The whole workflow, end to end

Parsing a CV into structured data and suggesting target roles is useful, and it is where every later stage gets its quality. But the target itself is a decision about someone’s life, and inferring it from history serves people who want continuity and fails everyone else.

Let the tool propose and the person choose. Then hold that choice as the authority for everything downstream.

Capturing constraints at this stage costs ten minutes and prevents most of the frustration that follows. Minimum salary, locations, notice period, what they will not do, and which employers to exclude are all things no system can infer and every later stage needs.

This stage automates cleanly. Searching multiple sources, deduplicating, filtering by eligibility, ranking by relevance and flagging postings that look stale or unserious is all mechanical work done better by a machine.

Keep the rejections inspectable. A screening step that silently discards a category is a failure the candidate cannot see.

This is also the stage where automation is most clearly better than a person rather than merely faster. Nobody checks forty sources every morning, and a role seen on the day it is posted rather than a week later is a materially different application.

• Profile and targeting — assisted, person decides
• Discovery and screening — automated, inspectable
• Documents — assisted, always reviewed
• Submission — confirmed by the person
• Interviews and offers — the person, throughout

Tailoring can be automated up to a draft. What cannot be automated is accountability: the candidate is answerable for every claim, in a conversation where an interviewer may ask about it directly.

Submission gets a confirmation step because it is irreversible and outward-facing. Everything before it can be retried; this cannot, and it arrives under their name.

How the confirmation is presented decides whether it works. A dialog saying "submit this application?" gets clicked; a screen showing the actual document, the actual answers and a highlighted diff against the canonical CV gets read, and the second one is the only version that catches anything.

This is the least discussed stage and possibly the one with the best return. Knowing what was sent where, what needs chasing, which reply is unanswered and which deadline is tomorrow is work people do badly for entirely understandable reasons.

It also compounds over a long search. By week eight a manual tracker has drifted from reality, and the candidate has lost track of which CV version an interviewer holds — which is the exact information they need the morning of a conversation.

Follow-up timing is worth automating specifically because it is so easy to get wrong in both directions. A polite chase after two weeks is normal and frequently forgotten; three chases in ten days is not, and a system enforcing a schedule prevents both.

Preparation automates well: summarising the company, generating likely questions from the posting, recalling what was written in the application, reminding what was said in earlier rounds.

The interview itself is the point of the whole process — a person evaluating a person. Attempting to automate participation misunderstands what is being assessed, and is likely to be treated as dishonest.

The preparation pack is more valuable than it sounds, because the information is genuinely hard to reassemble by hand. The exact CV that was sent eleven weeks ago, the free-text answers written at the time, and what the first-round interviewer asked are all things a candidate would otherwise walk in without.

A tool can supply market context, list the terms worth clarifying and help think through a counter. That is genuinely useful, and it is the limit.

Negotiation is a conversation with a future employer where the relationship starts. Automating it damages the thing it is trying to optimise, and nobody should hand it over.

The most useful thing a tool can do here is unglamorous: list what to ask about beyond salary. Notice period, start date, equity terms, review cycle, remote policy in writing — people forget these under pressure and regret it for years, and a checklist is entirely sufficient.

Automation works where the task is mechanical, reversible and invisible to the employer. It fails where the task is a judgement about someone’s life or a communication that someone will read as personal.

A tool claiming to automate all of it is either describing only the first half or is proposing to send unreviewed material under a candidate’s name. Both are worth knowing before adopting it.

Three questions separate them quickly: what does it do without asking, what does it decide on my behalf, and what happens when it gets one wrong. A product with clear answers has thought about the boundary; one that finds the questions unusual has not drawn it at all.