Friction Filters Effort, Not Ability. Spend It Where It Pays Back.

What Does Friction Actually Filter?

Whatever the step is expensive in. A form that costs twenty minutes of typing filters on the willingness to type for twenty minutes. For two decades that was a usable proxy, because writing time was scarce for everyone and roughly tracked how much a candidate wanted the job. The proxy worked by accident, and the accident has ended: drafting is now the cheapest thing in the process.

That leaves three kinds of cost, and only two of them still sort anyone:

• Text cost. Words, essays, cover letters, motivation paragraphs. Free to produce at any length. Filters nobody, and annoys the people who write their own.
• Attention cost. Reading something specific, forming a view, committing to an answer. Still scarce, because a candidate applying to fifty roles cannot spend fifteen real minutes on each.
• Access cost. A login, a proctored window, a video recording, a psychometric account, a request for a portfolio nobody has. Sorts on circumstance, not on the work.

Attention is the only one of the three that separates candidates on something you wanted to know. Access cost is the trap: it looks like rigor and behaves like a demographic filter.

The economics only work when the expense falls harder on the applicant you do not want, which is the whole definition of a signal. A form fails that in both directions now, since the cost is identical for everyone and near zero for anyone running an assistant. A question about the work fails it in one direction only, which is why it still sorts.

Why a Longer Form Selects Against the People You Want

Because the candidate who fills it out by hand pays the full price and the candidate running an assistant pays nothing. Agents now complete application forms, write the essay boxes and answer follow-up email, so every field you add is a tax on exactly one group. It cannot shrink the automated share of the pile, because that share pays nothing to fill it. What it changes is who is left inside, and that runs the wrong way.

Added friction carries a legal exposure too, and that one arrives as a complaint. Under 29 CFR 1630.11, the US regulation implementing the ADA (2023 edition), an employer must select and administer employment tests so that the results reflect the skill the test is meant to measure and not a candidate's impaired sensory, manual or speaking skills, except where those skills are what the test measures ¹. A timed form, a recorded answer or a rigid interface can fail that without anyone intending it. The fix is cheap: of the employers who told the Job Accommodation Network what an accommodation cost, 61% said it cost nothing and 33% reported a one-time median expense of $300 ². That is a self-selected sample of employers who already asked for help, but the direction is not in doubt. Confirm with counsel before relying on any of this for your own process.

Automated steps carry their own price in willingness. In a Pew Research Center survey of 11,004 US adults fielded in December 2022, 66% said they would not want to apply to an employer that uses AI to help make hiring decisions ³. That is stated intention rather than a measured drop-off, and nobody in the sample was standing in front of a job they wanted. Treat it as a reason to describe an automated step carefully, not as an abandonment rate. Whether to accept an application an agent filled out is the policy question underneath all of this.

Put the Friction Where It Produces Evidence

One step, fifteen minutes, role-shaped, and the candidate keeps something. That is the whole specification. A short scenario with a real constraint from the team, a two-question read of a document, one artifact plus the decision inside it: any of these produce material a screener can judge, which a form field never did. Say that AI is allowed, because the interesting part is the judgment rather than the drafting.

The payback test is what separates this from an essay box. Ask what you would have had to run later anyway. If the step replaces a first-round screen, it saves the candidate a scheduled call and saves you the call itself. If it is additive, it is friction with no return and it belongs after a human conversation, with payment if it runs past an hour.

What the candidate keeps matters more than it sounds. A result, a short piece of feedback, a copy of what they produced: any of these converts the fifteen minutes from a toll into a trade. It also changes the tone of the rejection, because there is something concrete to point at.

Assessing everyone up front instead of screening resumes at all is the version of this that goes furthest, and it is a real option when the volume is large and the roles are uniform.

Check Every Step Against Two Questions

Run each step in your application through two questions, in order. First: does this cost different candidates different amounts, in a way connected to the work? Second: does the candidate get anything for paying it? A step failing the first is not a filter. A step failing the second is attrition you chose. Steps failing both are the ones teams add during a volume panic.

Applied to the usual list:

• Cover letter box. Fails both. Free to produce, returns nothing, and reads the same in every application.
• Ten screening questions the ATS added by default. Fails both, and quietly rejects people on knockout logic nobody reviewed.
• Account creation before the form. Fails both, and costs most on a phone.
• A fifteen-minute role-shaped task, AI allowed. Passes both, if you read every answer and reply.
• A timed test with no accommodation route. Passes the first and fails the second, expensively.

The honest version of this decision is a reading budget, not a barrier. Whatever you add, someone has to read the output, and the number of steps you can add is capped by that. Cutting a very large pile to a shortlist is the same arithmetic from the other end.

Audit what you already have before adding anything. Count the fields in the current application, then ask which of them has ever changed a decision. Most teams find two or three, and a long tail nobody has read since the form was built. Deleting that tail buys back the time the new step will cost, which makes the change net neutral on your side and clearly better on the candidate's.

See a sample report >