Article Assessment Research
How to assess AI skills in business roles with evidence your legal team can review: Codility AI for Business for assessment leaders
On this page
In short
An AI skills assessment holds up in a hiring decision when four things are true. The candidate does real work with an AI tool, the score is produced deterministically, your hiring team makes the final decision, and your I/O psychology and legal teams can review the evidence behind it.
Codility AI for Business is built on those four points. The candidate produces a business deliverable with the AI Assistant. Fixed rules score the deliverable against an expert-defined answer space, and no AI model scores the candidate. The reviewer sees every prompt and response.
Three of the reference points assessment teams work to moved in 2026. Colorado repealed and re-enacted its AI Act, with new duties from January 1, 2027. The EU moved its high-risk hiring obligations to December 2, 2027. And a final rule to rescind the federal Uniform Guidelines went to review on October 16, 2026. The professional standards for validating a selection procedure did not move. This page maps AI for Business to them.
Key takeaways
- If a score has to hold up to a candidate, a regulator or a client, ask first how it is produced. In AI for Business, fixed rules produce the score, and your hiring team makes the final decision.
- If Legal is your blocker, start from how you will use the score. New York City’s rules turn on how much weight a score carries. Colorado’s law, from January 1, 2027, turns on whether the output is a “non-de minimis factor”.
- If your I/O psychology team is your blocker, ask for the Technical Manual by name. Then ask which roles each piece of validity evidence covers.
- If you measure current teams as well as candidates, use the same tasks for both. They run in Skills Intelligence as well as in Screen and Interview.
What does Codility AI for Business measure?
Codility AI for Business measures whether a person can get a piece of business work done with AI. The person gives the AI Assistant context, directs it, checks its output against the source data and delivers a result. The deliverable earns the score. The prompts show how the person worked.
AI for Business runs on the same platform as Screen and Skills Intelligence, where Codility has run validated assessments since 2009.
Each task ends in a deliverable: a verified report, a prioritization decision, an account brief or a customer reply. Tasks run 5 to 60 minutes. You can choose general business tasks, which suit any role and level, or tasks tailored to sales, marketing, finance, operations and customer roles.
Some tasks contain deliberate errors in the data or in the assistant’s output. Those errors test the behavior assessment teams ask about most: does the person check the AI’s work before using it? Candidates have the source files open beside the brief, so the check is always possible.
Programs often stall on a definition: what the skill means in your own roles. If your organization already has a framework for AI skills, Codility can map the assessment to it through Professional Services.
Vendors sell four formats for testing AI skills, and each one leaves different evidence behind:
| Format | What the candidate does | How the score is made | What a reviewer can show Legal |
|---|---|---|---|
| Self-report or knowledge quiz | Rates themselves, or answers questions about AI | A fixed answer key | What the candidate knows about AI tools |
| AI-led conversation or interview | Answers an AI interviewer | Often an AI model | Whatever audit evidence the vendor publishes |
| Scenario questions | Picks the best response to a described AI situation | A fixed answer key | Judgment in a described situation |
| Work task with an embedded AI toolAI for Business | Produces a deliverable with the AI tool | Fixed rules against an expert-defined answer space | The deliverable, every prompt and response, and the scoring rule |
AI for Business is the last row. For a vendor-by-vendor view, see our ranking of AI skills assessment tools for business roles. For the measurement behind it, see our AI proficiency assessment.
How is an AI for Business task scored, and where is AI used?
Put the AI in the task and keep it out of the score.
The candidate works with the AI Assistant, and fixed rules produce the score. The rules compare the deliverable with an answer space defined by assessment experts, so the same response always earns the same score. Codility never lets an AI model score a candidate.
Every prompt and reply is saved with the result for human review, and the prompts do not change the score. Codility provides the assistant inside the workspace, which keeps the record complete, so customer-provided models are not supported.
No AI or machine learning makes an automated hiring decision. Your hiring team makes it, with the deliverable, the score and the full conversation in front of them.
An AI-produced score is a second model you’d have to explain, once to your I/O psychologists and again to Legal. With the AI in the task, you measure how the person worked with it, and you can write the scoring rule down.
The saved conversation also briefs the next round: an interviewer can ask why the candidate used, changed or set aside the AI’s output.
Which AI-in-hiring laws apply when an AI skills score informs a hiring decision?
Whether a law applies depends on how you use the score, as well as on how the vendor built it. In New York City, the test is how much weight the score carries. In Colorado, from January 1, 2027, the test is whether the output is a non-de minimis factor in the decision. In the EU, a system used to evaluate candidates is high-risk.
