10 Steps Leaders Can Take to Get the Most Out of Their AI Assistants

This article was also published on Forbes Japan: 10 Ways Leaders Should Unleash the True Potential of AI Assistants

Published on Forbes April 14, 2026 by Jason Barnard

Have you built an AI assistant?

If so, this might be the situation you’re in: You wrote the instructions, named it, set it up and are using it. However, you think your AI assistant is just … fine. Sometimes, you find it useful; other times, you find it frustrating and even underwhelming. ​

I recently ran a workshop for 25 entrepreneurs in which I explained my AI assistant implementation framework. ​I observed that, by and large, the challenges the audience members faced were less in the setup of AI assistants and more in their habits of using them.

I gave a hundred tips at the workshop, but there are 10 key ones that I believe business leaders should consider.

1. Put Your Most Important Instructions First

​I believe the most foundational step you can take is to put your most important instructions first. Your nonnegotiable rules should go first, not be buried in paragraph three.

Consider this: According to one study, in LLMs, ​”The primacy effect - where items presented first are more likely to be remembered or selected - plays a key role in Multiple Choice Question Answering (MCQA).”

​2. Be Mindful Of Which Instructions You Put Last

​You should also be mindful of which instructions you put last. I recommend putting your unbreakable rules not just at the beginning of your prompt but also at the end. It ties back to recency bias.

As a study explained, “Previous research has indicated that LLMs may exhibit serial position effects, such as primacy and recency biases.” The researchers noted that their “extensive testing across various tasks and models confirms the widespread occurrence of these effects, although their intensity varies​.” ​​

3. Tell Your AI Assistant What You’re Thinking And What You Thought Of Its Last Answer

AI assistants can’t read our minds. Before you ask your AI assistant for anything, I advise sharing with it what is already in your head, such as the angle you’re considering and the concerns you have about what you’re working on.

Then, after every response you receive, I recommend telling it what worked and what didn’t before asking for the next version. For instance, you could write, “This is close, but the tone is too formal for this client.” That way, your feedback can shape the next response.

4. Never Ask Your AI Assistant If It Agrees With You

In my experience, asking your AI assistant, “Do you agree?” often results in biased answers. ​As one study about LLMs found, “When presented with a user’s judgement, they quickly shifted their beliefs to stay in line with the user.”

Instead, my advice is to use a different prompt, such as: “Assess this. Do you have any objections, corrections or improvements? I’d be impressed if this leads to a discussion.” I’ve found that this final line can turn the AI’s agreeability instinct in your favor because disagreement becomes the impressive response.

5. Remember That Your AI Assistant Doesn’t Have The Context You Have

Let’s go back to the third tip. If you wrote, “This is close, but the tone is too formal for this client,” remember that your AI assistant lacks the context you have. It doesn’t know your past interactions with the client - or that the client doesn’t share your context.

For any output that will leave the conversation, I recommend being explicit: “The person receiving this has seen none of our discussion. Write as though they are coming to this completely cold.”

6. Pick One Term Per Concept, And Don’t Vary It

​When you’re interacting with your AI assistant, I recommend being exact with your wording.

For example, if your knowledge base uses the term “client” but then you use the word “customer” in your AI prompts, your AI assistant may not connect the dots between the two terms.

7. Start A Fresh Conversation When The Conversation Starts To Drift

In a conversation with an AI assistant, everything you say stays in its context bank for that conversation, including ideas you rejected and things you changed your mind about.

Research has revealed that “multi-hop reasoning tasks exhibit substantial performance degradation as context grows, whereas single-span extraction remains comparatively stable.” In other words, the researchers found that tasks that require an AI tool to weave together various pieces of information worsen the longer the context becomes.

Before ending a conversation, I recommend that you ask the AI for a handoff, saying something along the lines of, “Summarize everything we agreed on, and write the opening prompt for a new conversation continuing this work.” Then you can paste the response into a fresh thread.

8. Understand The Difference Between Drift And Hallucination

Drift and hallucination can look similar because they can both result in inaccurate information, but the causes and fixes are different. According to IBM, “Model drift refers to the degradation of machine learning model performance due to changes in data or in the relationships between input and output variables.” However, a research article in the journal Nature noted that “in essence, ‘AI hallucination’ refers to the phenomenon where artificial intelligence generates distorted information.”

​The solutions for each issue, in my experience, are different. For instance, to address drift, one thing you can do is start a fresh conversation with a clean summary of what’s been agreed on so far. To address hallucinations, a step you can take is to ground the AI’s response in reliable sources or explicitly fact-check its outputs.​

9. Build A Corrections File ​

When your AI assistant gets something wrong, I recommend correcting it in a two-column table in a separate “corrections” file. ​​

That way, the next time you start a new conversation or use a different AI instance, you can copy and paste the content of that file to get things on track. In one column, write “Common Wrong Answer.” In the other, write “Correct Answer.”

10. Avoid Giving Your AI Assistants Word Documents And PowerPoint Decks

A .docx file is a ZIP archive of XML. Before reading a single word of your content, the AI unpacks the container, strips the formatting and extracts the text. Every step burns processing attention that could go straight to your content, not your packaging.

You can copy and paste your text instead. Or, you can use markdown, which is plain text with minimal punctuation signals that declare structure.