From students to CEO, in 2026 the ability to successfully prompt AI tools is essential.
CX Network's 2026 research into the state of CX asked practitioners to select the changes they believed generative and agentic AI will bring to their organization's CX function. As many as 36 percent said it would be the requirement for teams to upskill – the most selected response.
AI literacy requires many skills, but DataCamp's 2026 State of Data & AI Literacy Report – based on a survey of 517 business leaders in the UK and US – found 67 percent of leaders consider skills in prompt engineering and steering AI systems to be important or very important.
To help practitioners hone this vital skill, this article delves into the misconceptions around prompting AI, explains the factors that lead to AI outputs failing to meet expectations and shares tips for how to improve AI outputs.
The basics of prompting AI
If you ask an LLM how to prompt it for the best results, they will all tell you to be explicit and specific. LLMs are answer engines, designed to deliver an output. The the more the model is forced to infer meaning, the greater the opportunity for expectations to misalign.
This means that for the best results, users should provide examples, tell the LLM the role it should take, include context, explain how the output will be used and specify the length and tone.
Harvard University summarizes the top actions as follows:
Persona: Who the AI acts as
Task: What it does
Context: Background details
Format: How it should present the answer
Users may also receive better results if they use positive language over negative and structure prompts clearly. Breaking complex tasks into individual steps also helps, and this is called prompt chaining. It isn't about giving the LLM five individually specified tasks in one prompt, but giving each subsequent instruction in response to the output.
Applied to a real-world scenario, this transforms a prompt from "write me a report", to
"Write a 500-word report explaining the benefits of renewable energy for small businesses. Use simple language and include three real-world examples."
If you're prompt chaining a more complex task, the sequence may progress as:
- Create a slide deck for a strategy meeting using the information provided
- Add visual references that enhance the layout on slides 3 and 7
- Create conclusion slides for each section
- Create a final conclusion for the entire deck
Some additional tips from ChatGPT and Claude include:
Encourage step-by-step reasoning: For anything involving logic, math, or multi-part analysis, ask the model to "think" through the task step by step before giving a final answer. This can dramatically improve accuracy on complex tasks.
Structure your prompt clearly: Use delimiters (XML tags, markdown headers, triple quotes) to separate instructions, context, and examples. Something like ... and ... helps the model parse what's what, especially in longer prompts.
Treat your first prompt as a draft: If the output isn't right, look at why. Is it due to ambiguous instruction? missing context? wrong format specified? Once it's clear, adjust. Prompting is much more empirical than it looks.
Include constraints: Tell the AI what to avoid or include, for example: Maximum 300 words, use British English, avoid jargon, include references, or use a professional tone
From vocabulary to visuals
AI prompts don't have to be communicated in words. Olivia Lucas, who currently holds the position of senior service designer at EA, prompts with vocabulary and visuals when using LLMs.
She recently used Claude to create a journey vision instead of turning to the traditional CX tools. In part, this was because she had vast amounts of data, which needed to be transformed into a sharable visual that could be iterated on by multiple stakeholders. Another reason was because she wanted to collaborate with the tool to refine her ideas and "give the LLM space to create something for us".
On the secrets of a successful prompt, Lucas tells CX Network there are two hacks that are just as important as the instructions: adjusting the settings applied to each tool and knowing which tasks different LLMs are more likely to excel or fail at.
"Prompting is like another form of storytelling. The question is how do I tell the LLM enough to get the task done and give it the right context to do that task," she says.
"I found what worked decently well is approaching it in a storyboard format and telling the LLM these are the different pieces of the story and this is how I see them fitting together. That helped a lot. The clearer I am with what I want in the beginning, the easier it is to get the output," she says.
As an experience design specialist, sometimes Lucas uses images to enhance her prompts, sharing diagrams she has drawn or visuals she has found online. "Any inspiration you can share will improve the output," she explains.
One of the biggest considerations for Lucas is the number of prompts required to reach a workable output. Lucas is mindful about the environmental impact of AI and ensures she uses it responsibly. The more adept she becomes in prompting the tools, the more she can minimize her own personal impact.
"I'm aware of how much energy AI needs and I don't have a good solution for that, but I avoid using LLMs for random queries and when I am using them for work I plan beforehand, so it can reduce that overall impact," she explains.
Why do some AI outputs fail to meet expectations?
