Most of the agent announcements I’ve covered this fall start with the agent. Zoho started underneath it. Today, Zoho announced a major update that turns Zoho Analytics into what it calls an agentic analytics platform, and the most important part isn’t the agents. It’s the business context layer they sit on.
That layer has two parts. One’s an ontology that maps a company’s core entities and how they relate to each other. The other is a semantic layer, which Zoho describes as a business glossary for AI that defines what terms and metrics actually mean for that business. Both are built into Zoho Analytics, so you don’t need a separate data platform or warehouse to get them.
I think that’s the right order, and it’s rarer than it should be. I keep a nine-question test for telling a real context layer apart from a set of integrations. One of those questions is whether the system knows what a field means, not just where it lives. Zoom’s context layer announcement last week didn’t really address it. Zoho built its whole announcement around it.
Zoho’s example in the analyst pre-briefing made it concrete. Picture a coffee chain with transactions in a cloud database, memberships in its CRM and campaigns in HubSpot. The company builds an ontology of its core entities, then a semantic model for its loyalty program that defines how a Gold or Bronze tier is actually calculated. When someone asks an agent about loyalty, the agent works from the company’s own definition instead of guessing.
Zoho also says its semantic models are compatible with Apache Ossie, the open specification for sharing semantic definitions between tools. So a company’s definitions aren’t locked inside Zoho, and that deserves credit. One thing to keep in mind is that the spec is still early and doesn’t yet carry access policies or row-level security. The meaning can travel today, but the rules about who’s allowed to see it don’t travel with it yet.
What Zoho showed
On top of that foundation sits Ask Zia, which Zoho has rebuilt as a super agent that coordinates six specialized agents. They cover data analysis, preparation, modeling, visualization, coding, and actions, all working from the same semantic layer and conversation. You can run it on Zoho’s own model or bring your own. Teams can also build their own agents on the same definitions and connect them to outside systems through MCP. In the demo, an agent created purchase orders in SAP that sat waiting for approval. All of it is available headless, through MCP servers, a command line, and a Claude plug-in, and agents can be embedded in other apps with Zoho’s permissions still applied.
Zoho also spent real time on governance. A data catalog shows lineage, so you can see what changes downstream when something upstream moves. An observability dashboard shows which agents ran, who used them, what they accessed, and what it cost. And a sandbox keeps development apart from production, with versioning so you can roll back a change.
Pricing didn’t change, and every plan now includes Zoho LLM credits. The Basic plan starts at $24 a month billed annually, with two users and 10,000 credits, and Enterprise is $445 a month for 50 users and 100,000 credits. Building AI into the plan instead of metering it separately is buyer-friendly. The question I’d want answered is what a credit actually buys, and what happens to the bill when agents start doing more of the work. The cost tracking in the observability dashboard should help there.
Most of this is available now in the cloud version of Zoho Analytics. The ontology and semantic model, the foundation the whole release rests on, arrive as previews rolling out globally over the next four to eight weeks, along with pixel-perfect reporting. The on-premises version follows in the first quarter of 2027.
Where the work goes
The most candid moment in the briefing came from Zoho’s own presenter. He made the point that defining an ontology is the easy part. Versioning it and keeping it current is the hard part, and it will remain a job for a well-staffed team. I’ve been calling that the Friction Tax. The work doesn’t disappear when an agent can answer the question. It moves to the people keeping the definitions right. Zoho deserves credit for saying so out loud, and buyers should budget for it.
The design choice to watch
In the briefing, Zoho explained that agent memory is per user today. If a user corrects the agent enough times, their correction takes priority over the semantic layer. That means the governed definition everyone is supposed to share can be quietly overridden one person at a time. My daughter likes to remind me that you can be loud and wrong. In this case, whoever corrects the agent most persistently ends up with their own version of the truth.
Zoho says it’s still deciding how account-wide learning will work, so this is a good moment to weigh in. My suggestion is to treat repeated corrections as a sign that a definition might be out of date, and route them to whoever owns the semantic model as a proposed change. That way the definition gets better and everyone benefits.
The SAP demo raises a related question. The purchase orders waited for approval, which is how it should work. Then the agent suggested automating the whole workflow on a schedule. Every handoff used to be a chance for someone to check the work, and the approval step is the first thing automation offers to take out.
Where it’s heading
Zoho says the next step is moving from what it calls designed context to dynamic business context, where the platform keeps watching how a business’s data changes over time. It also plans to learn from the outcomes of the actions its agents take. That’s interesting next to Zoom, which talked about dynamic context last week from the other direction. Zoom starts from the conversation and works toward the record. Zoho starts from the definitions and works toward what’s changing. Both are heading toward the same place, and both will run into the same question. Once the context starts updating itself, who approves the change?
What I’d ask before buying
If you’re evaluating it, start with the preview timeline, since the layer that makes the rest of it trustworthy is the last piece to arrive. I’d ask Zoho which source wins when two connected systems disagree about the same customer. I’d ask whether permissions and audit trails travel with the context when an outside agent pulls it through MCP. And I’d ask what a credit buys once agents are running at scale.
Zoho is positioning Analytics as the intelligence layer for its whole stack. If the semantic layer holds up, that’s the right foundation to build on. The open question is the same one every vendor is facing this fall, which is whether the definitions stay governed once everyone starts teaching the agents.
