Two things happened to hiring at the same time, and together they broke a process that had worked badly but predictably for thirty years.
The first is volume. Applying is now nearly free and largely automated, so a mid-level opening that once drew forty applications draws several hundred, many of them generated. The second is that the resume stopped carrying information. Cover letters are written by the same tools for every candidate. Portfolios can be produced without the skill they claim to demonstrate. The credential filter that everyone quietly relied on has lost most of its discriminating power.
Skills-based hiring is the right answer to this. It is also, in most of the implementations we see, executed in a way that adds cost, annoys good candidates and predicts nothing.
The short version
- Screening is now the bottleneck, not sourcing. Any advice that starts with getting more applicants is solving last decade's problem.
- Test the job, not a proxy for the job. A short, realistic, paid-if-substantial work sample beats every puzzle, personality instrument and timed quiz we have compared it against.
- Assume candidates use AI, because they do. Design the exercise so using it is fine and the judgement still shows. A test that only works if the candidate is offline is testing your monitoring, not their skill.
- If software screens or ranks people, you own the outcome. Not the vendor. Document what it does and keep a human in the loop.
- Distributed hiring widens the pool and narrows the margin for sloppiness. The written process is what makes remote assessment work at all.
Why do most skills-based hiring programmes fail?
Three failure modes, and we have seen all three inside the same company.
The exercise does not resemble the job. Abstract reasoning puzzles, timed algorithm questions for people who will never write an algorithm, a personality profile treated as a gate. These measure something, occasionally something stable, but not whether this person will do this work well.
The exercise is too long. A four-hour unpaid assignment does not filter for skill, it filters for people who are not currently busy. That is close to the opposite of what you want, and your strongest candidates decline first because they have options.
Nobody defined what good looks like before the responses arrived. Without a written rubric, the scoring happens in the reviewer's head, which reintroduces every bias the programme was supposed to remove while adding the overhead it was supposed to justify.
What to test, by role
| Role | A test that predicts | A test that does not |
|---|---|---|
| Operations or admin | Here is a messy real inbox and a half-written process. Triage it and tell us what you would change | Typing speed, generic aptitude tests |
| Customer support | Answer three real tickets, including one where the answer is no | A personality questionnaire |
| Sales | A short live call against a realistic objection, then a written follow-up | Asking them to sell you a pen |
| Software engineering | Small change in a real repository, with the tests, or a live review of code they did not write | Timed algorithm puzzles for a CRUD role |
| Marketing | Rewrite one of our actual pages and explain the changes | Portfolio alone, unverifiable in 2026 |
| Finance | Reconcile a deliberately broken sheet and report what is wrong | A spreadsheet shortcuts quiz |
| Management | Walk through a real decision they made that went badly, with specifics | Hypothetical leadership scenarios |
The pattern: give them a real artefact from your business, with the confidential parts removed, and ask for judgement rather than recall. Judgement is the thing that is hard to fake and the thing you are actually buying.
How do you assess when candidates use AI?
You stop treating it as cheating and start treating it as the working environment, because it is.
Three techniques that hold up:
Ask for the reasoning, not just the output. The output can be generated. A ten-minute conversation about why they made a particular call, what they considered and rejected, and what they would do if a constraint changed cannot be, and it takes minutes to run.
Put a deliberate flaw in the material. A wrong figure, a contradictory requirement, an instruction that would break something downstream. Generated answers sail past it confidently. People who know the work stop and ask. This single technique has been the most reliable signal we have used.
Make the last stage live and specific. Not a whiteboard interrogation. A short working session on a real problem, with a colleague, as the work would actually happen.
The part that is now a legal question
If a tool screens, ranks or scores candidates, the obligation sits with you as the employer, and it does not transfer to the vendor because their sales deck mentioned fairness. In practice that means three things:
- Know what the tool actually does. What it scores on, what data it was built from, and whether the vendor has tested it for adverse impact. If they cannot answer plainly, that is the answer.
- Keep a human in the decision. Automated rejection with no human review is the shape regulators are most interested in, and it is also how you lose good candidates for reasons nobody can reconstruct.
- Keep records. What the tool did, who reviewed it, why the decision went the way it did. The EEOC has published guidance on automated systems in hiring, and several states now require documentation of automated decisions independently of employment law. We covered that wider surface in the US state privacy and AI patchwork.
Running this with a distributed team
Most of the roles this applies to no longer sit in one building, which makes the process harder in one specific way and easier in every other way.
Harder: you cannot form an impression over three days of someone being around, so the assessment has to be explicit. That is a feature. The impression was never very predictive and was where most of the bias entered.
Easier: the exercise, the rubric and the scoring are all written down and can be reviewed. Remote hiring done properly is more auditable than the in-person version ever was.
Two practical rules. Write the rubric before you post the role, because writing it clarifies what the job actually is, and it is common to discover at this point that two stakeholders wanted different jobs. And give every candidate the same exercise and the same time, so the comparison means something.
If you are building a team across time zones, the management practice matters more than the hiring practice, and we set that out in building a distributed team. If you are considering hiring capacity rather than headcount, what changed about hiring offshore talent is the honest version, and how to hire a virtual assistant team is the operational one.
A process that fits in two weeks
- Write the rubric. Five to seven things the person must be able to do, each with what a good answer looks like. One page.
- Screen against the rubric, not the resume. Three targeted questions in the application. Generated answers are visible when the questions are specific to your business.
- Send the exercise. Ninety minutes maximum, drawn from real work, with one deliberate flaw in it. Pay for anything longer.
- Score blind, against the rubric, by two people. Discuss only after both have scored.
- Run the working session. Reasoning, not recall.
- Check references with specific questions. "Would you hire them again, and for what" produces more than "how did they do".
Two weeks end to end, and the candidates you want will still be available, which is the part the six-week process keeps losing on.
Frequently asked questions
Should we drop degree requirements entirely?
Drop them where the degree is not a licence and not a genuine proxy for anything you are testing, which is most commercial and operational roles. Keep them where they are legally or professionally required. The gain is not ideological, it is that your candidate pool gets larger at no cost to quality.
Do we have to pay for work samples?
Pay for anything substantial, anything you would use, and anything over about ninety minutes. A short realistic exercise is a normal part of a process. A free deliverable is not, and candidates talk.
How do we handle the volume of generated applications?
Ask two or three questions that require knowledge of your specific business and cannot be answered generically, and weight those in screening. It filters far better than any detection tool, which we would not rely on.
Is skills-based hiring slower?
It is faster, if you resist adding stages. The mistake is bolting an exercise onto an existing five-stage process instead of replacing three of those stages with it.
Where to go next
If the honest answer is that you need capacity rather than another full-time hire, that is what our virtual assistant services exist for, recruited, trained and managed rather than handed over. If the hiring process itself is the thing that needs rebuilding, that sits with business operations consulting. Either way, talk to us.
