Zoho, the privately held business software maker, is rebuilding Ask Zia, its conversational assistant inside Zoho Analytics, around a multi-agent architecture that lets users ask plain-language questions and have AI agents handle analysis, visualization, data preparation, modelling, coding and actions.
Rather than navigating dashboards or waiting on analysts, a business owner can ask Zia to investigate falling sales, surface likely causes, build a chart and help kick off a workflow in response, a shift the Techweez report frames as agentic analytics replacing static reporting.
Two new layers underpin that shift: a semantic layer standardising shared definitions for terms like "active customer", and an ontology layer mapping how customers, products and departments relate, so the assistant does not misread context it was never given.
Country Head for Zoho Kenya Veerakumar Natarajan said agents need "a robust intelligence layer" on top of raw data to work reliably, echoing the report's point that a strong data foundation underlies all agentic activity.
The platform runs on Zoho's own large language model alongside configurable external models, and opens to outside AI tools via the Model Context Protocol, including a Claude integration; it now supports more than 30 additional data connectors atop 500-plus existing ones, covering SAP, PayPal, Adobe Commerce, Twilio and Notion.
Zoho said customer data is not used to train its models, though queries routed through third-party services via MCP fall under that provider's own data policies, and administrators retain control over AI feature access even as approval requirements for high-impact actions vary by product configuration.
Earlier versions of Ask Zia had already added automated "Key Highlights" and an "Agent Thinking Process" showing its reasoning.
And by June query suggestions could run on Zoho's own language model across US, Indian and European data centers for Premium and Enterprise plans.
Preview versions of the ontology and semantic model are rolling out globally over four to eight weeks, with the platform also offering a bring-your-own-key option for external models including OpenAI.
