AI Skills Assessment: Test Work, Not Trivia

An AI skills assessment should ask a person to produce inspectable work for one bounded task in their role. It should not merely test whether they can recall tool vocabulary. LearnAIthing's assessment model gives the person a clear brief, then reviews the workflow, evidence, exceptions, ownership, and handoff. The result describes readiness for that task under recorded constraints. It does not assign a permanent label or turn one work sample into a claim about universal AI competence.

What does the AI skills assessment scorecard measure?

LearnAIthing recommends assessing one bounded workflow from the person's real role. The scorecard records what a reviewer can inspect, not what the person says they understand.

These five dimensions are LearnAIthing's categories, not an industry standard. The three scoring anchors also belong to LearnAIthing's model. Record the role, workflow, constraints, and assessment date beside the scores. That context prevents a result for one task from becoming a claim about every task the person may face.

The European Commission's DigComp provides a common understanding of digital competence and a basis for digital-skills policy. European Commission Joint Research Centre, Digital Competence Framework DigComp describes digital competence as relevant to finding and critically interpreting information, creating, editing, and sharing digital content, managing privacy, safety, and wellbeing, and using digital platforms and services. European Commission Joint Research Centre, Digital Competence Framework

DigComp does not endorse LearnAIthing's five dimensions. LearnAIthing cites it only to support examining several observable competence areas instead of reducing digital competence to tool recall. European Commission Joint Research Centre, Digital Competence Framework

What should the assessment brief contain?

LearnAIthing recommends giving the person a brief before reviewing the work. Use one bounded workflow from the person's role and keep the assessment tied to that workflow.

Copy this assessment brief:

1. Role and workflow. Name the person's role and the single workflow under assessment.
2. Objective. State what the workflow must accomplish.
3. Allowed inputs. List the inputs the person may use.
4. Constraints. State the boundaries that apply to the work.
5. Acceptance evidence. Define what the person must show so a reviewer can inspect the result.
6. Exception path. State the exception the person must identify, record, or handle.
7. Handoff recipient. Name the person who must be able to receive the work.
8. Assessment date. Record when the evidence was reviewed.

The bounded task matters. LearnAIthing does not recommend generalizing from one workflow to universal competence. The assessment answers this task-specific question: can this person turn this role task into a bounded, reviewable AI-assisted system under these constraints?

What belongs in the evidence pack?

LearnAIthing's evidence pack makes the work inspectable after the person finishes the task. Collect the artifacts below:

• A workflow map that shows the bounded task.
• The prompts, instructions, or configuration used in the workflow.
• Source references used for the work.
• Sample outputs from the workflow.
• Evaluation notes for those outputs.
• An exception record.
• A runbook for the handoff.
• A named owner.

In LearnAIthing's model, do not substitute a confident explanation for missing artifacts. Confidence can inform support, but it does not replace the observable work sample. The same rule applies to tool vocabulary. Naming a prompt pattern, model, or product does not provide the workflow map, output evidence, exception record, or runbook that this model asks a reviewer to inspect.

The evidence pack also keeps scoring tied to visible work. A reviewer can point to the artifact behind an anchor, and the person can see what remains absent or needs support. That makes the result useful for choosing the next task without pretending that the score describes the whole person.

How do the five performance dimensions work?

The dimensions below are LearnAIthing's first-party framework. They describe what to inspect in one bounded work sample.

Frame the work

Inspect whether the person makes the objective, allowed inputs, constraints, acceptance evidence, exception path, and handoff recipient clear. The task boundary belongs in the work sample. A reviewer should not have to infer what the workflow is meant to do.

Choose and constrain the AI use

Inspect the prompts, instructions, or configuration and their source references. Check whether the evidence shows where AI is used and which constraints bound that use.

This dimension includes constraints and risks because LearnAIthing's model requires them to be part of the work. NIST developed the AI Risk Management Framework to help manage risks to individuals, organizations, and society that are associated with AI. NIST, AI Risk Management Framework AI RMF 1.0 is intended for voluntary use and helps organizations incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems. NIST, AI Risk Management Framework

NIST's Generative AI Profile can help organizations identify generative AI risks and proposes risk-management actions aligned with organizational goals and priorities. NIST, AI Risk Management Framework

LearnAIthing cites these NIST resources only to support examining constraints, risks, outputs, and ownership as part of work. LearnAIthing does not present its rubric as a NIST framework. NIST, AI Risk Management Framework

Build the workflow

Inspect the workflow map, prompts, instructions or configuration, and sample outputs together. LearnAIthing recommends looking for a bounded system that another person can review. A loose collection of tool actions is not enough when the steps and outputs cannot be inspected as one workflow.

