Finding Your Next Job, with AI

After 45 the search is a targeting problem, not a volume problem — a method, a folder, and a first week you can start

Author

Behzad Samadi

Published

September 10, 2026

A job search after 45 is not the search you ran at 30, and the reason is not the one you have been told.

The unemployment rate for Canadians over 55 sits at about 5%, lower than for everyone else. That number hides the two things that actually go wrong. The first is duration: a quarter of all unemployed Canadians have now been out of work for 27 weeks or more, up from 17% before the pandemic, and the older you are the more likely you are to be in that quarter. The second is the pay cut: among displaced Canadians aged 50–75 who found work again, roughly half took a cut of 25% or more. Workers 45 and older accounted for something like 40% of the rise in Canadian unemployment between mid-2024 and mid-2025. The search after 45 is slower, and the landing is lower.

That reframes what an AI is for. It is not for sending more applications — volume is the one thing the market already has too much of. It is for two harder jobs: shortening the search, and defending the pay level. Both come down to the same move — making twenty-five years of work legible to a machine that reads first and to a person who reads second, without letting the machine invent any of it.

This article is the method. It is the companion to Planning to Age Well, with AI and it uses the same habits: plan in ratios, not identifiers; keep the plan as documents whose every version you keep; measure. Articles that follow take each step further.

NoteThis is a method, not advice

Nothing here is legal or employment advice, and none of it knows your situation. Before you sign a severance agreement, a release, or an offer, talk to an employment lawyer — most offer a fixed-fee review. Before you make a decision about the shape of your life, talk to a person who knows you. Use what follows to arrive at better questions and a first draft, and treat the rest as yours to check.

Two readers, opposite tastes

Every application you send now has two readers, and they want opposite things.

The first reader is a model. On the employer’s side, applicant-tracking systems have grown a ranking layer — Workday’s HiredScore, LinkedIn’s Hiring Assistant, Indeed’s Talent Scout — and in a 2026 study of 2,245 real résumés rewritten by seven different language models, a model asked to judge picked the machine-rewritten version about 97% of the time, even where human graders had rated the original as clearer. Models prefer machine prose, and they prefer their own machine prose most of all.

The second reader is a person, and the person is sick of it. In a 2026 survey of over 800 hiring managers, 80% said they reject generic AI output; the tells are “proven track record”, “detail-oriented professional”, “I am writing to express my interest”, and a letter with no numbers and nothing specific to the company. One recruiter described eleven consecutive cover letters that opened “I am thrilled to apply.” Sixty-three percent of the same managers accept AI-assisted writing that is clearly personal.

There is a third finding, and it is the one this whole method rests on. When AI cover-letter tools became widely available on a large job platform, callbacks initially rose for the people who used them. Then the signal decayed: within a year the relationship between how well a letter matched the posting and whether it got a callback had fallen by half, because employers stopped trusting the letter and shifted their weight to prior work history. The one thing a model cannot manufacture is a long, specific, verifiable record — and that is the one thing you have more of than anyone under 40.

So the rule: AI for structure and vocabulary, aimed at the machine; your own voice and specifics, aimed at the person; and every claim traceable to something you actually did. The three plans below are that rule, applied.

What the machine actually does

Two myths are worth killing before you build anything on them.

The first is that the tracking system rejects your résumé for missing keywords and that three-quarters of résumés are never seen by a human. Neither is true. In recruiter surveys, around nine in ten say the system does not auto-reject on content, and the “75%” figure has no source anyone can find. The only genuine automatic filters are knockout questions — work authorization, a licence, location, a salary range — and most recruiters use them. A missing keyword makes you unsearchable, not deleted.

The second myth is that formatting does not matter. It matters more than keywords. The tracking system’s silent failure is parsing: two-column layouts, tables, text boxes, headers and footers, and PDFs that are really images come out of the parser scrambled or blank. A 2008-era two-column résumé with your contact details in a sidebar may be arriving as a page with no phone number on it. Plan three deals with this; the fix is a single column.

What has changed is the ranking layer on top, and ranking is where age leaks in: graduation dates, tenure length, “years of experience” fields, and the vocabulary of tools that are no longer current. Nobody has to type your age for a model trained on past hires to infer it. Age discrimination in hiring is illegal in Canada under the Human Rights Code whether a person or a model made the decision, and in the United States a nationwide collective action — Mobley v. Workday — is proceeding on exactly this theory, with the software vendor named directly. None of that helps you this week. What helps is knowing where the leak is and closing it.

