The wall between business intelligence and data science is crumbling as AI agents write code, test hypotheses, and explain anomalies that once required data scientists' expertise. This change brings new opportunities – and new challenges.
Gartner predicts that by 2028, 60% of self-service analytics users will rely on general-purpose LLMs for ad hoc and exploratory analysis. Major data platform vendors, including Microsoft, Google and Databricks, are pairing new agentic AI interfaces with semantic, context and governance layers to extend data access to more users. This shift raises a few hard questions: Who owns the data definitions? Who reviews agent output before it reaches a decision-maker? Who is accountable when a plausible answer turns out to be wrong?
Managing probabilistic failure
Not too long ago, a dashboard request was considered business intelligence. Custom analytics for specific use cases, often done in R or Python, was considered data science. Analysts often sat between the two disciplines, directing requests to the appropriate team.
Technology is only part of the problem. The larger challenge is organizational. Agentic analytics upends definitions that were once fixed: what the data means, who is allowed to view or process it and who signs off on whether an answer is trustworthy and applicable to a particular analysis. These questions relate to process, authority, and trust as much as tooling.
"The useful boundary isn't BI vs. data science, it's deterministic vs. probabilistic," said Jimmy Carter, co-founder of Daemo AI.
For Daemo – a startup that deploys agentic analytics for government affairs and lobbying firms – anything that has to be auditable and repeatable runs as plain SQL against the warehouse. Deterministic processing for filtering, matching and monitoring sits in that layer because it is cheap, predictable and easy to govern. LLM agents sit on top and handle the probabilistic work of synthesis, prioritization and explaining why a number moved. The two layers do not break the same way.
"Deterministic layers fail loudly. Agentic layers fail plausibly," Carter said. "That second property is the one you have to govern around."
Deterministic failures are fairly straightforward to monitor and triage. A plausible but incorrect answer requires human review, a step that creates friction and can be skipped when the output looks finished or good enough. That makes review a political question as much as a technical one. Someone must be willing to delay a finished-looking deliverable, and leaders have to back up that decision. Carter said the real executive question isn't which tool owns production, but which review scaffolding an agentic workflow must pass before its output earns the right to go out without human approval.
"We treat that as an earned status per workflow, not a platform decision," he said.
One result is that agentic analytics changes what organizations must govern. Each workflow requires a designated individual with the competence and authority to judge results. When business users direct agents in natural language, the instruction set serves as the governance layer.
"The most important document in our system isn't really a dashboard spec, it's the written, versioned specification of what each agent may do, what it must escalate and when it has to say, 'I don't know,'" Carter said.
Determining where human checks belong
Chinni Krishna Abburi, business intelligence manager at Communities Foundation of Texas, is integrating AI agents into the foundation's analytics work after spending about seven years deploying BI across most of the major tools. Exploratory work is moving to general LLMs, while reporting for leadership or the board remains in governed BI.
"A figure in a grant report or donor report can't be 'probably right,' so the real question becomes where you put the human check, not whether you keep one," Abburi said.
The foundation's dashboards use real-time data, which builds trust. When an agent answers a question, someone can check the live dashboard to confirm the number themselves.
"Having that source of truth sitting right there is what makes people comfortable acting on what the agent says," Abburi said.
The team also keeps its AI agents scoped to specific datasets rather than the entire environment. This keeps answers accurate and reduces the risk of unauthorized access or inappropriate data use. Access control is central to agentic analytics. The right information must reach the right people and stop at everyone else. As agents make it easier for anyone to query the data, access control becomes as important as accuracy.
"You must design that in early. It's painful to add later," he said.
Govern the definition layer rather than the query layer
Aigars Pilmanis, founder of VolRadar, an options and volatility analytics platform, said that agentic analytics is shifting the boundary from capability to auditability. An agent can generate a hypothesis and suggest an explanation for an irregularity, but it cannot validate the business definition of the metric it queried.
That is why every figure on VolRadar's platform traces back to a documented formula and a single primary data vendor. The real risk in analytics is not merely wrong data; it also includes metrics whose sources, formulas and definitions are undefined.
"So, govern the definition layer, not the query layer," Pilmanis said. "Let agents explore anything whose inputs you can already defend."
Why data quality still determines agent reliability
Bob Brauer, founder and CEO of Interzoid, a data enrichment and entity resolution provider, said many organizations believe agentic analytics only affects who can analyze the data. But the reality is that agentic analytics is no different from any other AI deployment: The data feeding the agent must be accurate and verifiable.
"AI has put added pressure on just how important trusted data becomes for AI, because AI can only reason, or determine steps to take, from the information it's given," Brauer said.
Agentic analytics changes the structure of accountability for who is responsible for the decisions AI makes, as well as determining when a human needs to step in.
"A lot of organizations fall into the trap of adding more AI capabilities before they have the processes, data quality and governance in place that is required to support them and ensure successful outcomes," Brauer said.
Technology changes how work gets done. It doesn't change who is accountable for the outcome.
"Agentic analytics moves the boundary, but it does not erase it," said Patrick Gibbs, founder of Epiphany Dynamics, which builds and governs AI agent workflows for small and midsize businesses.
Gibbs divides ownership by consequence rather than by whether the tool can technically complete the task. Business users can own exploratory questions when the agent is read-only, works from approved data, shows its query and assumptions and labels the result as provisional.
BI teams should own recurring metrics, shared data definitions and implementation of access controls. Data science experts must own data products and workflows deployed to production. They might use AI agents to generate the code for those jobs, but that does not mean business users have the expertise to properly judge whether the methodology is valid.
Gibbs advised organizations to require every agentic analysis to retain its source data, generated code, assumptions and review history. They should also name one accountable owner before the output affects a dashboard, customer or automated action.
"If nobody can answer who signs off when the agent is wrong, the workflow is not governed yet," said Gibbs.
George Lawton is a journalist based in London. Over the last 30 years, he has written more than 3,000 stories about computers, communications, knowledge management, business, health and other areas that interest him.
