Too often, AI is used like a vending machine.
Insert question. Receive answer. Paste and move on.
It is fast, certainly. But it is also how you end up with a feed full of posts that sound as though they were written by the same enthusiastic robot.
The problem is not always that the writing is bad. In many cases, it is very good. Almost too good. The sentences are clean, the structure is logical and the conclusion arrives precisely where you expect it to.
The problem is that very little original thinking may have happened along the way. The AI tool did the work you were supposed to do, and you signed your name to it anyway, with little or no input or review.
There is a better way to use AI, and it is not simply about writing better prompts. It is about changing your relationship with the technology so that it does less of your thinking for you and more thinking with you.
Scaffolding, not a finished building
Think of AI as scaffolding, not the building itself.
Scaffolding is useful precisely because it is temporary. It holds things up and acts as a temporary frame for the structure while the real work goes on behind it.
Once the building can stand on its own, you take the scaffolding down. Ask yourself: do you admire the craftsmanship of the finished building or the scaffolding that once surrounded it?
Too often, we leave the scaffolding up and call it the building.
The first draft becomes the final draft. The first answer becomes the final answer. Nothing gets tested, refined or properly thought through. It gets rubber-stamped and shipped.
Used well, AI is genuinely useful scaffolding. It can turn a shapeless problem into something workable in seconds. It gives you an argument to interrogate, a structure to rearrange or a draft to improve.
But a first draft is a starting point, not the finished product.
Take this post. It began as a collection of notes and ideas that I wanted to develop into an article. I fed those thoughts into an AI thought partner. After some back and forth, it helped me create the scaffolding.
It did not remove the need for me to think. It gave me something against which I could think more clearly.
A framework for thinking with AI: CRIT
Geoff Woods, author of The AI-Driven Leader, offers a useful framework for structuring this kind of conversation. He calls it CRIT: Context, Role, Interview and Task.
-
Context
Do not begin with a bare question. Give the AI the real situation. Explain the background, the constraints, what you have already tried and what is actually at stake. The quality of the response depends heavily on the quality of the context you provide.
-
Role
Tell the AI what kind of expert or perspective you want it to adopt. That might be a marketing strategist frustrated by low-quality content, an ambitious business development professional, a sceptical CFO examining the return on an investment or a hiring manager who has reviewed a thousand CVs. The role determines the perspective from which the problem is examined.
-
Interview Where the value lies
Instead of allowing the AI to jump immediately to an answer, ask it to interview you first. Its questions should uncover assumptions, missing evidence, unresolved trade-offs and parts of the problem you have not considered properly. This is where much of the value lies. The interview forces you to think before the AI starts producing.
-
Task
Only after that do you define what you want the AI to do. The order matters. Context and Role shape how the problem is understood. Interview exposes the gaps in your reasoning. Task turns that deeper understanding into an output. Many shallow AI interactions begin with the Task and skip everything that should have come before it.
Understanding the problem is not enough
CRIT helps the AI understand the problem, but understanding is not the same as challenging.
If you want an AI that genuinely argues back, you must give it explicit permission to disagree with you.
Ask it to identify the strongest counterargument to your position. Tell it to expose the assumption on which your reasoning depends. Ask what evidence would prove you wrong.
Most importantly, tell it not to manufacture disagreement merely for the sake of appearing critical. You do not need a tool that contradicts everything you say. You need one that recognises when your reasoning does not hold up.
A basic instruction might look like this:
Thought-partner prompt
You are my thought partner, not simply my assistant. When I bring you a problem or a plan, do not give me an answer or a polished draft immediately.
First, ask questions that expose gaps, assumptions or issues I have not considered. For each recommendation I make, identify the strongest counterargument, the assumption on which my reasoning depends and the evidence that would prove me wrong.
Push back when you disagree and explain why. Do not manufacture disagreement, but do not accept a claim merely because I stated it confidently.
Support factual claims and counterarguments with links to the original or most authoritative available sources. Distinguish clearly between verified facts, reasonable inferences, personal opinions and unresolved uncertainty. If you cannot verify something, say so plainly. Never invent a source, quotation or statistic.
Only after I have worked through your questions should you help me shape a final answer. Even then, tell me candidly where the reasoning remains weak or dependent on unproven assumptions.
The requirement for evidence matters because confident disagreement is no more useful than confident agreement if neither is grounded in verifiable information.
That is deliberately basic.
A more complete version would specify the role the AI should adopt, the industry in which you work, the types of decisions you make and what you are trying to optimise for. You might also include examples of the questions you want it to ask.
The important point is that the prompt should not be treated as a finished product either. It is scaffolding too.
The same pattern works elsewhere
Once you recognise this pattern, you begin to see where else it can be applied.
As a study partner
If you are trying to learn something properly, rather than simply finish an assignment, you can create an agent with two distinct modes.
It does not give you the answer or complete the work for you. It asks questions until you have worked through the idea and explained it in your own words.
You deliberately switch modes, give it the work you produced and ask it to critique the argument, identify gaps and evaluate it honestly against the relevant criteria.
Keeping the two modes separate matters. If the same interaction is simultaneously coaching and grading you, the feedback may be influenced by the path the conversation took.
I used this approach during a recent course and found it invaluable. I uploaded the grading rubric as part of the agent’s instructions and asked it to evaluate my work in assessor mode against that rubric.
It did not eliminate the need for judgement, but it gave me a structured way to test whether my understanding held up.
As a proposal partner
The same logic applies at work.
Instead of creating an agent that immediately drafts a proposal, build one that first interviews you about the client’s actual problem. It should pressure-test your value proposition and challenge you with questions such as:
- What is the client’s actual problem?
