A new global survey of 700 data analysts and 700 IT leaders finds 96% are using AI, but only 49% rely on it most of the time. As agentic AI rises on roadmaps—with 59% expecting to deploy AI agents within 12 months—teams say the hardest part isn’t building mode
The numbers land with a kind of whiplash: 96% of IT pros and data analysts say they’re using AI for their work. Yet when it comes to actually leaning on it day after day, only half—49%—say they’re using AI tools always or most of the time.
The gap is setting the tone for what comes next. Agentic AI is already “high on the agenda. ” with close to six in 10 respondents—59%—predicting they will be actively employing AI agents within the next 12 months. And despite all the talk about speed and autonomy. at least half of respondents say they are willing to grant AI agents “unrestricted access” to their data.
That’s where the friction begins. The survey report doesn’t spell out the security implications of unrestricted access, but it does show a very specific safety instinct: 44% say it’s critical to include human oversight as part of such access.
In other words, many are eager to let agents do the work. But they’re not treating oversight as optional.
In production today, the most common agentic applications are practical and familiar. 59% of respondents use AI agents to draft standardized communications or summaries for stakeholders. 54% rely on agents for scheduling or routing workflow tasks—alert triage and process automation among them. Another 48% say agents generate standard reports or dashboards without manual intervention. while 45% use them to monitor key performance indicators and trigger alerts or actions.
The list also shows where teams still feel tethered to messy reality: 45% of respondents use agents for cleaning. preprocessing. or validating routine data sets. And even with automation spreading. routine analysis isn’t fully handed off—34% use AI agents to run routine statistical analyses or basic predictive models. Only 23% say agents automatically generate insights or recommendations from data.
That last number matters, because it matches the bigger story the survey keeps circling. “Foundational data work” still eats time. Respondents report spending close to six hours per week on cleaning and prepping data for ingestion by AI models or associated retrieval-augmented generation platforms. Of that group, 48% say they spend six to 10 hours weekly on such tasks.
The tools involved look less like a revolution than a continuation: spreadsheets are cited by 61%, business intelligence tools by 56%, and dedicated data preparation platforms by 51%.
The survey authors describe it as AI layering on top of existing workflows rather than replacing them—and the data fits. Even as real-time responsiveness becomes a selling point for modern systems, few organizations seem built for it. Only 20% report that moving from data analysis to a business decision can be done within a few hours. A mere 5% say they support real-time decision-making.
So what’s the biggest barrier when AI does reach the decision stage? It’s not raw capability. Respondents point to a more human problem: explaining AI outputs to business decision-makers. 55% say that’s the difficulty. Another 54% cite limited analytical skills among business users.
Three other roadblocks cluster around trust, responsibility, and fundamentals: 50% say data is not sufficiently clean, integrated, or governed. 49% say there is a lack of clarity on ownership or accountability for decisions. And 45% point to technical limitations of AI tools or infrastructure.
The survey also shows the cost of “making it work,” not just building it. Generating insights from AI isn’t a once-and-done process. Analysts spend almost four hours per week validating or correcting AI-generated outputs. One in six say they spend almost an entire workday—six hours or more—fiddling with AI results.
Add the close to six hours per week spent on foundational data work, and the survey puts the AI “tax” at almost two days per week.
A quiet truth sits under all of this: many organizations are eager to move toward agentic automation, but the operational reality still depends on people who can verify outputs and translate them into decisions business leaders can stand behind.
The survey authors sum it up in a single emphasis on human oversight. arguing that while AI can accelerate work. organizations still need human oversight to ensure outcomes are consistent. explainable. and trusted. And for now. the survey suggests that’s exactly where the rollout stops feeling smooth—and starts demanding more than just adoption.
AI adoption agentic AI IT leaders data analysts human oversight AI agents data preparation AI productivity cybersecurity implications decision-making
So 96% use AI but only 49% rely on it… sounds like they’re just testing and nothing is real. Also “agents” sounds like sci-fi, not my problem.
“Unrestricted access”?? That’s insane to me. Like people keep saying AI is safe but half the time they want it to just do whatever. I don’t trust that at all.
Wait, if 44% say human oversight is critical, how is 59% already using AI agents for drafting emails and dashboards? Sounds like they’re contradicting themselves or the survey is messed up. Also I’m pretty sure my job uses AI for like autocorrect, so idk.
I don’t get why they say rollout lags when “96%” are using it. Maybe they’re using it for stupid stuff like summarizing meetings and then acting like it’s rollout. And agents in 12 months… sure, until someone’s data gets routed to the wrong place or the alert spam starts.