Evaluate outputs and handle exceptions

Inspect the sample outputs, evaluation notes, and exception record. The person needs to show how they reviewed the outputs and what happened on the exception path named in the brief. An unhandled quality or safety exception remains visible in the result. A team roll-up must not hide it inside an average.

Document ownership and handoff

Inspect the runbook, the named owner, and the handoff recipient. LearnAIthing's model treats handoff as part of the work sample. The build is not independently observable for this bounded task if ownership and the instructions needed for handoff remain unclear.

For a team that wants capability transfer as the next step:

Explore the Deployed Business Operator path

Can self-assessment replace observed assessment?

No, not in LearnAIthing's model. A self-assessment can collect confidence and perceived gaps. It cannot replace the observable work sample.

Use the two forms of assessment for different evidence:

• Self-assessment records how confident the person feels and where they believe they need support.
• Observed assessment records what a reviewer can inspect in the brief, workflow, outputs, exception record, runbook, ownership, and handoff.

A mismatch between confidence and observed evidence is not a character judgment. Keep the discussion on the task. The evidence may show that the task needs tighter boundaries, that support is appropriate, or that a smaller practice task would produce a clearer next step.

For the capabilities that may sit behind a work sample, read what AI skills are.

How should you interpret outcomes without pass/fail theater?

LearnAIthing recommends one of three task-specific outcomes:

• Ready for a bounded build. The required evidence is observable independently for this bounded task.
• Ready with support. The work is observable when the person receives support on a dimension.
• Needs a smaller practice task. The current work sample does not yet make enough of the bounded task observable.

These outcomes describe the next task and support decision. They are not permanent labels. An unsuccessful attempt provides evidence about task sizing and support, not a judgment about the person's fixed ability.

Avoid compressing the scorecard into a theatrical pass or fail. Keep each dimension visible, attach it to its evidence, and preserve the recorded constraints. A person may build the workflow independently while still needing support with exceptions or handoff. That distinction tells the team what to review next.

Use the outcome to choose a bounded next step. LearnAIthing recommends using the evidence to size the task and identify support instead of carrying the prior outcome forward as a permanent status. An AI training plan for employees or AI training for teams can provide the next learning context.

How should a team roll up the results?

LearnAIthing's team roll-up should preserve the work behind each result. Show:

• Which role workflows have an evidence pack.
• Which of the five dimensions have observable evidence.
• Where support is needed.
• Who can review a build.
• Who can own a build.
• Which quality or safety exceptions remain unhandled.

Do not average away an unhandled quality or safety exception. Keep the exception attached to the relevant workflow, evidence, and owner. A summary number cannot replace that record.

The roll-up is a map of evidenced work, not a ranking of people. It helps a team see where a bounded build can proceed, where a reviewer or owner is available, and where the task needs support or a smaller boundary. Record the assessment date because every result belongs to a specific work sample reviewed at a specific point.

The team can also use the evidence map alongside an AI pilot program. LearnAIthing does not recommend treating the roll-up as proof that every person can operate every workflow.

FAQ

How do you test AI skills at work?

LearnAIthing recommends testing one bounded workflow from the person's real role. Give the person a brief with an objective, allowed inputs, constraints, acceptance evidence, an exception path, and a handoff recipient. Review the resulting evidence pack against the five dimensions and three scoring anchors.

What are the top AI skills to assess?

In LearnAIthing's model, assess the ability to frame the work, choose and constrain the AI use, build the workflow, evaluate outputs and handle exceptions, and document ownership and handoff. These are first-party categories for a bounded work sample, not a universal list or industry standard.

Can I use an AI skills self-assessment?

Yes, to collect confidence and perceived gaps. In LearnAIthing's model, self-assessment does not replace observed evidence from the person's work sample.

How much does an AI skills assessment cost?

LearnAIthing's public model does not prescribe a price. The reusable brief, evidence pack, dimensions, and scoring anchors in this article define what to assess without turning the article into a vendor quote.

What happens if someone does not complete the task?

Treat the attempt as evidence about task sizing and support. The appropriate outcome may be ready with support or needs a smaller practice task. Do not turn the attempt into a permanent judgment about the person.

Should an assessment include practice questions and answers?

LearnAIthing's model does not score trivia answers. It scores observable work and the evidence attached to a real role workflow. A fake answer key would reward recall instead of showing whether the person can build and hand over the bounded system.

What should a manager do after the assessment?

Use the recorded dimensions, evidence gaps, exceptions, ownership, and handoff needs to choose the next bounded task and the support it needs.

For a team that wants capability transfer as the next step:

Explore the Deployed Business Operator path

Written by Loïc Guyon (Tileo), an operator who learns AI by running businesses with it.