- What happens if the client does nothing?
- Why should the client choose you over the available alternatives?
Only after you have answered those questions should it help construct the proposal.
The resulting document is more likely to reflect a real commercial argument rather than a polished collection of generic claims. It is also more distinctly yours because the substance came from your thinking, not merely from the model’s ability to generate convincing prose.
The pattern applies to almost any recurring task in which you might be tempted to let AI bypass the difficult part: strategy documents, important decisions, challenging conversations or your own reflection on a mistake.
Why this can make your thinking better
Used in this way, AI functions as a sparring partner rather than a substitute.
Every time it questions an assumption instead of accepting it, you complete a small repetition of critical thinking that you might otherwise have skipped.
Done consistently, this can improve both the thinking and the output that follows from it. Proposals, decisions and articles begin to sound more distinctly like you because more of your tested reasoning survives into the final version.
Build a recognisable point of view not by publishing more material, but by allowing more of your own thinking to reach the finished work.
What about the jobs?
It would be dishonest to argue for greater use of AI without acknowledging that it will change work. In some roles, that change will be significant.
History does not provide a simple promise that every worker or occupation will benefit. What it does show is that technology can both automate existing tasks and create new forms of work.
of US employment in 2018 was in job specialities that did not exist in 1940.
Autor et al., NBERof children entering primary school may work in job types that do not yet exist, a popular estimate, not a WEF finding.
Cited in WEF Future of Jobs 2016Research led by MIT economist David Autor found that about 60 per cent of US employment in 2018 was in job specialities that did not exist in 1940. The same research also distinguishes between innovations that augment workers and those that automate tasks, noting that the effects of automation on labour demand have intensified in recent decades.
These concerns were already evident in major future-of-work research before the emergence of generative AI. The World Economic Forum’s 2016 and 2018 Future of Jobs Reports identified artificial intelligence, machine learning, big data analytics and cloud technologies as key drivers of business transformation and workforce change. They anticipated that new job categories would emerge, existing roles would be displaced or redefined, and the skills required by workers would change significantly.
The 2016 report also cited a widely referenced estimate that 65 per cent of children entering primary school would eventually work in job types that did not yet exist. Importantly, the report described this as a popular estimate rather than one of its own findings.
What those reports could not fully predict was the specific form that AI adoption would take. They anticipated that artificial intelligence and machine learning would become important drivers of labour-market change, but they did not foresee the speed with which generative AI tools would become accessible to millions of knowledge workers or how quickly AI-assisted content creation, research and decision support would enter everyday professional workflows. The broader prediction that technology would reshape work proved accurate; the precise mechanisms and pace of adoption were harder to foresee.
Using AI as a thought partner does not resolve the wider employment debate. It does, however, offer one practical way to focus on augmentation: allowing the technology to support research, structure and challenge while keeping responsibility for judgement with the person using it.
You already know more than you think
There is one final point that most guides to AI overlook: not every question requires AI assistance.
In the rush to “prompt better”, it is easy to forget that you have spent a career, or perhaps a lifetime, developing expertise, instincts and judgement.
If you already know the answer, trust yourself.
Use AI to pressure-test the reasoning, sharpen the explanation or identify a blind spot. Do not automatically use it to outsource the thinking you are already capable of doing.
The goal is not to make AI sound more like you. It is to use AI in a way that makes you think more clearly.
Let it question the structure, test the assumptions and expose the weak joints. Then take the scaffolding down and put your own name on the building.
Put AI to work, safely
Help your team think better with AI
Talk to IT.ie about rolling out Microsoft Copilot with the guidance, guardrails and training that keep judgement where it belongs.
Frequently asked questions
What does it mean to use AI as a thought partner?
It means using AI to think with you rather than for you. Instead of accepting the first answer, you ask the AI to question your assumptions, identify counterarguments and expose gaps in your reasoning before it helps you produce a final output. Responsibility for the judgement stays with you.
What is the CRIT framework for AI prompts?
CRIT is a framework from Geoff Woods, author of The AI-Driven Leader. It stands for Context, Role, Interview and Task: give the AI the real situation, tell it what perspective to adopt, ask it to interview you before answering, and only then define the task you want it to complete.
How do I get an AI tool to challenge my thinking?
Give it explicit permission to disagree. Ask it to identify the strongest counterargument to your position, the assumption your reasoning depends on and the evidence that would prove you wrong. Tell it not to manufacture disagreement, and require it to support claims with authoritative sources and to say plainly when it cannot verify something.
Why should an AI study partner and assessor be kept separate?
If the same interaction is simultaneously coaching and grading you, the feedback may be influenced by the path the conversation took. Working through the idea in study-partner mode, then switching deliberately to assessor mode against a rubric, gives a cleaner test of whether your understanding holds up.
Should I use AI for every question?
No. If you already know the answer, trust your own expertise and judgement. Use AI to pressure-test the reasoning, sharpen the explanation or find a blind spot, rather than automatically outsourcing thinking you are already capable of doing.
Related reading: If You Wouldn’t Post It on Social Media, Don’t Paste It Into AI.
Sources
- The CRIT Framework: Building an Unfair AI Advantage (The Tilt)
- Becoming an AI-Driven Leader: The Strategic Framework (Growth Faculty)
- New Frontiers: The Origins and Content of New Work, 1940–2018 (NBER)
- MIT News coverage of David Autor’s research
- World Economic Forum: The Future of Jobs (2016)
- World Economic Forum: The Future of Jobs Report 2018
Written by John Grennan with an AI thought partner used as scaffolding, as described above. Statistics are attributed to their original sources; the 65 per cent figure is a popular estimate quoted by the World Economic Forum, not a WEF finding.
