The AI User Is Disappearing. The AI Manager Is Arriving.

Same human. A bigger role.

Picture a Monday morning at a training institute. A new course needs a launch campaign. There are competitor pages to review, learner questions to sort, a course brief to check, posts to draft and a calendar to prepare. Someone must keep the whole job moving.

An AI user might open a chatbot and ask for ten campaign ideas. Then ask for a caption. Then ask for an email. Each answer helps, but the person still carries the plan in their head and moves the pieces between applications.

An AI manager starts with the assignment: prepare a campaign pack for this course, for this audience, using these approved facts, and return it for review. Where the software has the necessary tools and access, AI can handle connected parts of the work. The person defines success and decides whether the result meets it.

That is the shift this Sunday Read is about. The headline describes a changing habit, not a claim that ordinary AI users or human jobs are literally disappearing. “AI manager” is a way of working; you do not need a management title to practise it.

Your job is to define what good work looks like — and check that it has been delivered.

From answers to connected work

Think of five ways AI can fit into a job. A chatbot answers a question. A copilot assists inside a task. A workflow follows a set sequence. An agent can choose steps and tools within its instructions. An “AI worker” is an informal label for a system assigned a recurring area of work, with a person responsible for its results.

These are overlapping patterns, not a dated timeline or a ladder everyone must climb. A short question may need only a chatbot. A predictable process may work better with a fixed workflow. Adding an agent makes sense when the next step depends on what the system finds.

Anthropic’s guide distinguishes predefined workflows from agents that direct their own tool use. It also advises starting with the simplest working approach. That distinction matters when deciding what to delegate: a polished chat interface alone does not provide access to your files, business applications or publishing system.

• Read Anthropic’s explanation of workflows and agents

The skill is writing a usable assignment

Good prompting still matters. What changes is the size of the responsibility you describe. “Write a report” leaves AI guessing about the reader, evidence, format and decision. A useful assignment supplies those details before the first draft.

Start with the result someone will actually use. A manager may need a one-page decision brief and a source table. A trainer may need a lesson plan, practice file and answer key. A sales team may need a checked list of enquiries and draft replies. Name the deliverable rather than asking for a discussion of the topic.

Then provide the working context: approved documents, a good example, current rules and the audience. Say which source wins if two documents disagree. State how missing information should be marked. This reduces the chance that a confident guess becomes a business fact.

Finally, define “done”. A launch pack is complete when its claims match the course brief, every required asset is present, links work and unresolved questions are visible. Length or polish alone is a weak completion test.

• Outcome: what finished work should exist?
• Inputs: which files and facts should guide it?
• Boundaries: what may AI read, change or send?
• Quality: how will you judge the result?
• Handoff: what needs your decision?

The AI Manager Loop

Use a repeatable loop: goal, plan, act, review, improve. The person remains accountable across all five stages. The loop is our practical teaching framework, not an industry certification or a guarantee of reliable automation.

Goal: describe the outcome and why it matters. Plan: ask for the proposed steps, required inputs and likely gaps. For familiar low-risk work, the system can proceed within previously agreed rules. A new or sensitive assignment needs a closer look at the plan.

Act: let AI do the permitted work. Ask it to report blockers instead of quietly changing the assignment. Review: compare the deliverables with your acceptance criteria and inspect the evidence behind the important claims. Improve: correct the brief or workflow so the same mistake is less likely next time.

“Improve” does not mean the underlying model automatically learns permanently from every correction. Save the corrected template, rules or examples in the system you actually use. If you do not retain that knowledge, the next run may repeat the same error.

Before vs Now: one course launch

Consider a fictional institute preparing a weekend course launch. The inputs are a checked syllabus, trainer notes, audience questions, brand rules and confirmed schedule. The examples below describe a possible workflow, not measured results from an AI5 campaign.

In the first approach, a person asks for ideas, chooses one, requests captions, rewrites the tone, pastes dates, prepares a calendar and checks the whole pack. AI helps at each step, but the person must repeatedly supply context and coordinate every handoff.

In the second approach, the person assigns a complete campaign pack: a message angle, three post drafts, an email draft, a seven-day calendar and a claim-check table. AI works from the same source pack throughout. If a fee or date is missing, it marks the gap. Publishing remains a separate authorised action.

The review becomes specific: does each post describe the actual course? Does the promise match what students will learn? Are the dates consistent? Does the call to action lead to the right page? A campaign can sound excellent and still fail all four checks.

The payoff to measure is the time required to get a correct, usable pack. More generated content is not automatically a better result. If checking and repairing the pack takes longer than preparing it yourself, reduce the scope and fix the process.

Three more assignments you can practise

Trainer preparation: give AI a lesson objective, learner level and approved syllabus. Ask for a 60-minute lesson plan, one worked example, a short practice task and an answer key. Require it to flag any prerequisite the learners may not have. The trainer checks factual accuracy, pacing and whether the exercise can actually be completed.

Sales operations: use a small anonymised enquiry file. Ask AI to group requests by course interest, list missing details and draft a follow-up queue. Keep ambiguous records in a separate group. A counsellor checks the grouping and selects the next action. Do not turn an uncertain interpretation into a confirmed customer preference.

