After a company completes a generative AI training program, employees may be able to write polished prompts and produce credible summaries. Managers then face a more difficult question: Do these employees truly understand how to collaborate with AI, or did they simply obtain content that looks like the right answer? If an assessment considers only completion speed and the final deliverable, managers cannot tell whether employees identified errors, verified sources, or understood which results should not be delivered without further review.
Using AI effectively requires more than knowing how to ask questions. Employees must also know how to validate and correct the output—and take responsibility for the final deliverable. To determine whether employees possess these capabilities, companies should have them perform tasks in real-world work scenarios. Managers can then observe how employees handle AI-generated content, where they pause to review it, the reasoning behind their revisions, and whether they are ultimately willing to take responsibility for the results they approve.
Google Is Bringing AI into Interviews, but Prompts Are Not the Only Thing Being Assessed
Google is gradually revising its engineering recruitment process. In some “code comprehension” interviews, candidates are now allowed to use company-approved generative AI tools to read, debug, and optimize existing code. However, this approach is still being piloted and has not been adopted across all engineering interviews. Companies should therefore not treat it as a finalized standard process at Google.
This pilot is worth examining because interviewers look at more than the final answer. Assessment areas include prompt engineering, output validation, and debugging. After receiving suggestions from AI, candidates must verify that the content is correct, identify errors in the code, and confirm that their revisions meet the requirements without overlooking edge cases. HackerRank also recommends that candidates first understand the suggestions provided by AI and then explain their reasoning to the interviewer. However, this is interview advice from HackerRank and should not be presented as an established Google requirement.
Google has stated that 75% of new code is generated by AI and then reviewed by humans, with engineers making the final decision on whether to use it. In other words, allowing employees to use AI does not make their core professional expertise any less important. Without the ability to understand requirements, apply sound judgment, and debug problems, it is difficult to determine whether an AI-generated answer can be used. Managers considering this approach should first clarify what they want to assess: whether employees know how to operate the tool, or whether they can verify that its output is sound and safely incorporate it into a workflow.
Skip the Prompt Competition: Use Real Tasks to Assess Four Key Stages
To assess the AI collaboration skills of current employees, companies can begin by selecting a clearly scoped task for which the manager understands the expected outcome. Examples include reviewing sales quotations, comparing procurement bids, summarizing quality-control anomalies, drafting customer service responses, or reviewing code. The task should preserve the kinds of missing data, conflicting conditions, and exceptions that employees encounter in real work. This gives employees an opportunity to demonstrate their judgment rather than merely compete to see who can write the longest prompt.
The task can then be divided into four stages: AI output, human validation, correction, and approval. Managers can require employees to retain their prompts and the corresponding AI responses, document the reason for every revision, and explain which parts can be accepted and which must be returned for rework. Performance can be assessed across the following four dimensions:
- Understanding requirements: Can the employee clearly explain the task’s objective, constraints, and intended recipient before immediately feeding the information to AI?
- Identifying errors: Can the employee spot contradictions, incorrect citations, missing fields, and content that does not comply with company policies?
- Checking edge cases: Does the employee proactively verify monetary thresholds, exception clauses, requirements for different customers, and how to proceed when information is insufficient?
- Reasons for correction and approval: Can the employee explain the reasons and basis for each revision, as well as why they are willing to allow the result to proceed to the next stage?
The same assessment method can be used for recruitment, post-training validation, and departmental pilots. However, the task must reflect the actual responsibilities of each role. Procurement personnel do not need to take a code-debugging test, while engineers should not be assessed solely on document summarization. Managers should first choose a real task and clearly define acceptable answers, unacceptable errors, and the situations in which employees must stop and seek confirmation.
Managers Need Visibility into Employee Judgment—and Records That Can Be Audited Later
A one-time assessment can show how an employee handles a task at that moment, but it cannot prove that the employee will continue following the same process in day-to-day work. Pilot teams can therefore retain prompts, AI responses, human revisions, supporting evidence, and reasons for approval. Managers can then review unsuccessful cases to determine whether the problem arose because the employee did not understand the task, failed to verify the information, lacked a required review checkpoint in the process, or received AI-generated content that was fundamentally unsuitable.
These records should not be kept to monitor every word employees enter. Their purpose is to ensure that the company can clearly explain what happened when a problem occurs: who adopted the AI’s recommendation, who revised the content, who approved its release, and which stage the process should return to after an error is identified. Assessments should not be based solely on the number of prompts, frequency of use, or completion speed, because these figures do not reveal whether an employee’s judgment was correct. Companies can review several successful and unsuccessful tasks each month to confirm whether employees are genuinely following verification standards, and then decide whether to strengthen training, revise assessment tasks, or modify workflows.
Companies using EgentWrX can connect multiple tasks into a workflow and add human approval checkpoints before handoffs. After an AI Agent completes one stage of the work, the workflow pauses and waits for human confirmation. EgentWrX’s audit log records the actions performed by the AI Agent. These records can be exported, and their integrity can be verified using a hash chain. However, pilot teams should still separately retain prompts, AI responses, human revisions, and reasons for approval. Reviewing these materials alongside the platform records provides visibility into both what the tool did and how people exercised judgment.
Companies do not need to begin by designing a standardized assessment for the entire organization. Instead, they can select tasks with manageable risks and outcomes that managers are capable of evaluating. A small group of users can then complete the full process, after which the company can refine its assessment criteria based on the records produced. The assessment approach should be extended to other roles only after the team can clearly explain which results may proceed and which situations must always result in the work being returned for revision.
FAQ
Should companies assess prompting skills or professional judgment first?
Professional judgment should be assessed first, followed by whether employees can use prompts to obtain the information they need. Prompts are simply a way to interact with AI. Managers have a greater need to confirm whether employees can identify errors, know how to verify information, and explain why they accepted or rejected an AI recommendation.
Which types of work are best suited for an initial AI collaboration skills assessment pilot?
Start with tasks that have clear standards for success and for which managers can identify errors, such as reviewing quotations, comparing procurement bids, drafting customer service responses, or reviewing code. A process is not suitable for an initial pilot if responsibilities are difficult to separate, the data is highly sensitive, or any error would have an immediate impact on external customers.
Which assessment records should managers retain?
At a minimum, managers should retain the task requirements, prompts, AI responses, human revisions, supporting evidence, and reasons for approval. These materials allow managers to determine at which stage an employee identified a problem and whether the underlying cause was unfamiliarity with the tool, insufficient professional knowledge, or a missing review checkpoint in the process.
Can a single assessment prove that an employee has AI collaboration skills?
No. A one-time assessment reflects only the employee’s performance at that particular moment. Companies should apply the same standards to routine tasks and periodically review records from both successful and unsuccessful cases. This helps confirm whether employees continue to verify and correct results—and document their reasons for approval—when dealing with different data, exceptions, and workplace pressures.
References
- Google’s Engineering Interviews Are Changing! Candidates Can Use Gemini—and Are Tested on Their Ability to “Direct AI” — Business Next’s summary of Google’s interview process pilot, official statements, and information from HackerRank.
