Baseline Checklist

Minimal risk ≠ zero rules

Even one everyday AI tool — like Copilot in Outlook or AI meeting notes in Teams — comes with a few basic duties. These six points are the minimum, each linked to the EU AI Act.

0 / 6 in place
  1. 1

    Keep an internal AI inventory

    Strongly expected

    A simple list of every AI tool in use: name, vendor, purpose, who has access, and its assessed risk level.

    In practice

    One spreadsheet row per tool — e.g. “Copilot in Outlook · Microsoft · email drafting · all staff · Minimal”.

    Finance example

    A bank adds Copilot to the same register it keeps for cloud vendors under DORA.

    EU AI Act reference

    • Art. 6 & Annex III — You can only show a tool is not high-risk if you know it exists and have classified it.
    • Art. 26 — Deployer duties for high-risk AI start with knowing what you deploy — the inventory catches tools that drift upwards.
  2. 2

    Train staff in AI literacy

    Legal duty

    Everyone with a licence understands the basics: how generative AI works, that it can hallucinate, and its limits.

    In practice

    A short onboarding session or one-page guide before Copilot is switched on for a user.

    Finance example

    Advisers learn that a Copilot summary of a client email may drop a key figure.

    EU AI Act reference

    • Art. 4 — Deployers must ensure a sufficient level of AI literacy among staff — applies at every risk level since Feb 2025.
    • Art. 3(56) — Defines AI literacy: the skills and understanding to use AI in an informed way and be aware of its risks.
  3. 3

    A human owns every output

    Strongly expected

    AI drafts, people decide. Nothing goes to a customer or into a decision without someone reading it first.

    In practice

    House rule: “Read and approve before you hit Send.”

    Finance example

    A Copilot-drafted reply to a loan enquiry is checked by the case officer before sending.

    EU AI Act reference

    • Art. 14 — Human oversight is mandatory for high-risk AI — applying it voluntarily keeps minimal-risk tools from sliding up a level.
    • Recital 165 — Encourages applying high-risk safeguards voluntarily to other AI systems through codes of conduct.
  4. 4

    Set clear data boundaries

    Legal duty

    Staff know what must never be pasted into a prompt or summarised by AI.

    In practice

    A short “never” list: health data, inside information, passwords and API keys, unencrypted client contracts.

    Finance example

    Summarising a public market report: fine. Uploading a customer list with ID numbers: not allowed.

    EU AI Act reference

    • Art. 2(7) — The AI Act does not replace GDPR — data protection law applies in full to data processed by AI.
    • Art. 10 — Data governance rules for high-risk AI set the benchmark for handling data responsibly.
  5. 5

    Use enterprise licences, not shadow AI

    Strongly expected

    Only approved tools under a business agreement, so company data isn't used to train public models.

    In practice

    Ban pasting work content into private, free AI accounts. Check the vendor's data-processing terms.

    Finance example

    Copilot under the firm's Microsoft 365 contract: approved. A personal free chatbot account: not approved.

    EU AI Act reference

    • Art. 25 — Sets out responsibilities along the AI value chain — deployers should know what their provider guarantees.
    • Art. 53 — General-purpose AI providers (behind tools like Copilot) must give downstream users documentation on capabilities and limits.
  6. 6

    Write a one-page AI policy

    Strongly expected

    Short internal rules: approved tools, who approves new ones, and how to report mistakes.

    In practice

    One page on the intranet, owned by a named person, reviewed yearly.

    Finance example

    An insurer's policy names the CISO as approver and asks staff to report data slips within 24 hours.

    EU AI Act reference

    • Art. 50 — Transparency duties — the policy should say when AI-generated content must be disclosed to customers.
    • Art. 95 — Encourages voluntary codes of conduct for non-high-risk AI — your policy is your internal version.

    Ready-to-use template

    [Company Name] – Workplace AI Policy

    Version: 1.0 | Effective: [Date] | Owner: [CISO / Head of IT / Compliance]

    Applies to: All employees, consultants and temporary staff

    1. Purpose & scope

    This policy defines the acceptable use of everyday AI assistants (Microsoft 365 Copilot in Outlook, Teams, Word and Excel). We do not develop our own AI models, nor do we use automated decision-making that affects customers or employees.

    2. Approved tools only (no "shadow AI")

    • Permitted: company-provided enterprise tools with commercial data protection (your corporate Microsoft 365 Copilot account).

    • Prohibited: personal, free or unvetted AI tools (e.g. free ChatGPT, personal Claude/Gemini accounts, browser plugins) for any work-related task. Free consumer tools may use your data to train public models.

    3. What can and cannot go into AI

    • Allowed: drafting internal emails and documents, summarising internal meetings, analysing internal spreadsheets, proofreading.

    • Never: passwords, keys or IT infrastructure details; special category personal data (GDPR Art. 9); non-public financial results, M&A information or sensitive customer contract terms.

    4. Human accountability

    • You own the output: you are fully accountable for anything you share, even if AI drafted it.

    • Always verify facts, names, figures and sources – AI can "hallucinate".

    • No fully automated customer actions: a human reviews and sends.

    5. Transparency (EU AI Act Art. 50)

    • Disclose AI use when a client, partner or authority would reasonably expect the work to be your own.

    • Never present unedited AI output as professional advice.

    6. Incidents, new tools & review

    • Report sensitive data entered into an unapproved AI tool to [security@company.com] within 24 hours.

    • New AI tools or extensions must be approved by [Named role / IT].

    • This policy is reviewed annually and supports AI literacy under EU AI Act Art. 4.

Legal duty means the law requires it today (AI Act Art. 4 or GDPR). Strongly expected means it isn't a direct AI Act duty for minimal-risk tools, but you'll struggle to prove compliance — or catch a tool drifting into high risk — without it. Finance firms should also check DORA's register of ICT providers.

Not sure your tool is minimal risk?

Answer ten quick questions to check its likely risk level.

Take the self-assessment