Needs analysis — workplace track
Status note: Northwind Logistics is a fictional client created for this portfolio. Client-specific claims (e.g., “agents paste customer data into AI tools”) are hypothetical for the fictional client. Population-level evidence comes from published research, cited inline and in research/sources.md (secondary research; no interviews were conducted).
Last reviewed: 2026-10-25
1. The business goal and the stakeholder request
Stakeholder request (fictional, Northwind Logistics): “We need AI training for everyone by Q1 so we’re covered under the AI Act.”
Reframe. Article 4 of the EU AI Act requires providers and deployers to take measures to support AI literacy — and explicitly does not require any individual to reach a specific level (AI Act Service Desk, accessed 2026-10-25). A completion certificate is therefore the wrong target. The performance goal that actually matters:
Staff use AI outputs safely and correctly in real tasks — they verify before sending, protect data, and keep humans in decisions that affect people.
Business goal: safe, correct AI use in customer service, operations planning and HR administration, without data incidents and without wrong decisions reaching customers or applicants.
2. The performance gap
| Aspect | Current state (hypothetical, fictional client) | Desired state |
|---|---|---|
| Verification | Send AI drafts after a skim; fluent text feels correct | Claims, numbers and sources checked before sending |
| Data | Paste anything into the AI assistant to “give it context” | Classify data first; never paste customer personal data |
| Judgment on fit | Use AI for anything, including decisions about people | Choose tasks where AI helps; keep human decisions where people are affected |
| Escalation | Quietly delete a chat after a mistake | Report incidents; use defined escalation paths |
This gap pattern matches external evidence: use is already widespread and self-taught (75% of knowledge workers use gen AI at work — Microsoft/LinkedIn WTI 2024), training emphasizes productivity over risk (Interface EU), and L&D demand for “AI capability” is broad but unspecified (43.3% of postings vs. 8.3% with specifics — Teamed 2026).
3. Root causes
Sorted into knowledge, skill, motivation and environment. Environment causes get non-training fixes (flagged), because training cannot compensate for a missing policy.
| Cause | Type | Training response |
|---|---|---|
| Staff believe AI “looks things up”, so they trust fluent output | Knowledge | Module 1: how text generation actually works (prediction, not lookup) |
| No verification routine exists | Skill | Module 2: teach a 5-step routine + job aid |
| Staff can’t distinguish safe vs. unsafe inputs | Skill + environment | Module 3: traffic-light classification. Non-training fix (recommended): publish the data classification policy the modules reference |
| “Everyone does it” normalization of pasting data | Motivation | Module 3: consequences + realistic alternatives; manager messaging recommended |
| No defined escalation path for AI mistakes | Environment | Module 4 teaches the (fictional) path. Non-training fix (recommended): the organization must actually define one |
| Compliance framing makes training feel like paperwork | Motivation | Scenario-based content about real decisions, not policy recitation |
4. Task analysis: where judgment is needed
| Task (Northwind context) | AI helps with… | Judgment steps where errors occur |
|---|---|---|
| Draft customer reply (Lena) | Tone, speed, first draft | 1. Paste request text (data?) → 2. Check claims (SLA times, prices) → 3. Check name/order details → 4. Send decision |
| Summarize operations report (Joris) | First-pass summary | 1. Numbers in summaries can be invented → 2. Check every figure against the source → 3. Act on the summary, not the source (risk) |
| Screen incoming CVs (Sofia) | Not recommended without safeguards | 1. Affected persons → 2. Bias in ranking → 3. Human must decide → 4. Escalate to DPO/lead before any AI use |
5. Constraints
- 20 minutes per module (self-paced, in work time)
- Mobile-friendly (shop-floor and on-the-road access)
- Multilingual workforce — plain English at B1 reading level (~grade 8); sentences short; idioms avoided
- No LMS packaging in v1.0 (SCORM noted as a next step)
- No real customer data in examples
6. Critical behaviors the program must change
- Verify before sending — claims, numbers, sources, names (Modules 2 + job aid)
- Classify data before pasting — never paste prohibited data; know what restricted means (Module 3)
- Keep a human decision where people are affected (hiring, discipline, compensation, safety) (Module 4)
- Escalate and report suspected data leaks and AI errors instead of hiding them (Modules 3–4)
- Use the job aid at the moment of need — performance support beats memorization for infrequent tasks
These five behaviors drive the learning objectives and the alignment matrix.