There's a report out from SmarterX, called the 2026 State of AI for Business Report, and it's a genuinely big one, over 2,100 professionals surveyed in H1 2026, across roles, industries, company sizes, the whole spread. CEOs, directors, specialists, marketing people, engineers, pretty much everyone who touches AI at work in some way.

Most of it isn't really about training data at all, and it's worth saying that upfront. A big chunk covers things like job displacement fears, how far along companies are in actually adopting AI, whether they have a roadmap or a governance policy, which tool their org pays for. All useful stuff, just not really what this piece is about.

But buried in there, a few findings kept jumping out as things end users are basically telling the teams building AI tools and agents, without meaning to, even though the survey never once frames it that way. Nobody was asked "do you trust the data a model was trained on." Nobody was asked about that at all, actually. Which is kind of what makes it interesting, when people describe frustrations in their own words, without anyone leading them toward a specific answer, you get to see what's actually bothering them underneath the surface complaint.

And honestly, that's the whole value of a survey like this for anyone building AI tools and agents. It's a huge, honest, mostly unfiltered sample of what real end users are running into, day to day, without any of it being framed as feedback for a specific team to act on. Most companies never get feedback this candid or this large. So it felt worth pulling out the three findings that seemed most tied to what a model was actually trained on, and looking at what they're really pointing at underneath.

LLM Output Accuracy Problem

Nowadays, people are starting to treat AI as their personal assistant, and even world's most knowledgeable guru, with all the knowledge in the world inside one tool. Which is great, but what people struggle with is actually relying on the data these tools produce. "Accuracy, misinformation, and hallucinations" shows up as a top concern in the survey, while "getting quality or accurate outputs" is named as one of employees' daily struggles. And here's the thing, "efficiency, productivity, and time savings" is the number one thing people say they're excited about with AI, at 28%, ahead of everything else. But you can't really get any of that efficiency out of a "guru" you don't trust, one that actively gives you misinformation you have to go check yourself. Double-checking the answer kind of defeats the whole point of asking in the first place.

Neither of these is really a prompting problem or a UI problem. It's a training data problem wearing a UX complaint's clothes, and it gets sharply worse outside the English-speaking world.

The 2026 International AI Safety Report tested a model on cultural knowledge questions and found it scored 79% on American culture and 12% on Ethiopian. Another example, researchers at Georgia Tech found Arabic-language models still recommending alcohol in Islamic contexts, because the underlying training text was Arabic translated from English, so the cultural bias traveled with the language instead of getting corrected by it. Overall, the pattern held across 83 languages tested, so it's really not just an occasional gap.

We all kind of know why this happens. It's a lack of training data AI can actually learn from, and a lack of variety in that data too. But an end user doesn't see it that way. Someone hitting this rarely thinks the training data was skewed. They just probably conclude the tool doesn't work well for them.

Ethics and Ownership Questions Behind AI Tools

Here's the thing about trust. Nearly half of the people in this survey, 48%, said they feel neutral, negative, or just unsure about AI's overall impact. Not the tech itself failing them exactly, more like a general unease sitting underneath everything. And when you actually break down what that unease is made of, two smaller categories stand out, "ethics, bias, and fairness" at 6%, and "fear, trust, and ethics" at 5%. Small numbers on their own, but they're probably pointing at two very specific, very real questions that people don't fully have the words for yet. One is ownership, if AI generates something, who actually owns it, the person who typed the prompt, the company that built the model, or whoever's voice or style or face it kind of resembles.

(Side note, the Supreme Court actually touched part of this back in March 2026. They declined to even hear a case, Thaler v. Perlmutter, where a guy tried to copyright an image his AI system made entirely on its own with no prompting, no edits, nothing human involved at all. The ruling that stands says no human authorship means no copyright, for anyone. But that's the extreme version of the question. The messier, way more common version, someone actually prompting and tweaking and iterating, still doesn't have a clean answer. Anyway.)

The other question is origin, how did the model even learn to do this in the first place, whose writing, whose art, whose voice went into it, and did anyone actually agree to that.

These two questions are just hard to phrase in a multiple-choice survey, so they end up compressed into vague categories like "ethics" and "fear," when really each one is doing a lot of quiet work underneath.

Here's what's interesting though. Nobody in this survey said that they didn't trust the training data. That phrase probably doesn't even occur to most people because most people don't really think about AI in terms of what it was trained on at all. They just experience the result. So, distrust gets filed under "I don't trust AI" as this big vague blob, when really, if you actually dig into it, a lot of that distrust is coming from something much more specific and much more traceable, the data itself, where it came from, whether anyone checked it, whether it's even any good.

Agents Are Popular, While Reliability Isn't There Yet

Last but not least, agents, the world's favourite thing right now. 40% of people in this survey said agentic AI is the trend they're following most closely, way ahead of anything else on the list, and it's the second most requested training topic at 51%, right behind basic workflow integration.

Here's the thing, though, actual reliability data on agents right now is honestly rough. Recent industry research puts agent failure rates in production somewhere between 70% and 95%, depending on the task. Part of that is just pure math. Each step an agent takes has some small chance of going wrong, and those small chances add up fast the more steps there are. A task with ten steps ends up failing way more often than a task with two, even if the agent's actually pretty good at each individual step.

The interesting thing also is where exactly the failures happen. Research using the WebArena benchmark found agents fall well below human performance mainly on tasks that need flexible navigation, recovering from something unexpected, or reasoning across multiple steps that don't go according to plan. Basically, agents are decent on the clean, scripted version of a task and fall apart the moment reality gets messy, a page loads weird, a step fails halfway, something needs correcting mid-task.

Which, if you think about it for a second, is a data problem more than anything else. Most of what trained today's models to be good conversationalists is text, and text is naturally tidy, someone already cleaned it up before it got published, but an agent needs the opposite. It needs to have seen what a mistake actually looks like, what correcting a mistake mid-task looks like, what a real, physical, multi-step attempt at something actually looks like from start to finish, including the parts that don't go according to plan. Text really can't show that, no matter how well it's written. Video can, because video is the one format that actually captures continuity over time, cause and effect, one thing happening and then the next thing happening because of it. That's just true of the medium itself, not really a solvable problem with better prompting or cleverer text data. Nobody in this survey was asked about any of that. But the excitement around agents and the actual reliability numbers on agents are describing the exact same gap from two different directions.

What Teams Building AI Tools and Agents Should Take From This

Put all three of these next to each other and a pattern kind of emerges, even though the survey itself never draws it out. The accuracy problem traces back to training data that's too narrow, mostly English and mostly Western, so the model just doesn't have enough of the actual world represented in what it learned from. The trust problem traces back to not really knowing where the data came from in the first place, who made it, whether they agreed to it being used this way, whether anyone can actually answer that if asked. And the agent problem traces back to training data that's too clean, too scripted, missing all the real, messy, multi-step stuff that reliability actually depends on.

Three different symptoms, but honestly the same root thing underneath all of them. A good model, or a good agent, is really built on data that's genuinely diverse, across cultures, languages, situations, people, environments, all of it, rather than a narrow slice scaled up really big. It's built on data where the origin is actually known, so trust is something you can actually verify instead of just hoping works out. And it's built on data that captures real continuity, real attempts, real mistakes and corrections, the whole messy version of a task rather than the tidy one.

None of that is a small ask, and honestly most of the industry is still figuring out how to get there. But it's at least a clearer target than just "more data," which is kind of where this conversation has been stuck for a while now.

If you're working through any of this yourself and want to see what genuinely diverse, traceable, real-world training data actually looks like, you can request a free sample set and judge it for yourself.