AI requirements
What the client actually asked for, extracted from PR MIEW.pdf — a Q&A round on My Workplace
Legal, answered by the client, with the legal process validated by Rui Neves Ferreira against the
law.
Read this as evidence, not as a plan. Every numbered item below is something the document states. Anything undecided sits in Open questions and is not a default to build on. The 8-phase legal process itself lives in Phase Steps & Inputs.
At a glance
| # | Requirement | Type |
|---|---|---|
| AI-1 | AI participates in all 8 phases, at a different moment in each | scope |
| AI-2 | AI never fills the informação base necessária | boundary |
| AI-3 | AI drafts the deliverables formais, blocked until base info is complete | feature |
| AI-4 | Cross-case learning — tell the handler what to watch out for | feature |
| AI-5 | Direct Sim/Não verdict, always with grounds, plus a confidence signal | feature |
| AI-6 | Detailed reasoning with traceable sources is mandatory | constraint |
| AI-7 | Human review is always required | constraint |
| AI-8 | Verdict visibility is role-gated, Instrutor and above | constraint |
| AI-9 | Background assistant first; chatbot is a nice-to-have | interface |
| AI-10 | May ground on historical data, with a GDPR disclaimer | data |
The human/AI boundary — AI-2 + AI-3
The single most important line in the whole document. Each phase splits in two, and the AI only ever touches the lower half.
For Abertura the deliverables are, in export order:
| Order | Deliverable | Note from the document |
|---|---|---|
| 1 | opening cover | — |
| 2 | Participação Disciplinar | Explains what originated the case. The case's opening document. |
| 3 | Despacho Inicial e Nomeação de Instrutor | "e por aí fora para cada secção" |
| 4 | Termo de Abertura | same pattern |
Two hard constraints fall out: generated deliverables have a fixed position in the export, and generation is blocked until the phase's base fields are complete.
Where AI intervenes — AI-1
Every phase, but not at the same point in each. Each one has base information (human) and a document-production part (where the AI helps): "a IA possa ajudar a escrever ou detalhar cada processo".
Who sees the verdict — AI-8
Users of the AI are all direct participants, but the Sim/Não verdict stops at Instrutor. The cut line matches the app boundary: the three roles below it have no admin app at all, only the Portal.
The verdict itself — AI-5
Four separate asks, worth keeping distinct:
| Ask | Strength |
|---|---|
| A direct opinion — "há fundamento: Sim/Não" — not a compilation for a human to judge | must |
Always carries its fundamentação | must |
| A confidence level / "temperatura da análise" beside it | wanted |
| The reviewer can rate the answer — likes it or doesn't | wanted |
The remaining constraints
| # | What the document says |
|---|---|
| AI-4 | "…com base em todos os casos analisados, ou já processados, ajudar a compreender, a quem vai tratar do processo, o que deverá ter em conta." An advisory layer over past cases, aimed at whoever is handling this one. |
| AI-6 | Presenting the reasoning is essencial, so there is always a reference point for where the information came from. Provenance is a requirement, not a nicety. |
| AI-7 | "Sim terá que existir sempre" — a human validation step before any action, without exception. |
| AI-9 | Primary interface is a background assistant — "mais o assistente em segundo plano". A chatbot is explicitly a nice-to-have. A step-by-step form was offered and not chosen. |
| AI-10 | The AI may use previous-case backoffice data, not only this case's uploads. It must ship a disclaimer that the model is informed by that data ("based on X processed cases"). GDPR is flagged as an open risk with no decision taken. |
Open questions
These block design work. They are not defaults.
| # | Question | Owner | Status |
|---|---|---|---|
| Q1 | Does the historical corpus influence the answer, or only feed statistics? | — | ✅ decided since: D1 — it influences |
| Q2 | How to satisfy GDPR when feeding prior-case data to the model | — | ✅ decided since: D2 — anonymise at ingestion |
| Q3 | Is the workflow fixed, or per-client customisable? | — | ❌ open |
| Q4 | Examples of past cases that count as a "good analysis" | Rui Neves Ferreira · Ana Rocha Alves | ❌ open — blocks the precedent corpus |
| Q5 | What sample data exists today, in what format (PDF, Word, scans) | Rui Neves Ferreira · Ana Rocha Alves | ❌ open — blocks the precedent corpus |
| Q6 | Is the knowledge base good enough for the model to answer without leaning on its own parametric knowledge? | Rui Neves Ferreira · Ana Rocha Alves | ❌ open |
| Q7 | Who owns the mandatory human validation, per phase | — | ❌ open — AI-7 confirms it exists, not who |
Not stated in the document
Called out so nobody reads them in by accident.
| Often assumed | Reality |
|---|---|
| AI orchestrates deadlines and sends alerts | prazos e alertas appears only inside the question. The answer redirected to "all 8 phases" plus a background assistant. Never committed to. |
| A model or provider was chosen | Not named anywhere. |
| The like/dislike signal feeds retraining | How it's consumed afterwards is unspecified. |
Next: how these are served — Knowledge architecture.