Which Roles Absorb a Career Changer Who Doesn't Know Your Field

Where a wrong answer comes back, and where it never does

Roles where a wrong answer surfaces fast and cheaply can absorb a career changer who does not know the field yet. Support work with a resolution the customer confirms, coding against a test suite that fails loudly, analysis a reconciliation checks the same afternoon, first-draft work a named reviewer signs before it moves.

Roles where a plausible wrong answer travels for weeks before anyone opens it carry the gap instead, and the person carrying it is usually not the hire.

Deciding which roles actually need AI skills right now is the prior question, and the sorting here starts after that answer, on the roles already named.

Seniority is not what separates the two lists.

Customer support, on the first list, has a measured case. Across a staggered rollout of a generative AI assistant to 5,179 agents at one software firm, access raised issues resolved per hour by 14% on average. The average splits by experience: a 34% gain for novice and low-skilled agents, and almost nothing for the experienced ones ¹.

Look past the headline at what that setting supplied. The outcome was issues resolved per hour on problems that have correct answers, and a wrong answer met a customer within minutes.

The novices gained in a setting that had both, using an assistant trained on that firm's own transcripts.

Now take a task with neither: a pricing memo, a coverage determination, a schema change nobody reviews, a figure in a client deck. A confident wrong answer there is not caught in minutes but by whoever acts on it.

Add one column beside the task list to do the sorting

Take the role task by task, and it sorts itself. Write out what the job actually does, then put one column beside it: how long a plausible wrong answer survives before something outside the person catches it. Minutes, if a test fails. A day, if a reconciliation runs. A quarter, if the first real reader is an auditor. Never, if the only available check is knowing the field.

The task list does not have to be written from nothing. O*NET, the US Department of Labor's occupational database, publishes task lists occupation by occupation. The entry for financial and investment analysts, 13-2051.00, opens with advising clients on aspects of capitalization, such as amounts, sources or timing, and goes on to analyzing the performance of companies facing financial difficulties ².

Strike what your job does not do, add what it does.

Federal selection guidelines ask for the same decomposition, for a different reason, when an exercise is defended by its content. Under the 1978 Uniform Guidelines on Employee Selection Procedures, a selection procedure can be supported by a content validity strategy to the extent that it is a representative sample of the content of the job. The same section calls for a job analysis of the important work behaviors ³.

One page starts both jobs: placing the hire, and the job analysis behind an exercise built from the same list. Whether that is enough to defend the exercise is a question for counsel.

Turning a list into stages is its own move, and every requirement needs a stage that tests it. The latency column does the sorting first. A role whose tasks all sit in the minutes-to-a-day band can take someone who does not know your field, whatever the title says.

A role with three tasks in the never band cannot, even when the other twelve look junior, unless someone who knows the field checks those three. Count the never-band tasks before you count the years of experience the posting asks for.

Speed transfers; knowing which sentence is wrong does not

Speed and structure carry. Judgment about your field does not, and the tool makes its absence more expensive. Someone with real AI habits arrives able to frame a task, get a usable first draft in minutes, and keep a working record of what they tried. None of that tells them which confident sentence in the draft is wrong here.

In a pre-registered 2023 field experiment with 758 Boston Consulting Group consultants, those given GPT-4 for 18 realistic consulting tasks were measured against a control group without it. Inside the model's capability the gains were large: 12.2% more tasks completed, 25.1% faster, and more than 40% higher grader-scored quality, with consultants below the average performance threshold gaining most ⁴.

One more task sat on the far side of that capability, and there the model worked against them. Right answers came from 84.5% of the control group, and from 60% and 70% under the two AI conditions, 19 percentage points down on average ⁴.

More coaching, worse result on that task: the group given a prompt-engineering overview fell 24 points, the group given none 13 ⁴.

The drop is evidence that working consultants could not reliably tell which side of the line the task sat on. A career changer has even less of the field to judge that line by, and no tool habit substitutes for it.

The filter you would be loosening was never carrying much. A meta-analysis of 81 independent samples put the corrected correlation between prehire work experience and later job performance at .06, and experience with relevant tasks, jobs or occupations specifically did no better, at about .07 ⁵.

Read the limit next to the number. The finding covers experience as measured at hire, the crude quantity a resume screen reads, and a small subset of its studies shows that experience predicting a little better in a job's first weeks. It says nothing about licensure or a minimum qualification the law sets.

Whether the remaining gap becomes a hurdle or a trade on the scorecard is a separate decision, and making it both is the common error.

Give them a packet with one claim they cannot verify

Hand the candidate source material with something in it that cannot be verified from what they were given, and watch for the sentence where they say so. A career changer with the habit you want will produce the draft fast, then stop at the claim the packet does not support and name it.

One without the habit produces the same draft and no stop.

The stop is the measurement. Producing the draft is cheap for everyone in the pool now, so the draft separates nobody. The exercise is hiring someone who can tell when AI is wrong, asked of a candidate who does not know the field yet.

Build the exercise out of the tasks you marked. In the job knowledge meta-analysis behind the 2022 selection re-analysis, all 164 studies together produced a mean observed validity of .22. The 59 using knowledge tests built for the job in question produced .31 ⁶.

The .22 and the .31 come from subsets of one meta-analysis rather than from a controlled trial, and knowledge tests presuppose candidates who already hold the knowledge.

Carrying the finding over to a career changer is a generalisation from that one comparison, and the most it supports is building the domain half of the exercise from the job's own tasks.

Sorting roles this way can freeze people out of the places where domain knowledge gets built. Every senior person in your field learned it somewhere, usually in a job where their errors did travel and somebody caught them.

A rule that only places the domain-naive where mistakes are cheap builds no reviewers.

The reviewer is also the thing you are spending.

The filled-in column sends the decision back to the task list. Name the never-band tasks in the first conversation as the part of the job the hire cannot check yet, and say who checks them instead. If a reviewer holds them, put those hours in that person's calendar before the offer goes out.

If nobody can be named, the column already answered the question.

See what gets scored >