The line in the posting

If you are looking in Ontario, you have a lever nobody has written up from the candidate’s side. Since January 1, 2026, an employer with 25 or more employees must state in a public posting whether it uses AI to screen, assess or select applicants; must post a salary range; must say whether the vacancy is real; may not require “Canadian experience”; and must tell anyone it interviews whether a decision was made, within 45 days.

Read a posting for those lines before anything else. The AI-disclosure line tells you whether to write for a ranking model first (a clean parse, the posting’s exact vocabulary) or for a person first. The salary range tells you whether to apply at all. The “is this a real vacancy” declaration is your defence against the 18–22% of listings that are estimated to be ghost jobs. And the 45-day rule turns silence into information: a posting that goes quiet past 45 days is a data point, not a mystery.

What AI is good at here — and what it is not

The failure mode is the same one as in financial planning: trusting it for exactly the wrong things.

It is good at:

  • Turning a career into structure. Twenty-five years of work live in your head as stories. A model is very good at interviewing you and turning the stories into an inventory: what you did, for whom, what happened, what it proves.
  • Translation. The thing you called “ran the plant floor” is called something else in a 2026 posting. A model can map your evidence onto the posting’s vocabulary — and, told to, will say plainly which requirements you do not meet rather than papering over them.
  • Research. What the company announced last quarter, who its competitors are, what its customers complain about, what the role pays in Ontario according to posted ranges and public wage data. Closing that gap is where most of the value is, and it involves no writing on your behalf at all.
  • Rehearsal. A model that has the posting and your inventory in context can ask you the questions you will actually get, one at a time, and push back on vague answers. Voice mode makes this a conversation rather than typing.
  • Adherence. A twenty-minute weekly review with your search’s own files in front of it is the cheapest way to not quietly stop.

It is not good at:

  • Applying for you. Auto-apply tools breach LinkedIn’s terms — the platform reported flagging 23.5 million automated sessions in a single quarter of 2026 — and they answer the knockout questions with generic filler. A 25-year LinkedIn profile is your professional identity; it is not worth risking for volume you do not need. The two most popular open-source job-search projects both refuse to submit applications on your behalf, on principle.
  • Knowing what you did. Left alone, a model will improve a bullet by inventing a number. For someone with a long record, an invented metric is the most expensive sentence in the document — it is the one a background check or a second interview will find.
  • Deciding whether a step down is right. It will produce a confident paragraph either way.
  • The phone call. Referred candidates are hired at something like ten times the rate of cold applicants. A model can draft the message; it cannot be the former colleague who sends it.
  • The interview. This is where machine-polished résumés die. Employers increasingly report that claimed skills are harder to verify and are moving weight from the résumé to live work — which, if you can do the job, is the best news in this article.

Giving an AI your career

The funds plan has a rule for pasting financial details into a hosted model: plan in ratios, not identifiers. A career needs the same rule, and for the same reason — anything you paste into a consumer chat model has left your machine, and the consumer tiers of ChatGPT, Gemini and Claude train on conversations by default unless you turn it off. Turn it off (in ChatGPT: Settings → Data Controls), and know that the switch is not retroactive.

Then translate before you paste:

  1. Employers become codes. “Regional operations manager at Maple Leaf Foods, 2014–2022” → “Ops_Manager at Employer_B (national food manufacturer, ~8,000 staff), 8 years.” The scale and sector are the signal; the name is an identifier.
  2. Compensation becomes a multiple. Your target salary is 1.0×. Your last salary is 0.9× or 1.2×. Never the number.
  3. Some things are never pasted. Your SIN, your address, the names of referees, anything covered by a severance agreement or an NDA, and the real reason for a gap if it is medical or caregiving — a gap can be “family responsibilities, 2023–2024” in every document and every prompt without a further word.

For the one step where the whole career goes in — the evidence interview below — a local model on your own laptop is a reasonable alternative. An eight-billion-parameter model running under Ollama or LM Studio is slower and less polished than a hosted one, and entirely adequate for asking you questions and writing down the answers. Nothing leaves the machine.

Plan one: the evidence

The goal is a single file that holds everything you have done, in enough detail that every later document can be traced back to it. It takes an afternoon, and it is the afternoon most people skip.

Do not start from your old résumé. Start from a conversation. Give a model the sanitized timeline — roles, sectors, years — and have it interview you, one role at a time, for the things a résumé leaves out: what was broken when you arrived, what you changed, what the number was afterwards, who noticed.

