Published on the 18/06/2026 | Written by Heather Wright

Not every decision needs an AI agent…

As organisations explore AI agents to automate workflows and make data-driven decisions at scale, Fay Fei is warning local organisations to take pause, with blanket rollouts doing more harm than good.

Fei, a Gartner director analyst, says in many enterprise environments, agentic analytics can introduce additional cost, complexity and risk – particularly where decisions are routine, tightly regulated or based on immature data foundations.

“Bad insights become bad decisions at scale. And that’s a much bigger problem.”

Analytics and business intelligence platforms are increasingly integrating AI agents to transform how data is analysed and insights are delivered, and businesses across Australia and New Zealand are increasingly experimenting, often layering new capabilities onto existing data environments. Many already have well-established reporting platforms in place and are now looking to extract more value – particularly through natural language interaction and automated insights.

But Fei told attendees at this week’s Gartner Data and Analytics Summit in Sydney that the push to add agency to analytics workflows needs tighter discipline. She says AI agents fundamentally change how analytics operates, and that shift is not suited to every use case.

At its core, agentic analytics moves organisations away from the traditional linear BI workflows – where data is cleaned and transformed, queried and visualised in dashboards, step-by-step, before insight is interpreted manually – into systems where agents handle much of the heavy lifting. These agents can set multi-step processes, automating data preparation and generate insights with minimal human intervention and in a more conversational based interface.

“The biggest trend here is that Ai agents are going to proactively detect changes in context and adjust insights and actions,” she says, noting that the key thing about agents – differentiating them from traditional automation – is they are goal-driven, not task-focused.

That shift has clear appeal, but also introduces new requirements around governance, trust and data readiness that many organisations are still grappling with.

Where agents don’t belong

Fei was clear that agentic analytics isn’t a universal fix, and is ‘overkill’ in many cases and actively counterproductive.

“There are some situations where it makes sense to just adopt a traditional reporting tool or BI, rather than agentic analytics.”

Among those are highly regulated environments, which need strict compliance and auditability. Traditional analytics workflows remain better suited where decisions must be fully traceable, deterministic and auditable. In areas such as large financial reporting, organisations need to understand exactly how outputs are produced – something that can be challenging with systems built on large language models.

Similarly, routine business processes offer little justification for introducing agents, Fei says. Where the objective is simply to view static metrics or track recurring KPIs, without the need for adaptive decision, agentic analytics adds unnecessary overhead without delivering additional value.

“These are low ROI use cases,” Fei says, noting that agentic systems typically require significant investment across compute, integration and governance.

In these scenarios, organisations risk over-engineering solutions and introducing complexity where existing tools already meet business needs.

Data maturity still dictates success

Beyond use case selection, the success of agentic analytics hinges on the underlying data environment.

Fei is clear: Organisations with immature data foundations are unlikely to succeed. Instead, she recommends addressing data integration and quality issues before attempting to deploy AI agents at scale.

“AI won’t solve your data issues. It only amplifies them,” she says.

“If the data is immature itself and the overall integration readiness is low… you won’t be successful with agentic analytics.

That has implications for organisations attempting to build more advanced capabilities on top of existing reporting systems. While there is clear momentum to make data more accessible and interactive, the underlying governance, semantics and quality still need to be in place.

The cost of getting it wrong

Agentic analytics doesn’t just change how insights are produced, it changes where risk sits.

Fei notes once organisations introduce AI agents into analytics workflows, the stakes move beyond reporting errors. “The truth is AI will make bad analytics organisations fail faster.”

Impact will be amplified. “The impact is not just bad insights but bad decisions at scale. That’s a much bigger problem.”

She says that helps explain why 32 percent of respondents in a recent Gartner survey believe truly usable AI agents remain a distant reality. The same survey showed only 19 percent of those surveyed are using AI agents to some extent on a daily basis.

Avoiding agent washing

Fei also called out agent washing. “Not everything offered or claimed as an AI agent is indeed an AI agent.”

She says agent washing is emerging as a prevalent trend, with vendors rebranding legacy features as AI-powered or agentic capabilities, often without meaningful innovation. In some cases products rely on basic natural language queries, scripted workflows or pre-defined reports, while presenting them as autonomous systems.

The absence of key features such as multi-step reasoning, hypothesis testing or proactive recommendations, signals the tools are not agents.

That widespread agent washing is threatening to undermine customer trust and stall true innovation, and is creating market friction and pervasive AI fatigue among B2B buyers.

Fei urged organisations to adopt an evidence-based evaluation when it comes to agents, evaluating them on real-world scenarios, and implementing task-driven acceptance tests. That includes assessing how systems handle enterprise data volumes, complex integrations and governance requirements. She also recommended engaging in detailed roadmap discussions with vendors and urged prioritising explicit transparency features, such as code visibility, confidence scores and semantic grounding, when selecting agentic offerings.

The path forward

Rather than pursuing broad deployment, Fei points to a more structured, refined, AI adoption model.

The first step is defining where agentic analytics truly add value, focusing on use cases with clear decision logic, strong data foundations and a need for adaptive insights. She suggested picking three to five low-risk high-value use cases as starting points.

From there, Fei says organisations can take a stage approach, establishing cross-functional teams, testing shortlisted platforms, running proof of concepts with task-driven acceptance tests and ultimately scaling wins, raising autonomy where trust is earned and adopting multiagent platforms.

Fei’s underlying message is clear: Agentic analytics represents a significant evolution in how organisations interact with data – but it is not a replacement for all existing analytics practices.

“It’s not everywhere for all your use cases,” she says.