When to Stop Prompting: When AI Problems Require Better Systems

When an AI system produces a disappointing answer, the first instinct is often to rewrite the prompt. Add more context. Give it an example. Spell out the rules. Try again. Sometimes that works—but there’s a point where prompt refinement becomes a way of avoiding the real problem. If an AI system still struggles after repeated rounds of instruction, leaders need to ask a different question: Is the prompt actually the bottleneck?

That question matters because AI performance depends on far more than the words sent to a model. The quality and availability of data, the tools a system can access, the workflow surrounding it and the model’s own capabilities can all shape the result.

For leaders, the challenge is knowing when to stop tweaking and start redesigning. If the problem is the way work gets done, the answer may be a new workflow. If the system lacks reliable information, better data may matter more. If the task requires actions or specialized capabilities, different tooling or a different model may be necessary. And sometimes the right answer is to rethink the product or process altogether. 

Members of the Senior Executive AI Think Tank—a curated group of executives and practitioners with expertise across machine learning, generative AI and enterprise AI applications—share how they recognize those inflection points and what leaders can do when a better prompt is no longer enough.

“Prompting stops being the answer the moment the failure is no longer about how you’re asking, but about what the model can access, do or be trusted with.”

– Vandana Singh, Sr. Technical Program Manager Lead at Amazon

SHARE IT

Use Prompting for the Right Problem

Vandana Singh, Sr. Technical Program Manager Lead at Amazon, says leaders should first determine whether the problem is actually one that prompting can solve.

“Prompting is the right lever only when the problem is generation, like drafting, summarizing or reformatting,” Singh says. “It stops being the answer the moment the failure is no longer about how you’re asking, but about what the model can access, do or be trusted with.”

That becomes especially important when an AI system doesn’t have the information or capabilities required to complete the task. In those situations, adding more instructions can add complexity without addressing the underlying limitation.

“If the model lacks the right context, the fix is better context, not more, since noise drives hallucination,” she says. “If it can’t act, the fix is tooling to call the right APIs. If the output can’t be trusted, add a verification step.”

The same principle applies when the problem originates in the data or the model itself.

“If it’s wrong because the underlying data is poor, fix the data. If it’s hitting a capability ceiling, switch models,” Singh says. “And if you’re prompting around an awkward fit entirely, the real answer is a different product or process.”

For leaders trying to decide when to stop iterating, Singh offers a straightforward signal: “Stop tuning the prompt when the bottleneck is access, action, data, capability or trust.”

Reduce the Model’s Reasoning Burden

Tal Lev-Ami, Co-Founder and CTO of Cloudinary, takes the argument a step further, suggesting that repeated model reasoning can itself be evidence that too much complexity has been left for the AI to handle.

“If your model is doing a lot of reasoning, it’s a sign your system isn’t designed well enough yet,” Lev-Ami says.

He points to several possible causes behind inconsistent results, including the condition of the data and how tightly the workflow constrains the task.

“Inconsistent outputs often signal unstructured or ungoverned data, a workflow that isn’t constraining enough, or a task requiring more determinism than a model can be expected to reliably deliver,” he says.

Rather than expecting the model to resolve every ambiguity, Lev-Ami says organizations should put more structure around the work before the prompt is sent.

“The best AI systems are those where the model has the least reasoning to do, because structure, context and constraints have done most of the work before the prompt is sent,” he says.

“Prompt engineering is a surface fix. The deeper fix is almost always upstream”

– Yogesh Malik, CEO of Way2Direct B.V.

SHARE IT

Fix the Problem Upstream

Yogesh Malik, CEO of Way2Direct B.V., says increasingly complicated prompts can be a sign that an organization is trying to solve the wrong problem.

“A better prompt stops being the answer when the underlying problem is data quality, workflow design or a mismatch between the task and the model’s actual capabilities,” Malik says.

For leaders, one warning sign is the need to keep adding instructions simply to compensate for unreliable results.

“If you find yourself writing increasingly complex prompts to compensate for inconsistent outputs, you are treating a system problem as an instruction problem,” he says.

That makes diagnosis more important than another round of prompt refinement.

“Leaders should ask whether the AI is failing because it does not understand the request, or because it does not have access to the right information, or because the task requires a structured process rather than a conversational interface,” Malik says.

His conclusion is direct: “Prompt engineering is a surface fix. The deeper fix is almost always upstream.”

Start With the Work, Not the Prompt

Daria Rudnik, Team Architect and Executive Leadership Coach at Daria Rudnik Coaching & Consulting, says leaders should begin even earlier—before deciding how AI should be used.

