Almost every company is using AI now, but far fewer are getting much out of it. McKinsey's 2025 State of AI survey found that 88% of organizations use AI in at least one business function, up from 78% the year before, and yet most of them still can't point to a measurable effect on profit.
A lot of this gap comes down to how you ask an AI assistant questions. The interface to these tools is ordinary language, which feels like it should make everything simple, right up until you watch someone type "write me a marketing plan" and get three paragraphs of beige nonsense in return.
Andrej Karpathy, who helped found OpenAI, summed up the shift in 2023: "The hottest new programming language is English." The catch is that you still have to be good at it. You get what you describe, so learning to describe well is most of the job.
This guide is part of our Using AI for Business series, and today, we’ll walk through how to ask AI assistants better questions for the work you actually do, with prompts you can copy and adapt as you go.
Why Knowing How to Ask AI Questions is Important
LLMs generate a response by predicting what should plausibly follow your input, and that's really what's happening under the hood. So if your input is vague, the most plausible continuation will be something generic.
The most useful way to think about this comes from Wharton professor Ethan Mollick, who has spent the last few years studying how people actually work with these systems. His advice: "the best way to work with it is to treat it like a person." You wouldn't drop a one-line task on a new hire with no context and expect them to nail it, and the AI assistant is in the same position.
Luckily, none of this requires special syntax or secret keywords. It's mostly a matter of telling the tool what any competent colleague would want to know before they started.
Nine Steps to Ask an AI Assistant Better Questions
Quick note: all nine of these won’t be necessary on every task, so think of them as a menu you order from. Each one below comes with a prompt you can lift straight into a chat and adjust.
1. Give the AI assistant context about the task
Before you ask for the thing itself, spend a sentence or two on the situation you're in and who the result is for. That turns a request that could mean a hundred different things into the specific one you need.
The reply won't be flawless, but it'll be close enough to edit instead of rewrite, which is the whole point.
2. Tell the AI the format and length you want
Left to its own instincts, LLMs like ChatGPT and Claude tend to over-produce, often handing you a wall of prose where five bullets would have done. So, be specific about both structure and word count:
Once you've described the container, the tool puts its effort into the content rather than guessing at how you wanted it packaged. This also saves tokens used for reasoning and redirects it to the actual task.
3. Show the AI assistant an example of what you want
Models are pattern-matchers by nature, and they copy a style remarkably well once they can actually see one. Describing the tone you want in words is fine, but pasting in a real example you like does the job far better.
One strong example will be much more effective than a whole paragraph of adjectives about the “mood” or “vibe” you're going for. If you can give it two or three samples rather than one, better still, because the tool starts inferring the pattern they share instead of copying a single sample too literally.
Just make sure the examples are genuinely representative, since the AI assistant will reproduce whatever quirks happen to be in them.
4. Give it a role and a job to do
Telling the tool what perspective to write from changes the entire character of the response. A flat "what do you think of this budget" gets you agreeable, forgettable notes, while asking it to read the same budget as a skeptical CFO produces something sharper and genuinely useful. For example:
5. Hand it the real material instead of a description of it
If you want the AI assistant to improve something, give it the actual content. Don't say paraphrase the report or upload the file. The instant you describe your material instead of providing it, the tool starts working on an imagined version of your problem, which is both inaccurate and inefficient.
Most tools let you upload the file directly, which is easier than pasting for anything long. A word of caution: strip out anything sensitive, like customer details or figures under NDA, before it goes anywhere near a chat window.
6. Break a big request into steps
When you ask for two things in one breath, like "analyze this feedback and write me a plan," the tool rushes both, and the plan suffers because it's built on an analysis it did in a hurry and never showed you. Instead, get the first part, check it, then build the next part on the version you approved. Start with the analysis on its own:
Once the themes look right, you move to the second stage:
Each stage is smaller and easier to check before an error compounds into the next one.
7. Give the AI assistant a goal, then let it ask you questions
Most of us under-specify without realizing it, which is why one of the best moves is to hand over the objective and then ask the tool to interview you before it writes anything. It surfaces the gaps you didn't think to fill, and you end up with a far better brief than you'd have written cold:
This flips the usual script. Instead of you guessing at what context the tool needs, the tool tells you what's missing.
