You're Using AI to Save Time. What If You Used It to Do Things You Never Thought Possible?

Most people use AI to get through their to-do list faster. There's a much more interesting possibility: using it to take on work they would never have attempted in the first place.

Ask an AI assistant to summarize a document, draft an email, or fix a bug, and you can save a few minutes. Ask it to help you investigate a problem you don't understand, test an assumption you've held for years, or build something outside your area of expertise, and the nature of the work starts to change.

The difference isn't necessarily the model. It's the ambition of the problem you're willing to attempt.

Here's a practical way to make that shift.

The hidden limit isn't always what AI can do. It's what you think to ask.

Imagine a small software company losing customers.

The founder opens an AI assistant and asks:

"Write an email to customers who haven't logged in recently."

The model produces a perfectly reasonable email. The founder sends it. A few customers return. Work completed.

Now imagine a different question:

"Help me understand why customers stop using our product. Examine the available usage data, group customers by behavior, identify where engagement starts to fall, and propose experiments that could prevent cancellations. Separate what the evidence shows from what we're guessing."

This request is harder. It requires data, investigation, and verification. The answer might reveal that customers aren't leaving because the product is bad. Perhaps they never reach the feature that makes it valuable.

The founder who asked for an email gets an email. The founder who asked about the underlying problem might discover a way to change the business.

That distinction matters.

AI can make individual tasks cheaper and faster. But some of its most interesting uses come from changing which problems we attempt, how deeply we investigate them, and how many possible solutions we can explore.

The first question is no longer just, "Can AI do this task?"

It's, "What would I try if the cost of attempting it were dramatically lower?"

1. Start with the outcome, not the task

Most requests to AI begin halfway through the thinking process.

We have already decided what needs to be produced, so we ask the model to produce it.

Write the report. Build the dashboard. Create the presentation. Fix the error. Draft the proposal.

Sometimes that's exactly right. But sometimes the requested deliverable is merely a habit inherited from the way work used to be done.

Consider a weekly sales report.

You could ask AI to format the latest numbers into a presentation. That's useful. But why does the report exist?

Perhaps the actual purpose is to identify deals that are at risk, understand why the pipeline is slowing down, and decide where the sales team should spend its time.

Once you understand that purpose, an entirely different set of possibilities opens up.

You could ask AI to compare current performance with historical patterns, flag unusual changes, investigate likely causes, and prepare a short list of decisions for the team. You might even discover that a weekly presentation isn't the best way to accomplish the goal at all.

The report was the task. Better decisions were the outcome.

Try this

Before your next AI request, complete these two sentences:

• What I'm asking AI to produce is...
• The reason I need that output is...

Then ask whether the output is actually the best way to achieve the reason.

A useful prompt:

"I'm about to ask you to [task]. Before doing it, help me clarify the outcome I'm trying to achieve. Identify better ways to reach that outcome, including approaches that don't involve producing the deliverable I've requested. Explain the trade-offs, then recommend a path."

This doesn't mean every email needs a strategy session. Use it when the stakes are meaningful, the current process is expensive, or you've been repeating the same task without questioning why.

2. Ask for work you couldn't comfortably do yourself

Here's an uncomfortable thought: we often use AI within the boundaries of our existing abilities.

A developer asks for code because they know how to code. A marketer asks for copy because they understand marketing. An analyst asks for a spreadsheet because spreadsheets are familiar.

These are sensible starting points. But they can also create a ceiling.

Imagine a designer who wants to understand why visitors abandon a website.

They might know how to design a better landing page. They might not know how to analyze event data, construct a statistical comparison, inspect browser behavior, or design a credible experiment.

Previously, they would need to find the right specialists, convince them to help, and coordinate the work. Now, an AI assistant may help with much of that process, provided it has access to the relevant tools and data.

The designer could ask it to:

1. Identify what evidence is needed to understand the problem.
2. Explain which analytical methods are appropriate and why.
3. Help write queries to extract the relevant data.
4. Compare competing explanations for the behavior.
5. Design an experiment that could distinguish between those explanations.
6. Define what results would count as evidence that the experiment worked.

