the whole job

AI Act: the whole job, in order

If your staff use AI tools at work, you are probably a deployer, and some AI Act duties already apply to you. Here is the job as a plan: seven phases, in the order that holds up, with who carries each step, what it produces and the proof it leaves.

last checked 2 November 2026

how to read this plan

Each step has an owner, a kind of work and two outputs. The kinds of work are different jobs: choices that someone with authority signs, lists and settings that someone builds, and checks that someone keeps running. The first five phases apply to every company that uses AI tools. The people phase applies if you use AI on applicants or staff, and its duties start on 2 December 2027, after the July 2026 change. Under each step a small note names the legal basis, or says practice where no text sets it. Practice is still worth doing, and the difference tells you what a regulator could ask for.

  • decide and documentchoices, policies and records that someone with authority signs
  • build and configuresystems, settings and procedures that make the choice real
  • prove and operatechecks, logs and records that show it keeps working
01

settle who answers

Confirm that you are a deployer, name an owner and put the dates in one calendar.

  1. confirm that you are a deployer

    A deployer is a company that uses an AI system under its own authority, outside personal use. Check whether staff use any AI system for work. If the answer is no, write that down and check again once a year.

    • owner legal or DPO
    • decide and document
    what it produces
    a one-page note with the date and the reasoning
    the proof it leaves
    the note, approved by management

    the work page

    basis AI Act Art. 2(1), 3(4)

  2. name an owner for AI use

    Pick one person who answers for which AI tools the company uses and for what. Give that person the authority to say no, and decide who reports to the board on how the work is going.

    • owner board
    • decide and document
    what it produces
    a board decision naming an accountable owner
    the proof it leaves
    the minutes of that decision

    basis working practice. No legal text sets this.

  3. put the dates in one calendar

    Record the dates that touch a deployer: the bans and the literacy duty already apply, 2 December 2026 adds two bans, and 2 December 2027 starts the high-risk rules for the uses listed in annex III. Rules in this area have moved before, so note the day you checked.

    • owner legal or DPO
    • decide and document
    what it produces
    a dated list of the dates that apply to you, with the source of each
    the proof it leaves
    the list, rechecked on a stated date

    the work page

    basis AI Act Art. 113; Regulation 2026/1744 Art. 1

you are done with this phase when you can say in one page that you are a deployer, who owns AI use and which dates apply to you.
02

see what you use

List every AI tool in use, what it is used for and whether it touches people or personal data.

  1. list every AI tool in use

    Start from what you can export today: the software list from finance, the app list from your identity provider, the browser extensions on managed devices. Then ask each team which tools they use, including free ones and ones built into other software.

    • owner head of IT
    • build and configure
    what it produces
    an inventory of AI tools with an owner each
    the proof it leaves
    dated exports and the inventory

    the control: ai-inventory

    basis working practice. No legal text sets this.

  2. record purpose, data and users for each tool

    Next to each tool, write what it is used for, what data goes in, who uses it and whether its output reaches customers, applicants or staff. Mark every tool used to screen applicants or evaluate staff.

    • owner everyone
    • build and configure
    what it produces
    an inventory with purpose, data and users per tool
    the proof it leaves
    the inventory, confirmed by each team lead

    the control: ai-inventory

    basis working practice. No legal text sets this.

you are done with this phase when you can show a current list of AI tools, each with an owner and a purpose.
03

write the rules and train people

Agree which tools are allowed, for what, and train the staff who use them.

  1. write the rules for staff

    Decide which tools are allowed, what data may go into them, who approves a new tool and who staff ask when unsure. Say which results a person must check before they are used. Keep it to a page.

    • owner board
    • decide and document
    what it produces
    a one-page AI use rule, signed by management
    the proof it leaves
    the rule, with the date of approval and the version

    the work page

    basis working practice. No legal text sets this.

  2. run AI literacy training on the tools you use

    The duty is to take measures that support AI literacy of the staff who use these systems. The July 2026 text does not ask you to guarantee a set level for each person. A short session on your own tools and your own rules is a reasonable measure. Cover new joiners too, and keep the attendance list with the material.

    • owner HR
    • build and configure
    what it produces
    a training session and an attendance record
    the proof it leaves
    the record and the material with its version

    the work pagethe control: awareness-training

    basis AI Act Art. 4; Regulation 2026/1744 Art. 1

you are done with this phase when you can show written AI rules and a training record for the people who use AI tools.
04

check the red lines

Check your tools against the banned practices and the transparency duties, and find out whether any use is high-risk.

  1. check your tools against the banned practices

    Go through the list of banned practices once, tool by tool, and write down the result. Two bans start on 2 December 2026: AI that generates sexualised images of identifiable people without their consent, and AI that generates child sexual abuse material. Stop any use that falls under a ban.

    • owner legal or DPO
    • prove and operate
    what it produces
    a dated check of each tool against the list of bans
    the proof it leaves
    the check, signed by its author

    basis AI Act Art. 5

  2. tell people where transparency applies

    If you publish deep fakes, say so. If you use emotion recognition or biometric categorisation, tell the people it is used on. Find out from the inventory whether either applies, and if it does, write the notice.

    • owner legal or DPO
    • build and configure
    what it produces
    a list of uses that need a notice, and the notices
    the proof it leaves
    the notices as published or given

    basis AI Act Art. 50(3), 50(4)

  3. find out whether any use is high-risk

    Compare each marked use with the annex III list. Tools that filter applications or evaluate staff are on it. If a use is on the list, the next phase applies to it. If you decide it is not, write the reasoning.

