How We Ended Up Building Our Own Screening Pipeline
We went through three different Pre Employment Assessment platforms in two years before realizing none of them actually measured what we needed. That process cost us roughly forty thousand dollars and nearly burned out our HR team. Here is what I learned along the way. A Pre Employment Assessment is supposed to filter candidates before the interview stage. The reality is more complicated. You pick a tool, you configure it, you send it out, and then you spend three weeks arguing with your procurement team about why the scores look wrong for half your applicant pool. There are three main categories you will encounter. Cognitive ability tests measure reasoning speed and pattern recognition. Work sample tests ask candidates to perform tasks identical to the actual job. Personality inventories attempt to predict cultural fit and retention risk. Each category has a different failure mode.
I found that cognitive tests have the highest false positive rate for experienced hires. A candidate who has been doing the work for eight years might still score in the 40th percentile because the test was normed against recent graduates. That happened to me with a senior data role where my top two final candidates bombed a general reasoning test. We hired both of them. They became the strongest performers on the team. The test predicted nothing useful.
What to Do Before You Buy Anything
The first step nobody tells you about is writing a job analysis document. This is a one or two page breakdown of the actual tasks, decision points, and cognitive demands of the role you are hiring for. Without it, any assessment you purchase is a guess. I used to skip this. I would just look at what TestGC or SHL offered and pick something that looked reasonable. The candidates we screened that way were mediocre at best. Once I started spending three hours with the hiring manager to map out real job tasks, the assessment selection process became obvious. You stop looking for general intelligence measures and start looking for work samples that mirror the daily workflow.
Get the Full Details
How to Run a Work Sample Test
Work sample tests are the single most predictive category for job performance according to meta analysis data. The problem is building one takes significant effort. You cannot just make up a task and expect it to correlate with success. Here is my basic structure. I take the top five tasks from the job analysis. I convert each into a time bounded exercise of twenty to forty minutes. I grade them with a rubric built from my best current employees in that role. The rubric has four levels: does not meet expectations, meets expectations, exceeds expectations, and exceptional. Vague grading scales like poor to excellent are useless because two hiring managers will interpret them completely differently. We ran a simulated customer support case where candidates had to triage ten tickets in twenty minutes and write three full responses. The rubric scored accuracy, tone, escalation judgment, and completeness. That single exercise had a correlation of roughly 0.54 with six month performance ratings. That is stronger than any personality inventory we tested.
The Edge Case That Broke My Workflow
Two years ago I was running assessments for a remote engineering team across four time zones. I discovered that one of our top scoring candidates in India was actually using an AI coding assistant during the work sample test. The test platform did not detect it because it only monitored tab switches, not clipboard activity or secondary screens. The candidate got a perfect score. We brought them in for an interview and they could not explain a single line of code they had submitted. The workaround was brutal but effective. I switched to live proctored sessions for the technical portion and moved the work sample to a take home format with a mandatory written explanation component. Candidates had to submit their code plus a paragraph explaining their reasoning for each major decision. The AI assistance problem essentially disappeared because generating the explanation part was too much friction for most people. It also added a layer of depth to the evaluation that the original test lacked.
Common Pitfalls That Waste Your Budget
Using a single cutoff score across all roles. I saw a company set a 70th percentile minimum for every position. They rejected candidates for junior support roles who would have been strong fits because the general cognitive test did not differentiate well for entry level positions. The fix was role specific percentiles based on historical performance data. Junior roles needed lower thresholds. Senior roles needed higher ones. Another trap is treating personality assessments as screening tools. They are not. These tests predict retention and team dynamics better than they predict individual performance. Using them to eliminate candidates introduces bias without improving hiring quality. I moved ours to a final stage conversation starter instead of a pass fail gate.
Build It Yourself or Buy It
Commercial platforms like HireVue, Criteria, and Aptitude International offer decent out of the box packages. The customization is limited and the per candidate cost adds up fast if you hire more than twenty people a year. A custom work sample pipeline built on something like TestGorilla or even a shared platform costs roughly a fraction of that once you have the rubrics in place. I recommend starting with a commercial vendor for your first two hiring cycles to understand baseline metrics. Then move toward custom exercises once you have enough internal performance data to validate your rubrics. The transition usually happens around the thirtieth hire for a mid size team. The biggest thing I wish someone had told me is that assessment validity decays. A test that predicts well today will predict less well in three years as the job itself changes. We stopped revalidating our work samples after a product launch shifted our engineering workflow. The old exercises were still measuring something, just not what we thought they were. Schedule a yearly review of your assessment results against actual performance data. You will be surprised how quickly the correlation drops when you stop maintaining the tool.