You could be forgiven for thinking better prompts will deliver better outcomes. But according to Bret Alexander, partner, digital product and service innovation for EY Studio+, that is a persistent misconception.
"While prompting certainly matters, most organizations eventually discover that the quality of an AI's output is determined far less by how a request is worded and far more by whether the AI understands the underlying purpose behind the request," he explains.
"Prompt engineering tells AI what to do. Context engineering gives it the why. And just as with people, understanding why transforms execution from functional to impactful."
This is applicable to all LLM use cases, but in the realm of CX we know great experiences emerge from a deep understanding of customers – their needs, motivations, and desired outcomes – and "AI is no different," Alexander says.
"As that context layer matures, AI becomes increasingly effective at identifying patterns, diagnosing root causes, prioritizing opportunities, and accelerating improvement initiatives across content, design, product, technology, and service teams," Alexander continues.
Simon Sinek's Golden Circle is a leadership framework that argues inspiring leaders and high-performing organizations communicate the why before the what, i.e. they capture attention with purpose, not instruction. Alexander says the framework provides "a useful lens for understanding this evolution".
"Once AI understands the underlying purpose, it becomes far more effective at determining the how and generating the what. The future of customer experience will belong to organizations that invest in building this context layer, creating a compounding advantage where every customer interaction, insight, and enhancement makes the next AI-generated recommendation smarter, more relevant, and more impactful," Alexander says.
He concludes: "Ultimately, organizations will not create differentiated customer experiences because they mastered prompting. They will create them because they built a deeper understanding of their customers and made that understanding available to both people and AI."
Better prompting is only part of the solution
This sentiment is echoed by former attorney, award-winning author and AI and cyber risk advisor, Dawn Kristy, JD. Kristy says there are two primary reasons AI may be failing to meet expectations:
1. The AI may be filling gaps the user did not recognize: She explains: "When information is missing, AI may infer what the user intended or supply details that appear to fit the request." This means a lack of clear purpose, audience, context, or desired outcome "can produce a plausible but disappointing response". She adds: "Some assumptions may be useful, while others may lead to inaccurate or misleading output."
2. The problem may extend beyond the prompt: Kristy says the AI may lack the necessary information, access to an authoritative source, or the capability to retrieve or act on the material the user references. "Better wording cannot always correct missing context or a mismatch between the tool's access and the result the user seeks," she clarifies.
Even with these points tackled, there are two further misconceptions that many people have about prompting.
The first is that there is one perfect prompt. In fact, Kristy says effective prompting is often iterative.
"The first prompt begins the exchange and need not contain every possible instruction. Prompts with multiple instructions stacked upon one another may be as unclear as ordinary writing that does the same," she explains.
The other misconception is that effective prompts are mainly about grammar or following a formula. "Clear writing helps, but good prompting also depends on clarity about the task, the relevant context, and what you are asking the AI to produce. Uncertainty in the task often reappears in the answer," Kristy adds.
Top tips for prompting: Pause, think, verify and revise
Sharing her top tips for how to prompt and AI tool, Kristy offers two pieces of advice:
1. Treat prompting as a dialogue: Kristy says that overloading the first prompt can eliminate the opportunity for dialogue and feedback. "Shorten the prompt, when necessary, observe what the AI returns, and revise accordingly," she says. "A useful answer may require several exchanges, depending on the task and the context available."
2. Pause, think, verify, and revise: Kristy explains: "When the output is unexpected, pause and consider what appears to be incomplete or incorrect. Verify important information, then revise the prompt or reconsider whether the AI tool is suitable for the task."
Here's a handy template on how to prompt an LLM, courtesy of ChatGPT:
How to prompt LLMs: Key takeaways
As these powerful tools become more deeply embedded in daily life, learning how to use them to their full potential is becoming a vital workplace skill.
As the interviewees in this article advise, clear language is paramount to a successful output and this can be enhanced through the user's formatting and telling the LLM about the role it is expected to take in the task.
It is possible to share vast reems of information with an LLM, either through necessity for the task at hand, or to engage in two-way idea generation. It is even possible to use visuals to support a prompt.
However, context is the key to a successful output. As Alexander says: "Once AI understands the underlying purpose, it becomes far more effective at determining the how and generating the what."
Quick links
- 8 Essential steps to take when preparing data for AI
- The EU AI Act compliance checklist for customer service teams
- Agentic AI needs guardrails before it needs more intelligence