The designer doesn't magically become an expert statistician. The model can make unfamiliar methods more accessible, but its recommendations still need scrutiny.

That's the important shift: you no longer have to master an entire discipline before making a serious attempt at a problem that depends on it.

You need enough understanding to ask sensible questions, inspect the work, and recognize when outside expertise is necessary.

Try this

Choose a problem that has been sitting on your list because it requires a skill you don't have.

Ask:

"Help me tackle this problem even though I'm not an expert in [field]. Explain what a competent professional would do, what information they'd need, which methods they'd use, and how we'd check the result. Don't skip steps that require my judgment or specialist review."

Then actually attempt the problem.

The point isn't to pretend expertise doesn't matter. It's to stop treating a lack of expertise as an automatic reason not to begin.

3. Go deeper instead of merely asking for more

When an AI response disappoints us, we often ask for a longer version.

More detail. More examples. More ideas. More sophisticated language.

But a longer answer isn't necessarily a better answer.

Suppose a company wants to improve its website's conversion rate. It asks an AI model for recommendations and receives the usual suggestions: improve the headline, simplify the form, add testimonials, and make the call to action more prominent.

Nothing is necessarily wrong with that advice. It's just advice the company could have found almost anywhere.

A deeper investigation would start with different questions.

Which visitors are leaving? What do they see before they leave? Does the problem occur on particular devices, traffic sources, or pages? Is there evidence that users misunderstand the offer? Are visitors encountering an actual technical problem?

Those questions require a different level of work.

Instead of producing recommendations from general knowledge, the model would need to investigate the company's own evidence, test hypotheses, and distinguish observations from speculation.

The same principle applies to programming, scientific research, operations, and marketing.

A shallow request asks for an answer.

A deeper request asks for a defensible answer, grounded in evidence, with a way to check whether it's right.

Try this

After receiving an answer you consider mediocre, don't immediately ask for ten more ideas.

Ask:

"Give me a more rigorous version of this work, not a longer one. Identify the assumptions behind your current answer, the evidence that would challenge those assumptions, the best methods for investigating the problem, and the independent checks needed to validate the result. Then revise your answer using that standard."

Notice the emphasis on independent checks.

AI systems can make mistakes with impressive confidence. Asking a model to verify its own work can help catch some errors, but it isn't sufficient proof of correctness.

Whenever possible, use an external test: run the code, query the database, inspect the original source, compare the result against known facts, or ask a qualified person to review the work.

The harder the problem, the more important this becomes.

4. Bring in disciplines you wouldn't normally consider

Some problems remain difficult because we're looking at them through a single professional lens.

A marketer sees a campaign. An engineer sees a system. A founder sees a business opportunity. A researcher sees a question to investigate.

Each perspective reveals something and misses something.

AI can help you explore multiple perspectives without having to arrange a meeting with a dozen specialists.

Consider a company preparing to launch a new product.

A conventional launch plan might cover positioning, pricing, advertising, and distribution.

A broader investigation could examine the launch through several different disciplines:

• Behavioral science: What would prevent someone from changing their existing habits?
• Economics: What alternatives does the customer give up by purchasing this product?
• Reliability engineering: What could fail under real-world conditions?
• Security: How could someone misuse the product or exploit its weaknesses?
• Journalism: Which claims would a skeptical reporter investigate?
• Operations research: Where are the bottlenecks if demand grows much faster than expected?

These aren't six personalities for an AI to imitate. They're six different sets of questions and methods.

That distinction matters. Asking a model to "act like a world-class expert" can change the language of its response without changing the quality of its reasoning. Asking it to apply a particular method, inspect specific evidence, or perform a concrete analysis gives you something more useful to evaluate.

Try this

Describe the outcome you want and ask:

"Identify five disciplines that could contribute useful methods to this problem. For each one, explain what it would investigate, which technique it would contribute, and what new evidence it might reveal. Select the two perspectives most likely to change the final result and apply them."

You don't need to use every perspective. The goal is to find the ones that expose something your usual approach would miss.