    • owner legal or DPO
    • decide and document
    what it produces
    a decision per marked use, with reasoning
    the proof it leaves
    the decisions, approved by the owner of AI use

    basis AI Act Art. 26(1); AI Act Annex III

you are done with this phase when you can show a dated check of banned practices and a decision on each use that may be high-risk.
05

ask your suppliers

Put the right questions to every AI provider and keep the answers.

  1. send the same questions to each AI provider

    Ask each provider: Does it say whether the system is high-risk under annex III? Where are the instructions for use? Which logs can you export and for how long? How does it handle serious incidents and tell you? Is your data used to train the system? Does it mark generated content?

    • owner procurement
    • build and configure
    what it produces
    a standard AI questionnaire and the answers received
    the proof it leaves
    the questionnaire and the answers, dated

    the work pagethe control: supplier-management

    basis working practice. No legal text sets this.

  2. put the answers into contracts and the register

    Where a tool handles personal data for you, check whether the provider acts as a processor and whether the contract carries the required terms. Record each answer in the supplier register and add what you need, such as log access, to the contract at renewal.

    • owner procurement
    • build and configure
    what it produces
    an updated supplier register and a list of contract changes
    the proof it leaves
    the register, and the signed terms where they changed

    the work pagethe control: supplier-management

    basis GDPR Art. 28(3)

you are done with this phase when you can show answers from each provider and the contract terms that follow from them.
06

if you use AI on applicants or staff

Meet the deployer duties for high-risk uses before 2 December 2027.

  1. decide whether the use stays

    For each high-risk use, decide whether to keep it, replace it or stop it before the duties start on 2 December 2027. Compare the effort in the steps below with what the tool gives you.

    • owner board
    • decide and document
    what it produces
    a board decision per high-risk use
    the proof it leaves
    the minutes of that decision

    basis AI Act Annex III

  2. get the provider's instructions and follow them

    Ask the provider for the instructions for use and keep the current version. Take measures to use the system as they say. If you cannot say what the instructions are today, finding out is the first job.

    • owner procurement
    • build and configure
    what it produces
    the provider's instructions on file, and a note of how you follow them
    the proof it leaves
    the instructions with their version and the date received

    the work pagethe control: supplier-management

    basis AI Act Art. 26(1)

  3. give oversight to someone competent

    Assign human oversight to a named person with the competence, training and authority to act, with authority to act on what the system does. Make sure that person has the time.

    • owner board
    • decide and document
    what it produces
    a named overseer for each high-risk use
    the proof it leaves
    the appointment, and the training the person received

    basis AI Act Art. 26(2)

  4. tell workers and their representatives first

    Before you put a high-risk system into use at work, inform the workers affected and their representatives. Say which system, for what purpose and who to ask. Check whether your national law or an agreement adds consultation duties.

    • owner HR
    • build and configure
    what it produces
    a notice to workers and their representatives
    the proof it leaves
    the notice with the date it was given

    basis AI Act Art. 26(7)

  5. keep the logs you control

    Find out which logs the system produces and which of them sit under your control. Keep those for at least six months, unless another law sets a different period. Agree with the provider how you get the rest.

    • owner head of IT
    • build and configure
    what it produces
    a log retention rule for each high-risk use
    the proof it leaves
    a sample export showing the retention in effect

    the control: logging-monitoring

    basis AI Act Art. 26(6)

  6. monitor the system and report serious incidents

    Monitor operation against the instructions. If you identify a serious incident, report it to the provider and to the authority. Add this route to your incident procedure and name who sends the report.

    • owner head of IT
    • build and configure
    what it produces
    a monitoring routine and a reporting route with named roles
    the proof it leaves
    the procedure, and a dry run of the report

    the work pagethe control: incident-management

    basis AI Act Art. 26(5)

  7. keep personal data in these systems under GDPR

    Applicant and staff data in these tools stays under GDPR. Add each tool to your risk assessment with the risk to the people concerned, and check the security measures fit that risk.

    • owner legal or DPO
    • build and configure
    what it produces
    each tool in the risk assessment with a security decision
    the proof it leaves
    the assessment, dated and approved

    the work page

    basis GDPR Art. 32(1)

you are done with this phase when you can show instructions on file, a named overseer, a notice to workers, logs and a monitoring routine for each high-risk use.
07

show it keeps working

Keep the inventory current, keep the proof together and report to management.

  1. keep the inventory current

    Review the inventory on a set rhythm and whenever a team adopts a new tool. Compare it with the software list from finance, and ask about anything that shows up in one and not the other.

    • owner head of IT
    • prove and operate
    what it produces
    a review rhythm and a dated review record
    the proof it leaves
    the review records

    the control: ai-inventory

    basis working practice. No legal text sets this.

  2. keep the proof in one place

    Collect the outputs of every step above in one folder: the inventory, the rules, the training record, the banned-practice check, the notices, the instructions and the provider answers.

    • owner head of IT
    • prove and operate
    what it produces
    an evidence folder organised by step
    the proof it leaves
    the folder, with each item dated

    basis working practice. No legal text sets this.

  3. report to management and recheck the dates

    Once a year, show management the list of tools, the open gaps and the dates ahead. Recheck the dates against the current text of the AI Act, because they were changed in July 2026. Record what management decided.

    • owner board
    • prove and operate
    what it produces
    a yearly report to the board
    the proof it leaves
    the report and the minutes

    the work page

    basis working practice. No legal text sets this.

you are done with this phase when you can hand a regulator or a customer one folder that shows each step in this plan.

who carries what

controls you will need

Each control has its own page: what it is, what to look for, where it goes wrong and the proof it leaves.

    check it yourself

    your next step

    List every AI tool in use at your company this week, including the ones nobody approved, and mark any used on applicants or staff.