“I see prompting as the final step,” Rudnik says. “Before trying to fix the prompt, leaders need to understand the work itself, starting with the most important question: What are we trying to achieve?”

That question shifts the focus from the quality of an AI response to the outcome the organization actually needs.

“The real question isn’t ‘Can we make AI produce a better answer?’ but ‘Is this the best way to achieve the outcome?’” she says. “And when it is, the questions are: Where can AI add value? What should remain with humans? What data is needed, and how should the workflow change?”

Only then does prompting enter the picture.

“Once those decisions are clear, it’s time for better prompting,” Rudnik says. “What I see a lot now is that teams are trying to use better prompts to compensate for a poorly designed process. The prompt should serve the workflow—not define it.”

Watch for Prompt Debt

Andre Shojaie, Founder of HumanLearn, identifies another sign that teams may be reaching for prompting too often: “prompt debt.”

“A better prompt stops being the answer when people spend more time teaching the AI how to navigate a broken process than fixing the process itself,” Shojaie says.

The warning signs can accumulate gradually as teams add instructions to handle more exceptions and employees develop their own workarounds.

“I would watch for ‘prompt debt’: instructions that keep getting longer, exceptions multiplying, and employees developing personal workarounds to get reliable results,” he says. “That is usually a signal to move up one level.”

That next level may involve examining the information, automation or technology surrounding the model.

“Is the right data available? Should a step be automated rather than prompted? Does the task need a specialized model or tool? Should the workflow itself change?” Shojaie says.

The larger point is that prompting shouldn’t become a substitute for system design.

“Leaders should treat prompting as one layer of system design, not the system,” he says. “If every improvement requires another paragraph in the prompt, the problem may no longer be communication with the AI. It may be the architecture around it.”

Build for Reliability, Not Just Responses

Will Conaway, President of Tuxedo Cat Consulting, says the distinction becomes especially important in environments where AI outputs can affect real-world decisions.

“A better prompt stops being the answer when the issue is no longer unclear instructions, but an unreliable system around the model,” Conaway says. “In healthcare, that may appear as inconsistent outputs, missing patient context, poor data quality, weak handoffs or decisions that require accountability beyond what an AI response can provide.”

Rather than automatically turning back to the prompt, leaders should identify which part of the system is creating the problem.

“Leaders should ask whether the problem is caused by communication, workflow friction, data gaps, model limitations or a process that should be redesigned entirely,” Conaway says.

That diagnosis can point toward solutions that have little to do with prompt wording.

“If clinicians must repeatedly correct the same errors, the solution may be better integration, validation rules, structured data, training or a different tool; not another prompt,” he says.

For Conaway, prompting is ultimately one component of a larger operating model.

“AI-native companies should treat prompting as one layer of system design, not the whole strategy,” he says. “The goal is to build safe, measurable workflows where technology supports judgment, reduces burden and improves outcomes.”

Recognize Structural Failures

Dhyey Mavani, Generalist, Agentic AI (C-suite priorities) at Abridge, draws a particularly sharp distinction between problems with language and problems with the system itself.

“A better prompt stops being the answer when the failure is structural, not linguistic,” Mavani says. “If the model lacks the right data, no wording fixes that. If the task needs tools, memory, approvals or deterministic checks, redesign the workflow.”

There is also a point when continued prompt refinement stops producing meaningful gains.

“If quality plateaus across strong prompts, compare models,” Mavani says.

And when users have to change the way they work simply to accommodate the AI, he says leaders should question whether the product itself is the right fit.

“If users still need to contort their process around the AI, question the product itself,” he says.

His broader principle is simple: “Prompting should optimize a sound system, not compensate for a broken one.”

“Every new rule, exception and ‘never do this’ is a sign you’re asking instructions to fix something instructions can’t fix.”

– Dr. Aditya Vikram Kashyap, Vice President of Firmwide Innovation at Morgan Stanley

SHARE IT

Test the Model Against the Task

Dr. Aditya Vikram Kashyap, Vice President of Firmwide Innovation at Morgan Stanley, suggests a practical test for determining whether a prompt is carrying too much of the workload.

“A better prompt stops being the answer when your prompt starts reading like a policy manual,” Kashyap says. “Every new rule, exception and ‘never do this’ is a sign you’re asking instructions to fix something instructions can’t fix.”

He proposes putting the prompt itself to a simple test: Could someone reasonably complete the task with only the information provided?

“Hand the same task to a smart new hire with only the information in the prompt,” he says. “If they succeed, keep iterating. If they fail, the prompt isn’t the problem.”