A prompt that gets a useful answer — run this locally if you can:

I am building an evidence file for a job search. Below is my career timeline with employers replaced by codes. Interview me about it one role at a time, starting with the most recent. For each role, ask me — one question at a time, waiting for my answer — what the situation was when I started, what I did, what changed as a result, and what evidence exists for that (a number, a document, a person who would confirm it). Push back when my answer is vague. When we finish a role, write it up as a Markdown section with a dated list of bullets, and mark any bullet whose evidence I could not name with [UNVERIFIED]. Do not add anything I did not say.

The output is evidence.md. Save it. It will be the most-read file in the folder, and the model’s job from here on is never to write from nothing — only ever to select and rephrase from this.

Two things the file will show you that the old résumé did not. The first is how much of your strongest evidence is more than ten years old; that is useful to know before a posting asks for “recent experience with”. The second is what you actually enjoyed — the entries you had the most to say about — which is the honest start of the question of what to look for.

Plan two: the target

Volume is the market’s problem. Targeting is yours. The single most useful habit in the open-source job-search tools — the ones people wrote for themselves, and used — is a written evaluation of every posting before deciding to apply.

A prompt that gets a useful answer:

Here is a job posting and my evidence file. Do not write anything for me yet. Evaluate the posting on eight lines, one sentence each: (A) what the role actually is, stripped of the posting’s language; (B) each stated requirement, marked strong / partial / missing against my evidence, with the evidence line cited — do not invent evidence; (C) the seniority the posting implies, and whether mine is above, at, or below it; (D) what the salary range, if posted, is as a multiple of my target of 1.0×; (E) whether the posting discloses AI screening, and what that changes about how the résumé should be written; (F) the three questions this employer is most likely to ask someone with my history; (G) anyone in my outreach file who could plausibly know this team; (H) a fit score from 1 to 5, and the one reason it is not a 5.

Save the answer as fit.md in the application’s folder whether or not you apply. Three months of these is a dataset: it will tell you which kinds of roles you keep scoring 4 on, which is a better description of what to look for than anything you would have written on day one.

Two things about where to look. Postings on the big boards arrive with hundreds of applicants — the average posting now draws about 244, up from 116 in 2022. The same roles sit on employers’ own career pages and the tracking systems behind them, usually for days before they are syndicated, and there are search engines that index those pages directly. And the roles that suit a long career — interim, fractional, “we need someone who has done this before” — are disproportionately filled by referral, which is plan four.

Plan three: the two documents

You have one master file and a posting you have decided to pursue. Two documents come out.

The résumé

Tailoring, done honestly, is selection and translation — never addition. The model chooses which evidence to lead with and phrases it in the posting’s vocabulary where the evidence supports the phrase. It does not get to improve a number.

A prompt that gets a useful answer:

Here is portfolio.yaml, my evidence file, and the posting. Propose changes to the YAML only. Rewrite the summary in three sentences for this role. Reorder the bullets under each role so the ones relevant to this posting come first, and rephrase them using the posting’s own terms only where my evidence supports the term. Every bullet must keep its source: line. Keep dates on all roles. Cover the last fifteen years in full and summarize anything earlier in one line per role. Remove education dates. Do not add, merge, or quantify anything that is not in the evidence. Show me the change as a diff.

Read the diff. Then render both layouts. The parser layout is single column, plain text, dates on roles, no education dates — the last two are the standard advice for older candidates, and they cost nothing. Upload the parser PDF to a free résumé-parser test before you send it anywhere; seeing what a machine makes of your PDF is the most useful sixty seconds in this article.

On the older question — whether to hide the length of your career — this article’s position is that concealment is a tactic, not a strategy. Trimming to fifteen years and dropping graduation dates removes the cheap age cues. Beyond that, you are choosing between two arguments: I could be a 5–7 year person for you, or I am the judgment you cannot get from a 5–7 year person plus a model. The second is true more often than people admit, and it is the one the market is drifting toward. A later article treats it properly; for now, keep both versions of the summary, and measure.

The note

Cover letters are the document employers have stopped trusting, so write a short one that could not have been written about anyone else.

A prompt that gets a useful answer, in two steps:

Step one. From the posting and the company’s own site only, list five specific facts about this company or this team — a product, a number, a recent change, a stated problem. No adjectives.