8. Ask the AI assistant to show its work
For anything involving numbers, logic, or a chain of reasoning, tell the tool to lay out its thinking before it hands you a conclusion. Two useful things happen at once. You get to check the steps rather than trusting a black-box answer, and the act of reasoning out loud tends to make the answer itself more reliable:
When the logic is visible, a wrong turn is easy to catch, which is better than discovering it later inside a confident-sounding paragraph after it's already caused a problem. This isn't only for math, either. The same move helps with any answer that rests on a chain of reasoning, like a recommendation or a plan with dependencies, where you'd rather see the logic before you buy the conclusion.
9. Ask for a few options, not a single answer
Whenever there's more than one reasonable way to go, like a subject line, a headline, or a positioning angle, ask for several distinct versions instead of letting the tool commit to one on your behalf. You get to compare and choose, and seeing these options often clarifies what you actually wanted in the first place.
The trick is to push for real variety, since a model left to itself might hand you six near-identical lines:
Common Mistakes When Asking AI Assistants Questions
The classic one is the bare question, the "write a sales email" with no product or audience attached, which can only ever produce filler. Nearly as common is stuffing four separate requests into a single message and then getting a shallow treatment, when the fix is simply to send them one at a time.
Leading questions are trickier, because they feel perfectly reasonable. If a user asks why a particular strategy is the best choice, the tool will dutifully build a case for it, and you'll probably come away more confident than the evidence actually supports. Ask instead for the strongest arguments on both sides for a more objective analysis.
Then there's the habit of treating the first answer as the finished one, which it almost never is. The draft that comes back after you've told the tool what's wrong with the first attempt is usually the one you'll keep. Another common mistake is trusting facts and figures you never checked, which leads us straight to the next point.
AI is Better at Some Questions Than Others
AI isn't uniformly capable. It's very good at some tasks and quietly hopeless at others, and sometimes, the two can look nearly identical. The researchers who ran one of the largest real-world studies on this gave the pattern a name: the jagged frontier.
A Harvard and BCG field experiment involving 758 consultants found that on tasks well suited to the model, the people using GPT-4 produced work rated over 40% higher in quality and finished it around 25% faster. On tasks outside the model's range, the effect reversed, and AI users became more likely to arrive at the wrong answer. The ones who came off worst were those who took the output at face value and stopped checking it.
The key is to ask yourself what kind of question you’re asking the AI. Drafting, summarizing, restructuring, brainstorming, and explaining all play to the model's strengths. But, precise multi-step calculation and anything that hinges on current facts it can't look up is less reliable.
Using AI That Already Knows Your Context
When you start asking AI assistants questions, you’ll notice that most of the effort goes into supplying context -- who the work is for, and the raw material itself. AI tools built into your workflow can remove this friction. Gmelius lives inside Gmail, and its AI, Meli, reads the email thread you're already sitting in. There's no background to paste, because the background is the conversation on your screen.
Ask Meli to draft a reply, and it works from the real exchange, the customer's last few messages, and any custom knowledge base. Meli takes care of the asking-well part for email, because it starts with everything you'd otherwise have had to explain. Gmelius agents even use email architecture for federation and connect multiple apps to execute tasks automatically, without asking the AI assistant questions.
If you're following the series, our guides on using AI assistants for work and how to use AI for productivity explore these ideas further.
Want AI that already knows the context you'd otherwise spend time explaining?
Start your free Gmelius trial and let Meli draft straight from the thread you're in.
Frequently asked questions
Do I need to learn prompt engineering to ask AI assistants questions?
No, you don't need to learn prompt engineering to ask AI questions. The phrase prompt engineering makes it sound like a specialized discipline, and for the vast majority of business tasks it isn't one. If you brief the tool the way you'd brief a sharp colleague and correct it as it goes, you've already covered almost everything.
How long should my questions to an AI assistant be?
Your questions to an AI assistant should be long enough to carry important context, and no longer than that. If you're repeating yourself, cut the length; if the goal or the audience is missing, add it in.
How do I stop an AI assistant from making things up?
You can reduce hallucinations greatly, though you can't switch it off completely. Give the AI assistant the real source documents so it's reading rather than recalling, and ask for citations on anything you plan to rely on.
Does being polite to an AI assistant actually help?
Not for the reason people assume. Sprinkling in "please" and "thank you" won't move the quality of what you get back. But the impulse behind good manners, the habit of treating the exchange as a real back-and-forth with a capable collaborator, happens to be the exact mindset that produces well-formed questions.
.png)
.png)
.png)
.avif)
