AI will not take the job tomorrow. It moves the tasks today

It is tempting to talk about AI as if it arrives from the outside and does something dramatic to entire jobs. For a small team, that is usually the wrong level.

The first change is more ordinary. A quote needs to be drafted faster. A meeting needs to turn into a list of decisions. A school administrator wants to stop digging through three documents before replying to a parent. A restaurant owner wants to see which bookings, purchases and staffing changes actually need a decision before Friday.

In June, OpenAI published an EU version of its AI Jobs Transition Framework. The useful point is not that anyone can predict exactly which jobs disappear. The report says the opposite: the categories are a planning map, not forecasts. It sorts work into four broad paths: work that may grow with AI, work with higher automation potential, work likely to be reorganized and work where change is slower.

For Hammer readers, the practical lesson is this: do not start with "which jobs will AI replace?" Start with "which tasks in our week should AI help with, and where must a human still own the decision?"

Source: The AI Jobs Transition Framework for the EU (OpenAI Economic Research, June 2026).

Today's signal: build a role map, not a twenty-page AI strategy

OpenAI's report uses ESCO, the EU taxonomy for skills and occupations, together with Eurostat data. It estimates that about 27 percent of EU employment sits in occupations where work is likely to be reorganized by AI. About 14 percent sits in the higher automation-potential group, 12 percent in the group that may grow with AI, and 47 percent in work with less immediate change.

It is easy to get stuck in the percentages. Don't.

For a three-person company, a school, a consultant, an association or a small shop, the value is not sorting the whole labor market. The value is borrowing the logic and making it smaller. Replace "occupation" with "recurring task".

Look at the next two weeks. That is usually enough:

• Customer questions that always need the same first triage
• meeting notes that never become tasks
• quote requests that start in email and end in manual copying
• student, course or participant follow-up scattered across documents
• bookings, schedule changes and purchases someone is keeping in their head
• reports written late because the sources were messy from the start

This is where AI becomes concrete. AI can take the first move in a clearly bounded task, while a human still owns the role.

Source: The AI Jobs Transition Framework for the EU (OpenAI Economic Research, June 2026).

A daily example: the meeting that already writes the next step

In the same coverage window, Google wrote that Gemini can now take notes in Google Meet for Google AI Pro and Ultra subscribers, as well as eligible Workspace customers. The feature transcribes, summarizes and saves notes in Google Drive. After the meeting, it sends a summary with action items.

This is not a grand strategy. It is a small but important workflow change.

If a meeting used to end with "someone should write this up", it can now end with a draft of decisions, next steps and owners. That is exactly why a role map matters. Gemini can write the note. It should not automatically promise a delivery date to a customer, change an invoice, move a student to another group or send external emails without review.

Humans should not write everything from scratch. Humans should own the parts that are sensitive, expensive, customer-facing, or hard to undo.

Source: Gemini can now take notes in Google Meet for Google AI Pro and Ultra subscribers (Google, June 29, 2026).

45 minutes: build your first AI role map

You can do this in a spreadsheet, Notion, a whiteboard or a plain document. The tool is not the point. The point is that the team stops talking about AI in general and starts talking about work that actually exists.

0 to 10 minutes: write down ten tasks from the week

Choose recurring work. Not "become more digital". Write "summarize customer meeting", "reply to first price question", "sort support email", "draft lesson plan from course goals", "compare purchase list with inventory".

10 to 20 minutes: sort each task into one of four boxes

Use the four paths from the OpenAI report, but at task level:

• May grow with AI: tasks you do too rarely today because they take too long, such as better follow-up after every course or customer meeting.
• Can be partly automated: repetitive steps with clear rules, such as the first triage of incoming requests.
• Needs to be reorganized: work where AI can draft, collect sources or suggest next steps, but people still need to bring relationship, accountability or judgment.
• Wait on this: tasks that are too sensitive, unclear, or dependent on spoken context right now.

20 to 30 minutes: write a stop point for each task

A stop point is where AI must hand over. Examples: "AI may draft the reply, but not send it." "AI may suggest a schedule change, but not book it." "AI may summarize the meeting, but the responsible person approves decisions and next steps."

30 to 40 minutes: choose one first workflow

Do not choose the most impressive one. Choose the task that is boring enough, frequent enough and easy enough to review. A good first workflow saves 20 minutes a week without creating new risk.

40 to 45 minutes: define the receipt

Every AI workflow needs to leave a short receipt: source, what AI did, what it did not do, who approved it and the next step. Without a receipt, AI becomes another black box.

Copy-paste prompt: let AI draft the first role map

Paste five to ten recurring tasks. Replace the brackets with your real context.

You are my workflow assistant. Help me create a simple AI role map for [team/company/school].

Here are recurring tasks from our week:
[list of tasks]

For each task, suggest:
1. Which category fits best:
   - may grow with AI
   - can be partly automated
   - needs to be reorganized
   - wait on this
2. Why it fits there, in two short sentences.
3. What AI may do in a first test.
4. Which stop point requires human approval.
5. Which sources AI must use.
6. Which receipt the workflow should leave after it runs.

Keep the suggestions practical. Do not write a large AI strategy. At the end, choose one first test that can run in 45 minutes with low risk and clear value.

Three small workflows that fit well first

Meeting to decision card

AI may summarize meeting notes, find decisions, list open questions and suggest owners. A human approves what was actually decided before anything moves on.

Customer question to first reply draft

AI may read the customer's question and suggest a reply using approved sources: price list, terms, previous FAQ and internal instructions. It may not change pricing, promise delivery or send the reply without approval.

Course or class follow-up to action list

AI may summarize anonymized notes, find recurring blockers and suggest the next teaching step. The teacher or course lead owns the assessment and the contact with students, participants or guardians.

When the role map becomes integration

Sooner or later, someone will want to connect AI to real systems: Drive, CRM, finance tools, learning platforms, email, calendar or ticketing. That is reasonable. This is also where the role map earns its keep.

Start with the smallest useful access. Use read access when it is enough. Put API keys in environment variables or a secret manager, not in chat. Give AI a bounded account or scoped permission. Redact unnecessary personal data. Put approvals before external emails, payments, bookings, price changes, publishing and deletion. Save a run log.

This is not a brake. It is what lets AI do useful work without the team losing control.

For Hammer, this kind of exercise often sits between Mindset Forge and Tool Forge. First the team needs to understand which tasks fit AI. Then it can build one small, reviewed workflow. If that workflow returns every week, it becomes Skill Forge: a habit, an instruction and a run log that people actually use.

Start with the role map. It is simple enough to do today, but concrete enough to show what AI should be allowed to do tomorrow.