Step two. Draft 150 words that connect two of those facts to two entries in my evidence file, cited by line. Do not use any of these phrases: proven track record, detail-oriented, I am writing to express, passionate, leverage, delve, thrilled, excited. End without a summary.

Then add one sentence only you could write — a name, a place, a thing that happened. In the study of AI-written letters, time spent editing the draft was the best predictor of a callback. That sentence is the edit.

Plan four: the phone call

Referred candidates make up a small fraction of applicants and a large fraction of hires; by most estimates they are hired at around ten times the rate of people who came in through a board. After 25 years you have the one asset that makes this work: people who have seen you do the job.

The model’s role here is small and specific. It reads evidence.md and helps you list, for each role, the people who were there — then you decide who to write to. It drafts the message. You send it, and you make it a request for twenty minutes of information, never a request for a job.

A prompt that gets a useful answer:

Draft a 60-word message to a former colleague — we worked together on [project] at Employer_B in [year] — asking for twenty minutes to hear how hiring works at their current company now. No request for a job, no résumé attached, one specific question I could ask them, and a line that makes it easy to say no.

Log every ask and every answer in outreach.md. The number of conversations you have had in a week is the leading indicator of this search; the number of applications is a lagging one.

Plan five: the interview

The interview is where the machine’s advantage ends and yours begins, and it is also where a résumé polished beyond what you can defend collapses. Rehearse the questions someone with a long career actually gets — not the ones in the generic lists.

A prompt that gets a useful answer — use voice mode if you have it:

You are interviewing me for the attached posting; my evidence file is attached. Ask me one question at a time, wait for my answer, then tell me whether the answer was specific — a situation, an action, a result with a number — or vague, before asking the next. Include, at some point: why I am leaving or left; the gap in 2023; whether I would really take a role at this level; whether I am overqualified; what salary I expect, which I will answer as a multiple of my target; and how current my tools are. Grade for specificity, not for tone.

Record your answers in the application folder. The point is not a script; it is that the third time you answer “why would you take a step down,” you will have an answer you believe.

Two more things. If a posting disclosed AI screening, expect an automated first interview and treat it as a parsing problem: short, concrete, keyword-bearing answers, no jokes. And if an employer offers a paid trial, a take-home, or a day on site, say yes. Work trials are coming back precisely because résumés have stopped being believed, and they favour the person who can do the job on Monday.

Keep the search alive

A weekly ritual, twenty minutes, with the folder open:

  1. Count the funnel: applications sent, responses, first interviews, later rounds, offers. Write the five numbers in log.md with the date.
  2. Tag each application with which steps used a model — evaluation, résumé, note, rehearsal — and watch whether the tagged ones respond at a different rate. This is the only way you will ever know whether any of this is working for you, as opposed to for someone in a survey.
  3. Change one thing at a time, for two weeks, then look again. Two summaries, two layouts, two sources of postings — one experiment per fortnight.
  4. Count conversations, not just applications.
  5. Save the week’s version — a commit, or a dated copy.

For calibration: across ten million applications on one platform, about 3% of applicants were interviewed and roughly one in 180 was hired; about 44% of applicants who hear back do so within two weeks. Your numbers will differ. The point is to have them.

The last item belongs to the mind plan. A long search is a cognitive and mood risk in its own right — a quarter of unemployed people over 55 have been out for more than six months, and the routine of work goes with the work. The folder is the routine: a file that changes every week is visible progress when nothing else is. Add one standing social commitment that is not about the search, and one genuinely new skill that you can name in an interview. Both count twice.

A first week

  • Monday. Make the folder — git init it if you use git, or add a history/ subfolder for dated copies if you don’t. Turn off training in whichever chat model you use, or install a local one.
  • Tuesday and Wednesday. The evidence interview, one role at a time. Save evidence.md.
  • Thursday. Move your résumé into portfolio.yaml, one bullet at a time, each with a source: line. Render the parser layout. Upload it to a parser test and look at what came out.
  • Friday. Evaluate three postings with the eight-line prompt. Apply to none of them yet. Write to two people for twenty minutes.
  • Saturday. Rehearse “why are you leaving” until you believe the answer.
  • Sunday. First entry in log.md. Five numbers, most of them zero.

That is the whole system. The articles that follow take each step further — the evidence file and the local model, privacy, what the parser reads, the two readers, the overqualified question, why never to automate the applying, the phone call, the interview, and how to measure — but nothing in them changes the shape of the folder you made on Monday.

Sources worth reading directly