AI Should Help Us Think, Not Think for Us

Every few months the conversation about AI seems to swing between two extremes. Either it is going to replace every knowledge worker, or it is a glorified autocomplete that does very little of value.

My experience, and I would argue the real value of AI, sits somewhere in the middle.

The most useful applications of AI in my work have not removed humans from the process. They have reduced the friction around work humans were already responsible for doing.

Triaging

One of the most practically useful AI tools I’ve built is a simple email triage agent.

The prompt is not especially sophisticated. It tells the agent who I work with, which stakeholders need particular attention, and what kinds of requests usually require action from me. It then looks for deadlines, explicit asks, and anything that may be blocking someone else’s progress.

On a typical day I receive dozens of emails. Some are genuinely urgent. Some require thoughtful responses. Others are newsletters, market updates, or industry commentary that I would like to read but which can wait until later.

The agent helps me separate those categories.

It’s not making decisions for me. It’s helping me navigate a large amount of information and focus my attention where it is most needed. The result is that important emails are addressed more quickly, colleagues are less likely to be blocked waiting for a response, and the less urgent material remains available for when I have the time and headspace to engage with it properly.

Responding

Another surprisingly useful application is something software engineers have long referred to as “rubber duck debugging”.

The idea is simple. Before asking another person for help, explain the problem out loud to a rubber duck. Oftentimes, the act of talking through the issue reveals the solution, one you already knew.

In my case, I literally have a yellow rubber duck replacing the escape key on my keyboard.

The challenge, of course, is that the duck never talks back. AI does.

When I am trying to think through a problem, formulate an argument, or approach a difficult decision, AI can act as an interactive sounding board. Its value is not usually in the answers it gives, but in the questions it asks: identifying assumptions, exposing gaps in reasoning, and forcing me to explain my thinking more clearly. I do have to be explicit that I am looking for challenge rather than praise, because AI often defaults to being agreeable.

That does not make it a substitute for discussing important decisions with colleagues. AI lacks the context, experience, organisational knowledge, and the nuanced understanding that people have. What it can do is help me arrive at those conversations with my thoughts already organised.

Reviewing

There is a particular challenge that comes with creating anything for an audience. Whether it is an email, a presentation, a strategy document, or a proposal, it is remarkably easy to become so immersed in the details that you lose sight of how the finished product will be received.

AI can provide a useful external perspective.

I upload documents (within the parameters of my organisation’s AI use policy!) and explain who the intended audience is. AI then evaluates whether the document is aligned with those expectations.

This works only because a human remains responsible for the judgement. AI does not know my stakeholders. It does not understand the politics, history, or relationships involved. It does help me zoom out to see the woods, not just the trees.

Prototyping

AI has cut the time it takes to turn an idea into something visual from days to hours.

As a communication tool, this is enormously powerful. A prototype can make an idea tangible. It can generate discussion, uncover misunderstanding, and create alignment more effectively than a written description alone.

This approach is not without risk.

People become attached to what they can see. Once stakeholders have interacted with a prototype, they may view it as the product rather than an illustration of the product. Expectations become anchored to a version that was never intended to be production-ready.

There is another risk as well. AI-generated prototypes often hide the complexity that sits beneath real software. Security, authentication, scalability, reliability, accessibility, governance, and integration rarely appear in the prototype. Those questions are still there; they are simply waiting for somebody to solve them.

Summarising

Perhaps the most obvious use case for AI is summarisation.

When AI has access to documents, notes, and knowledge repositories, it can rapidly bring together information that would otherwise take more time to gather manually.

This can be enormously helpful. It can also be wrong thanks to AI’s tendency to hallucinate.

For that reason, I almost always ask AI to cite its sources. If it makes an important claim, I want to know which document it came from and where I can verify it. The ability to trace information back to its origin is often more valuable than the summary itself.

The Human Still Owns the Outcome

Looking across these examples, the pattern is consistent. AI is most useful when it makes human involvement more effective, not when it pretends human involvement is unnecessary.

Used well, it can remove friction, accelerate understanding, and help us work with information more effectively. Used poorly, it can encourage us to outsource the thinking that gives our work value.

The goal is not to hand responsibility to AI.

The goal is to become better at exercising our own.