Weekly reporting: provide a checked export and the definitions of each metric. Ask for a summary, a comparison with the previous period and a table of exceptions. Require every total to trace back to the supplied records. The person reviews the calculations and determines what the changes mean for the business.

Each assignment combines knowledge of the job with a clear review method. You can start manually with a file-enabled assistant. Connecting apps, scheduling repeated runs and allowing changes to live records are separate setup decisions; writing a prompt does not create those capabilities.

What humans keep. What AI gets.

The useful division is about responsibility. People set direction, make decisions that affect others, judge trade-offs and own the outcome. AI can help gather material, produce alternatives, compare records and carry out permitted repeatable steps. The exact split depends on the task and the tools.

A system can propose a shortlist; a hiring decision still needs a person’s judgement and an appropriate process. It can draft a reply; someone must own the promise made to the customer. It can flag a strange number; that does not prove the cause of the problem.

The most practical boundary is to separate preparation from commitment. Preparing a draft, analysis or proposed edit is different from sending it, spending money or changing an official record. Give permission for the actions needed, and define where the system should stop.

The mistakes an AI manager must catch

The first mistake is giving a vague objective and judging only the writing. Smooth language can hide weak evidence, missing rows and unsupported assumptions. Ask to see the source table, exception list or calculation behind the summary.

The second is treating a successful demonstration as a dependable process. Try a normal case, a missing-input case and a contradictory-input case. Does the system ask, flag or stop when appropriate? A workflow that fails clearly can be easier to manage than one that silently fills gaps.

The third is granting broad access before you know what the task needs. Start with a copy of the relevant files and only the permissions required. Decide what happens if an app is unavailable or a source cannot be checked. Do not let an unavailable tool turn into a claim that the action was completed.

The fourth is accepting an AI self-check as the only check. A second pass may catch mistakes, but a model can repeat its own wrong assumption. For the facts that matter, compare against original records, use a calculation or inspect the actual output yourself.

Try this today: a 30-minute Sunday challenge

Choose one task you repeat every week: a meeting brief, a lesson plan, a content calendar or a simple report. Pick something with a small, clear output that you can judge yourself. Use sample or anonymised material if the working data is private.

Spend five minutes collecting the inputs and writing three acceptance criteria. Spend five minutes turning them into an assignment. Give AI ten minutes to prepare the output, allowing for your tool’s speed. Use the final ten minutes to review it and record the changes needed.

Do not aim for an entire automated department in half an hour. Aim for one useful assignment you can run again. Keep the starting brief, final result and corrections together. That small record is the beginning of your own working playbook.

Copy the brief below and replace the bracketed fields. If your assistant lacks file or app access, supply the relevant text or ask for a draft you can use manually. Never assume it has completed an action merely because it describes the steps.

A reusable AI work brief

Use this template for your first assignment. The review questions are part of the job, not an extra request after something goes wrong.

Goal: Produce [deliverable] for [audience and purpose].
Inputs: Use [approved files, examples and sources].
Rules: Follow [tone, format and business constraints].
Missing facts: Mark them; do not invent them.
Permissions: You may [allowed actions].
Stop before: [actions requiring my decision].
Method: Plan the work, complete permitted steps and flag blockers.
Quality checks: [three specific acceptance criteria].
Handoff: Return the output, evidence, assumptions and open questions.
Review: Check every requirement and state what remains incomplete.

Measure useful work, including the review

Compare the complete process with your usual method. Record preparation time, AI execution time, review time and repair time. Also note whether the final output passed each acceptance criterion. Use the same task type and a similar amount of input when comparing runs.

For example, a hypothetical report might take 45 minutes manually. An AI-assisted run could take five minutes to brief, five to execute and 20 to review and repair: 30 minutes in total. The saving is 15 minutes, not 40. Those numbers are an illustration; measure your own work before making a business claim.

Watch for quality costs as well as time costs. Were sources missing? Did a colleague have to recheck everything? Did the result create more follow-up questions? A shorter first draft does not help if the team spends longer resolving mistakes.

After three runs, decide whether to keep the workflow, change it or stop it. You may find that a simple template gives most of the benefit. Good management includes knowing when more automation is unnecessary.

Build the skill one assignment at a time

A useful learning progression is prompting, AI tools, AI workflows, AI agents, then managing AI work. Start by making instructions clear. Learn what your chosen tools can actually do. Connect a few repeatable steps. Only then consider a system that chooses some of its own actions.

For a beginner, the strongest first project is often a modest recurring task with a visible result. Save evidence of your work: the brief, source pack, draft, review checklist and final output. A portfolio built around checked results tells a stronger story than a folder of copied prompts.

The person who can explain the goal, organise the inputs, set limits and judge the output has a useful role in an AI-assisted team. Your knowledge of the work remains central. AI gives you more ways to put that knowledge to use.

This Sunday, choose one task and give it a proper brief. Review what comes back. Improve the instructions. Then repeat. That is how you begin managing AI work.

• AI Agents for Business: practise supervised business workflows
• Generative AI for Business: build your everyday AI skills
• AI5 free courses: start with a practical exercise