Enterprise AI Has a Prompting Problem
Enterprise artificial intelligence has a trust problem, not because the outputs are obviously wrong, but because they're not.
Consider two marketing team members working from the same platform, using the same underlying data, responding to the same business brief. One asks the AI to validate a market opportunity and gets the green light. The other asks it to stress-test that same opportunity and gets a list of reasons to pause. Both answers sound coherent, credible, and strategically defensible.
So it might be surprising that those two answers should be different. After all, a question built to confirm and a question built to challenge are not the same question. And a good system will answer each on its own terms. The problem isn't that the answers diverge; it's that often nothing on the screen tells you why: which framing shaped each one, which criteria the system weighed, or whether either answer ever strayed beyond what the data can actually support.
That is the compounding cost of unstructured prompting, and it's quietly becoming one of the most consequential risks in enterprise AI adoption.
The Art of the Prompt Was Always the Wrong Frame
When large language models (LLMs) first arrived, the prevailing wisdom was that mastering AI meant mastering the prompt: Ask the right question in the right way and the right outputs will follow. Prompting became known as an individual, technical skill, something employees would learn, refine, and eventually master.
That framing made sense when AI was experimental and being used by individuals without the need for consistency across the organization. It stopped making sense the moment AI evolved into an operational, organizational skill.
Companies do not scale through unstructured improvisation. They scale through systems, repeatability, and shared frameworks. Yet many organizations are still approaching AI as though all employees should develop their own prompting style, their own workflows, and their own interpretation of what good looks like.
At the individual level, that variability feels manageable. At the enterprise level, it becomes an alignment problem. This can also be described as the difference between technical skill and organizational competence. The latter is where real value can be added because it creates a consistent and shared way of doing things.
Different Answers Aren't the Failure
It's tempting to conclude that the fix is a system that always returns the same answer. It isn't, and for many of the most valuable uses of enterprise AI, it can't be.
For example, when an AI persona represents a customer segment, a market, or an audience, its opinions can change depending on the context or the prompt. Of course, this mimics human behavior where the same person can give different answers depending on how questions are framed or the environments change in which they are asked. In AI personas, shifts in responses from similar questions are often the result of subtly different predictions and probabilities.
At some level, that variation must be bounded. The aim is to keep it from swinging to just anything. The questions that actually matter are whether that range is understood, and whether the answer sits inside it. An answer that stays within what the data supports is the system working exactly as intended; one that drifts beyond it is the only real problem.
The Shift Worth Building Toward
The path forward isn't turning every employee into a skilled prompt writer, and it isn't chasing a single identical answer. It's building repeatable frameworks that anyone in the company can use, regardless of how comfortable they are with AI frameworks designed to do two things at once.
The first is to set the boundaries. Structured prompting, such as conversation starters embedded into a workflow, help define which questions get asked at each stage of a research or evaluation process, which criteria the system weighs, and which assumptions it surfaces. The aim is to keep an answer inside the range the data can support rather than letting it drift wherever the phrasing of a prompt happens to push it.
The second is to make the reasoning visible. This is the difference between black box AI, where a question goes in, an answer comes out, and nothing in between can be inspected, and what we at Stravito call glass box AI, where the framework, criteria, and guardrails are transparent and open to challenge. Put simply: Black box AI produces answers. Glass box AI is the standard toward which we build and to which hold ourselves—answers you can trust because you can see the boundaries within which they stayed. Glass box AI makes the AI's thinking transparent so your thinking can be separate from the system and confident because of it.
The Cost of AI You Can't Verify
Most organizations focus on whether AI gives them the right answer. The harder question is whether they'd know if it didn't.
A market sizing built on a flawed assumption looks identical to one that wasn't. A competitive analysis shaped by a leading prompt reads the same as one that wasn't. An answer that quietly drifted outside what the data supports sounds every bit as confident as one that stayed within it. When every output sounds credible, errors go undetected until a bad decision has already been made. At enterprise scale, that risk compounds fast.
Boundaries you can't see are boundaries you can't enforce. Visibility isn't a layer applied on top of the answer; it's what makes the answer usable in the first place.
An answer is shaped far more by what surrounds the model—the data it retrieves, the context it assembles, the guardrails it applies—than by the model itself. The organizations that pull ahead won't have the most powerful models. They'll treat prompting as shared operational infrastructure embedded into workflows, standardized across teams, refined over time. They'll also hold themselves to a simple standard: not making AI give the same answer every time but keeping every answer inside the range the data supports and making that range visible enough to trust.
It's the bar worth setting, because it's what will separate AI that sounds trustworthy from AI that earns it.
Thor Olof Philogène is CEO and founder of Stravito, providers of an AI-powered enterprise knowledge management solution that allows employees at Fortune 2000 organizations to store, discover, share and integrate consumer insights in seconds.