Agency

I use AI every day. It drafts, searches, compares, writes code and turns a rough page of notes into something another person can read.

It also makes weak thinking look finished.

That is where founders and knowledge workers can lose the plot. The work arrives faster, the sentences sound confident and the spreadsheet has all the right colours. Meanwhile, nobody can explain where the central assumption came from or why the recommendation was accepted.

Agency disappears through a series of convenient little decisions. Let the model frame the problem. Let it choose the sources. Let it produce the options. Let it recommend one. Let it write the explanation. By the end, your name is on the document and very little of your judgement is inside it.

A 2025 study of 319 knowledge workers collected 936 first-hand examples of AI use at work. Higher confidence in AI was associated with less critical thinking, while greater confidence in one's own ability was associated with more. The work also shifted towards verification, integration and supervision.

Staying in the loop means keeping hold of three things: the problem, the standard of proof and responsibility for the decision. The machine can do an enormous amount around them. Those three remain yours.

Think before you prompt

Before opening an AI tool, write five lines:

• What decision am I trying to make?
• What do I know from direct experience?
• Which assumption worries me most?
• What evidence would change my mind?
• Who carries the consequence if I am wrong?

This takes a few minutes and gives you an independent starting point. Without it, the first coherent answer can become the frame for everything that follows. A polished recommendation has a way of making other possibilities feel less serious.

Then give the model your note and ask it to attack the weak parts. Ask for missing evidence, the strongest objection, plausible alternatives and the conditions under which your preferred option fails.

“Tell me what to do” is a poor prompt for an important decision. “Find the unsupported assumptions in this plan and show me what evidence would test them” gives the system a useful job.

For creative work, make your own ugly first move. Write the opening paragraph. Sketch the product. Name the tension. Choose three references that carry the taste you want. A rough first move records your taste and intention before the model fills the page.

AI can then expand the field without defining the whole field for you.

Give every AI session a named job

An AI assistant becomes dangerous when its role drifts. It begins as a researcher, turns into a strategist and ends up approving its own recommendation.

Name the job at the top of the session.

Researcher: collect evidence, preserve links, distinguish primary material from commentary and flag conflicts.

Drafter: turn an agreed argument into a first version without adding claims.

Critic: identify weak logic, missing perspectives, hidden incentives and possible failure modes.

Operator: complete a bounded task within stated permissions, then leave a record of what it did.

Keep the decision outside these roles. A founder can ask for a pricing brief, scenario model or contract comparison. A named person still chooses the price, approves the plan or signs the contract.

Write the boundary into the prompt. “Prepare three options. Do not recommend one until I have reviewed the evidence.” “Draft the customer reply. Do not send it.” “Analyse the agreement. Quote the relevant clauses and flag anything that needs legal advice.”

The boundary matters most on tired days. A capable system will always offer to carry one more part of the load. Convenience is persuasive.

Keep proof and dissent inside the workflow

Confident language is not evidence. A citation is not proof that the cited page supports the sentence beside it.

For any claim that can move money, affect a person or damage trust, open the original source. Check the date, scope and definitions. Recreate an important calculation. If the answer depends on a contract, read the clause. If it depends on customer behaviour, speak to a customer.

Set a verification budget before the work begins. A social caption may need a quick sense check. A hiring recommendation, financial forecast or legal position deserves independent review. The amount of checking should follow the consequence of being wrong.

Do not ask the same system to approve its own work. Use primary documents, a colleague with different incentives, a specialist or a separate method. A second paragraph with slightly different adjectives gives you no meaningful dissent.

Teams need a dissent habit too. Collect human views before showing the AI summary in a decision meeting. Otherwise the summary becomes an anchor and the room spends its time editing the machine's frame.

Give one person permission to ask:

• Which source would prove this wrong?
• What did the system fail to see?
• Are we optimising the measure while harming the outcome?
• Who has not been represented in the data?
• Can we reverse the decision?

Useful friction keeps a team awake. Ten minutes spent challenging a plausible answer can save weeks spent implementing the wrong one.

Protect the inputs that remain yours

A model can process an extraordinary amount of material. It receives only the version of a customer that reaches its context. It may notice hesitation in a recording, but it has no first-hand relationship with that person and no responsibility for what follows.

Founders need direct contact with the world their systems describe.

Keep one customer conversation each week without an automated script directing every question. Read one important document in full rather than relying on its summary. Walk through the process you plan to automate. Watch where a person waits, invents a workaround or gives up.

Knowledge workers need unsynthesised input too. Read the paper, judgement, annual report or interview. Spend time with people who disagree with your professional circle. Make notes before asking for a summary. A diet of generated explanations produces the intellectual equivalent of airport food: tidy, familiar and difficult to remember.

Protect a block of work where chat, search and notifications are closed. Forty-five minutes is enough to discover whether you understand the problem without assistance. Write, calculate, draw or think through it on paper.

Keep an embodied skill as well. Cook, repair something, play an instrument, work with wood, practise a sport. Reality gives immediate feedback and refuses to be impressed by confident wording. A golf ball has never accepted a persuasive excuse for going into the lake.

Audit the loop on yourself

AI systems improve through feedback. Your way of using them deserves the same attention.

At the end of each week, choose three AI-assisted pieces of work and write down:

• What did the system do?
• What did I decide?
• Which parts did I verify?
• Where did I correct it?
• What will I do differently next time?

Track a small set of signals for a month: time returned, errors caught after delivery, decisions reversed, corrections from colleagues and the number of important claims checked against original material. Add one qualitative signal: did the work teach you anything?

Watch for behavioural warning signs. You open the tool before you can state the problem. You struggle to explain a recommendation without the transcript. Your writing starts to sound the same across every subject. You have fewer conversations with customers and colleagues because summaries feel efficient. A blank page becomes uncomfortable without an assistant waiting beside it.

When one of these appears, reduce the system's authority. Move an automated action back to approval. Write the first draft yourself for a week. Remove the meeting summary and compare your own notes. Change the prompt so the model asks questions before producing an answer.

Design the workflow so agency survives deadline pressure and a very persuasive response.

A rhythm you can start on Monday

A practical routine does not require a new productivity system.

Each morning, define the important decision or creative problem in your own words before opening AI. During the day, label the system's role and set its boundary. Verify consequential claims at the source. Keep one protected block for unassisted thought.

At the end of the week, review three pieces of work and one decision. Look for time saved, errors introduced and judgement exercised. Speak to one person who experienced the outcome. Their response is often a better signal than the dashboard.

Once a month, remove one automation that has made the work faster and the thinking weaker. Increase authority only where the evidence shows that quality, control and reversibility have improved.

AI should return time, widen the range of options and remove administrative sludge. If it leaves you less able to explain your work, defend a choice or notice when the frame is wrong, the price is too high.

Use the machine for speed, breadth and repetition. Keep the problem definition, the standard of proof and the final responsibility in human hands. Then the loop still belongs to you.