5. Turn one-off requests into systems that improve

There is a final step that often gets overlooked.

Once AI helps you solve a problem, ask whether you should have to solve it manually again.

Suppose you spend an afternoon analyzing customer feedback. You identify recurring complaints, group them into themes, and produce a useful report.

Next month, you'll probably do the same work again.

But what if the first analysis could become a repeatable process?

You could create a workflow that collects new feedback, categorizes recurring issues, flags unusual changes, links each finding to its source, and prepares a draft report for review.

The first attempt produces an answer. The second creates a process. The third might improve the process based on what you learned.

This is where the value can compound.

Instead of saving twenty minutes every time you perform a task, you may be able to redesign the task so that less manual work is required in the first place.

Of course, automation introduces its own risks. A workflow can repeatedly produce the wrong answer just as efficiently as the right one. It needs clear boundaries, error handling, appropriate permissions, and checks for the decisions that matter.

Try this

Whenever you complete a task that you're likely to repeat, ask:

"Which parts of this workflow are repeatable? Design a system that could perform those steps using the tools and data available. Identify the permissions required, likely failure modes, quality checks, and points where a human must approve the result. Start with a small version that we can test before relying on it."

Don't automate a process merely because you can. First make sure it's worth doing, then make sure it works.

A practical framework for your next AI session

These five approaches can sound abstract until you use them on a real problem.

Here's a simple sequence you can save and reuse.

Step 1: Name the outcome

What do you actually want to change, achieve, understand, or decide?

Avoid describing the deliverable too early.

Step 2: Find the missing capability

What expertise, information, tool, or access would make the problem easier to solve?

Be specific. "I need better analysis" is less useful than "I need to determine whether customers who skip onboarding are more likely to cancel."

Step 3: Choose a method

Ask the model to propose a suitable approach and explain why it fits the problem.

For a data question, that might mean a statistical test. For a software problem, it might mean reproducing a bug and writing a regression test. For a business decision, it might mean comparing hypotheses against customer evidence.

Step 4: Build in verification

Decide in advance how you'll know whether the result is correct.

A persuasive explanation isn't the same thing as a verified result. Use independent evidence wherever possible.

Step 5: Capture what you learned

If the process worked, document it. If it failed, identify why.

The next attempt should start from what you learned rather than from a blank conversation.

You can combine all five steps into one prompt:

I want to achieve [specific outcome].

My current approach is [what I normally do], and my constraints are [time, budget, tools, access, or other limitations].

First, identify assumptions in my approach and alternative ways to reach the outcome. Then determine which information, methods, and expertise are needed. Propose a plan, explain the trade-offs, and begin with the highest-value step you can perform using the available tools.

Distinguish facts from hypotheses. Don't invent missing information. Show me how we can independently verify the result, and identify any steps that require human judgment or approval.

Once we've finished, explain which parts of the process could be reused or automated safely.

The prompt isn't magic. Its value comes from changing the structure of the work: from producing an immediate answer to investigating a goal, taking action, and checking the outcome.

The biggest obstacle might be your imagination

There's a reason people keep returning to familiar uses of new technology.

We know how to evaluate them. We understand the workflow. We can predict the result.

Asking an AI assistant to rewrite an email is comfortable. Asking it to help investigate a new business opportunity, build a tool you've never built before, or challenge a belief you've relied on for years is less comfortable.

You might discover that the idea was bad. The model might fail. You might learn that the problem requires expertise or resources you don't have.

That's normal.

The useful change isn't believing AI can do anything. It can't. The change is becoming more willing to investigate what might be possible before deciding that it isn't.

Start with a problem that matters to you. Give the model the evidence and tools it genuinely needs. Ask for methods rather than impressive-sounding answers. Test the result. Learn from the attempt.

Then try something a little harder.

The most interesting question about AI isn't how many tasks it can remove from your day.

It's what you might finally be able to attempt with the time, knowledge, and capabilities it puts within reach.

Further reading: How to Be More Ambitious: A Tutorial for Humans with Models, by Felix Rieseberg. This article develops a separate framework and examples around the broader question of how people can approach work differently with AI.