Your legal team therefore has two questions: which laws cover the places you hire, and what the vendor documents.
| Law | What brings a score into scope | Status on November 3, 2026 |
|---|---|---|
| NYC Local Law 144 | A score you rely on solely, weight more than any other criterion, or use to overrule human decision-making, under the city’s rules. A bias audit and candidate notice follow | Enforced since July 5, 2023 |
| Colorado SB26-189 | An output that is “a non-de minimis factor” in an employment decision. Notice before use, an explanation within 30 days of an adverse outcome, and human review. No impact assessment is required | Applies to decisions on or after January 1, 2027 |
| Illinois Human Rights Act, as amended by HB 3773 | AI use that results in discrimination against protected classes, plus notice to applicants and employees | In force since January 1, 2026; notice rules not final |
| California automated-decision system regulations | Covers “computer-based assessments or tests”. Anti-bias testing counts as evidence, and records are kept for four years | Effective October 1, 2025 |
| EU AI Act, Annex III | Systems intended “to evaluate candidates” are high-risk | Obligations apply from December 2, 2027, after the Digital Omnibus, Regulation (EU) 2026/1744, entered into force |
For that review, Codility provides documentation written for your legal, data protection and AI governance teams. It covers how tasks are scored, how AI is used in the assessment and where human review sits. Your team checks that posture against the rules for your locations. Codility is SOC 2 Type II audited and ISO 27001 certified, and you choose EU or US data hosting when you sign the contract.
What validation evidence should you ask for, and what does Codility provide?
Ask for evidence against the five considerations in SIOP’s 2023 guidance on AI-based assessments, and for every document, ask which roles it covers. A vendor can have strong evidence for one job family and none for yours.
| SIOP 2023 consideration | What to ask the vendor for | What Codility provides | For AI for Business tasks |
|---|---|---|---|
| Scores predict future job performance accurately | A criterion-related validity study for your job family | Content validation documented in the Technical Manual, and local validation studies run with each customer | Criterion evidence for your roles comes from your local study |
| Scores are consistent | Reliability figures and the scoring rule | Fixed-rule scoring, with reliability reported with McDonald’s omega in the Technical Manual | Scoring rule public; reliability provided on request |
| Scores are fair and unbiased | Adverse-impact analysis at your cut score, and a language review | Adverse impact monitored at every measured cut score, for candidates who complete the optional demographic questionnaire. An independent cApStAn linguistic audit found 65% of tasks at or below level B1 | Monitoring covers the platform |
| Operational considerations and appropriate use | Where AI is used, who decides, what candidates see | AI in the task only, fixed-rule scoring, every prompt saved, people make the decision | Public on the AI for Business page |
| All steps documented for verification and auditing | A technical manual and its update cycle | A 92-page Technical Manual, updated at least every 6 months and shared with prospective and current customers | Provided on request |
The four-fifths ratio comes from the federal Uniform Guidelines on Employee Selection Procedures, at 29 CFR 1607.4D. The EEOC sent a final rule rescinding the Guidelines to federal review on October 16, 2026, and the rule had not been published by November 3. Treat four-fifths as the long-standing professional benchmark. Two professional references for validating a selection procedure are unchanged. One is SIOP’s Principles for the Validation and Use of Personnel Selection Procedures, fifth edition. The other is the Standards for Educational and Psychological Testing.
Codility says in public where the science stops.
Its post on holding itself to the SIOP standard says that “scoring an AI interaction in a way that is deterministic, explainable, and fair is genuinely hard”, and that “we would not claim to have fully solved it.”
More on the method: Codility’s assessment science.
Does AI for Business work for entry-level and non-technical candidates?
Yes. General AI for Business tasks suit any role and level, and the shortest take 5 minutes. That fits early-career and campus programs, where candidates have little work history to show.
Language load is checked too. An independent linguistic audit by cApStAn found 65% of Codility tasks at or below level B1 on the Common European Framework of Reference for Languages.
Early career hiring is one of Codility’s longest-standing areas. Before you set a bar for entry-level candidates, run the tasks with current entry-level employees first.
Where does AI for Business fit next to your existing assessments?
AI for Business sits alongside the cognitive and situational judgment tests you already run and adds evidence of how a person works with AI. Codility offers no self-report or personality instruments, so your current battery keeps that job.
It runs in three places:
- Screen For hiring before the interview, with results in your applicant tracking system (ATS)
- Interview Where the candidate works on a real task with the AI Assistant
- Skills Intelligence For measuring the AI skills of current teams
All three share Codility’s reporting, integrations, security and assessment science. You decide whether the score is a gate or one weighted factor.
How do you pilot AI for Business and set a cut score?
Pilot it in four steps:
- Start with one cohort in a 2-week team evaluation.
- Run the same tasks with current employees in Skills Intelligence to see what good work looks like in your roles.
- Set the cut score with your assessment team, and check adverse impact at that cut score.
- If your counsel wants criterion evidence for your own roles, run a local validation study with Codility.
If you already have an AI skills framework, map the tasks to it in step 2, before anyone sets a cut score.
When is AI for Business not the right fit?
AI for Business is the wrong tool if you need any of these:
- A personality or self-report instrument
- Mechanical, electrical, hardware or CAD tasks
- Learning paths attached to the assessment
- Candidates using their own AI model, or one your company provides
FAQ
No. Fixed rules score the deliverable against an answer space defined by assessment experts, so the same response always earns the same score. Every prompt is saved for human review and does not change the score.
See the evidence for yourself
Request the Technical Manual and a walkthrough of an AI for Business task with our assessment science team.
Request the Technical Manual