Skip to content

What an AI job application assistant should and shouldn't do

Language models are good at drafting prose and bad at knowing facts about you. A trustworthy assistant uses AI for the first job and keeps it away from the second.

AI5 min readUpdated

On this page

Ask a language model to fill out a job application and it will. Asked how many years of Kubernetes experience you have, it will give a confident number whether or not you have ever run a cluster. That is the core problem with AI job application assistants: the model is fluent, and fluency looks like knowledge.

This guide splits an application into the parts where AI earns its place and the parts where it should never answer on its own, and gives you a way to judge any assistant against that split.

Three kinds of application questions

Almost every field on an engineering application falls into one of three groups. They need very different handling.

Question types and who should answer them
TypeExamplesRight source
Identity and contactName, email, phone, LinkedIn, GitHub, portfolioYour confirmed profile only
Structured factsSponsorship, work authorization, notice period, salary expectations, years of Go, relocationYour facts or answers you saved; blank if unknown
NarrativeWhy this role? Tell us about a project you're proud of. Cover letterAI draft grounded in your facts, then your edit

The first two groups have exactly one correct answer, and it lives in your head or your records. A model cannot improve on it, only get it wrong. The third group has many acceptable answers, and that is where a model's ability to produce a clean first draft is useful.

Where AI actually helps

  • Narrative answers. A first draft of why you want to join Northbeam Labs, connecting your payments work to their billing platform, saves real time, as long as every claim traces back to something you did.
  • Job analysis. Reading a long job description and separating hard requirements from nice-to-haves is a reasonable language task. It's even better when the output is checked against your facts.
  • CV emphasis. Reordering bullets so your observability work leads for an SRE-leaning role, without adding anything new.
  • Rephrasing. Tightening a 300-word answer to fit a 150-word limit.

Where AI fails, and why it matters

Models fill gaps with plausible text. On a job application, plausible and wrong is the worst combination. A few failure patterns show up again and again:

  • Inflated years. The posting asks for 5+ years of TypeScript, so the model writes 5, because that makes the answer fit.
  • Invented employers or projects. A narrative answer says you led the migration to Kafka at your previous company, when you only consumed from an existing Kafka cluster.
  • Wrong legal answers. Sponsorship and work-authorization questions answered from the job's preferences rather than your situation.
  • Borrowed numbers. A line about cutting latency by 40% appears because it sounds like a typical CV bullet.

Grounding: the test that separates good assistants from risky ones

The useful question to ask any AI assistant is: where did this sentence come from? A grounded assistant can point to a source for each factual claim: a line on your CV, a fact you confirmed, a saved answer. An ungrounded one can't, because the claim came from the model.

Consider a draft for Alex Rivera applying to a platform role at Harbor & Vale. Alex's confirmed profile says: 6 years of Go, led the move from EC2 to EKS at Paperplane, on-call rotation lead for 18 months. An ungrounded draft might say Alex led a team of 12 through a migration that cut infrastructure costs by 35%. None of those numbers are in the profile. A grounded draft sticks to the record: at Paperplane, Alex led the move from EC2 to EKS and ran the on-call rotation for 18 months, which is the kind of work the platform team describes. Every claim in it traces to a fact.

How to evaluate an assistant in ten minutes

  1. Pick a real job posting that asks for a skill you do not have.
  2. Run the assistant on the application and look at the question about that skill. Did it claim experience? Leave it blank? Ask you?
  3. Look at a narrative answer. Highlight every number and every named technology. Can you find each one on your CV?
  4. Check the sponsorship and work-authorization answers against your actual status.
  5. Find out whether the tool submits anything itself, or whether you click submit.
  6. Read its data policy: does your CV train its models? Can you export and delete everything?

If an assistant fails step two, stop there. A tool that invents one skill will invent others.

Using AI drafts well

Even a grounded draft is a starting point. Read it out loud. Replace generic phrases with something only you would say: a specific incident, a trade-off you made, a reason you care about this company's problem. Recruiters read a lot of lines about being passionate about scalable systems. A sentence about the time you cut Quillstack's CI time in half by splitting the test suite (if that happened) is worth more than a paragraph of that.

Before submitting, run a final check across the whole form: narrative answers should agree with structured ones, and both should agree with your CV. There's more on keeping versions consistent in the CV tailoring guide, and on data handling in security.