sits at level 2, and the ones claiming level 4 are usually level 3 with an optimistic description. That is not timidity; it reflects where the consequences change character.

Up to level 2 a mistake costs you time. From level 3 a mistake is a permanent record: a wrong answer submitted to a company you wanted to work for, attached to your name, with no undo. The asymmetry is what keeps the line there.

It is worth noticing that the asymmetry is about you rather than about the vendor. The product experiences a failed submission as an error rate; you experience it as one specific employer holding one specific bad impression, which is not a number that averages out.

The technical obstacle is not writing the application. It is that application forms are unbounded in variety — free-text questions about motivation, salary expectations, eligibility declarations, occasional legal attestations.

A system answering "why do you want to work here" without you is inventing a motivation on your behalf. A system answering a right-to-work question wrongly has made a false declaration in your name. Neither is a model quality problem that improves with a better model.

That distinction is the crux. Problems that better models solve get easier over time; problems where the system lacks the information — because it is about you, your circumstances or your intentions — do not get easier no matter how capable the model becomes.

The pattern that works is full autonomy on the mechanical parts and a hard stop on the consequential ones. Search, filter, score, tailor, pre-fill — all unattended. Anything that constitutes a statement about you, or that cannot be retracted, waits for a person.

This is exactly the design that emerged in every other agent domain that touches the real world, and for the same reason: limit what the system may do rather than trusting it to decide well.

Pre-approval is the useful refinement between gating everything and gating nothing. Deciding once that the agent may submit to these employers, at this seniority, up to this many a week, with this document, makes your judgement reusable rather than requiring it per application.

• Unattended: discovery, filtering, scoring, drafting, form pre-fill
• Gated: final submission, free-text answers about you, any declaration
• Never: eligibility or legal attestations, salary commitments

A level 4 claim implies the system can recognise when it is out of its depth, and that is the capability least often demonstrated. An agent that handles ninety per cent of forms and confidently mishandles the rest is more dangerous than one that handles seventy and stops.

Ask what it does with a question it has not seen. Stopping and asking is correct; guessing plausibly is the failure that produces an application you would not have sent and will not find out about.

The same applies to verification after the fact. A system that cannot show you the exact text it submitted has not merely omitted a feature — it has removed your ability to know whether the ninety per cent claim is true.

Discussions of autonomy focus on whether the application is good. The larger risk is what a bad one does to a relationship you wanted, because an employer does not see a system error — they see a person who applied carelessly.

That cost is concentrated rather than averaged. A system with a two per cent error rate across two hundred applications has produced four bad impressions, and if one of them is at the company you most wanted, the aggregate statistic is no comfort at all.

It also persists. Applicant tracking systems retain history, so an application submitted badly in March is visible to the same recruiter in November, attached to your name, with no context explaining that a tool filled the form. Nobody reading it knows or cares that you were not the one typing.

• Errors are concentrated on specific employers, not spread thinly
• Applicant tracking history persists across future applications
• An employer sees carelessness, not a system failure
• Applying to several unrelated roles at one company is visible to them
• A withdrawn application is still a record that you applied

If you are evaluating a product that claims to apply for you, the useful questions are about failure rather than capability. What happens when a form has a question it has not seen? Can you see exactly what was submitted, afterwards? Can you revoke access cleanly?

A product that cannot show you the exact text submitted in your name is not one to grant autonomy to, regardless of how well it performs when everything goes right.

Start it narrow regardless of the answers. Let it run at a low weekly cap on roles you are exploring rather than ones you want, read everything it produced for the first fortnight, and widen only once you have seen how it handles something unusual.