The reason for the failure then becomes the next diagnostic question.

“If they lack the facts, fix your data. If the task is secretly five tasks, redesign the workflow,” Kashyap says. “If they need to look something up or take action in another system, add tooling. If they understand the job but can’t do it reliably, try a different model.”

And sometimes the AI output isn’t the real objective.

“If a perfect answer still wouldn’t move the business, the problem is the process or product itself,” Kashyap says.

The temptation to keep prompting is understandable because it is often the easiest fix to try. But Kashyap says leaders should aim for something broader: “The goal isn’t a perfect prompt. It’s a system where an average prompt still gets great results.”

Design Around the User

Goran Paun, Principal and Creative Director at ArtVersion, approaches the question from a user-experience perspective. He says organizations shouldn’t make employees responsible for absorbing the complexity of an AI system.

“The goal should not be to teach people how to become better prompt engineers,” Paun says. “It should be to design systems that require less prompting in the first place.”

That becomes key as users repeatedly explain, refine and correct their requests.

“From a UX perspective, we are already seeing fatigue from constantly having to explain, refine and correct what we want,” he says. “At some point, the system has to absorb the complexity instead of pushing it back on the user.”

That also means being selective about where AI belongs in the workflow.

“Teams also need to be more deliberate about which tool fits which task,” Paun says. “Not everything needs an AI workflow.”

In some cases, the best solution may not involve AI at all.

“Sometimes doing the work directly is faster, more accurate and better at preserving the original thought and meaning,” he says.

Orchestrate the Work When the Task Gets Complex

Uttam Kumar, Engineering Manager at American Eagle Outfitters, says prompting becomes less useful when the work itself requires multiple systems and actions.

“Prompting stops being the right tool when a task requires complex, multi-step actions across different platforms,” Kumar says.

For example, a retail employee may need to work across several operational systems to answer a customer or make a business decision.

“If a store associate or supply chain manager has to stitch together inventory checks, logistics data and customer history manually through chat boxes, prompting is failing them,” he says.

In those situations, Kumar says leaders should shift their attention from the conversation to the workflow.

“Leaders must pivot toward workflow design and specialized agentic tooling,” he says.

The objective isn’t simply to generate a better response but to remove unnecessary manual steps from the entire journey.

“By automating the journey through robust orchestration, companies remove manual friction, letting retail teams focus on high-value customer interactions instead of fighting with text boxes,” Kumar says.

Actionable Strategies for Better AI Systems

• Identify the real bottleneck before refining the prompt. Determine whether the issue involves access, action, data, model capability or trust before investing more time in prompt optimization.
• Move complexity out of the model when possible. Use structure, context and constraints to make the task clearer before it reaches the model.
• Fix upstream problems at their source. If inconsistent results stem from poor data, workflow design or a mismatch between the task and the model, address that underlying issue rather than adding more instructions.
• Start with the desired business outcome. Define what you’re trying to achieve, where AI adds value, what should remain with humans and how the workflow needs to change before optimizing the prompt.
• Watch for prompt debt. Increasingly long instructions, multiplying exceptions and employee workarounds are signals that the architecture around the AI may need attention.
• Design for reliability, not just useful responses. Build in appropriate integration, validation, structured data, training and human judgment when the consequences of errors require more than a better prompt.
• Distinguish linguistic problems from structural ones. If the system needs better data, tools, memory, approvals or deterministic checks, redesign the workflow instead of continuing to refine the wording.
• Use a simple test to diagnose the problem. Ask whether a capable new hire could complete the task using only the information contained in the prompt; if not, determine what information, workflow, tooling or model capability is missing.
• Design the experience so users don’t carry the complexity. AI systems should reduce the need for users to repeatedly explain, refine and correct their requests—and teams should consider whether AI is even the right tool for a particular task.
• Orchestrate complex work instead of forcing it through chat. When a task requires multiple systems, data sources and actions, workflow design and specialized tooling can remove manual friction more effectively than increasingly elaborate prompts.

Design the System, Not Just the Prompt

The Senior Executive AI Think Tank members point to a simple shift in mindset: When an AI system keeps falling short, the answer isn’t always to give it better instructions. Sometimes the better move is to step back and look at everything around the model—how the work is structured, what information is available, which tools are connected and whether AI is even the right fit for the job.

That may be where the bigger opportunity lies. As AI moves deeper into everyday work, leaders will have to spend less time asking, “How do we get a better answer?” and more time asking, “How should this work?” The prompt will still have a place, but the organizations that get the most from AI may be the ones that know when to stop tweaking the words and start redesigning the system.