The No Nonsense Guide to Zia Models, Zia Agents, and Zoho MCP for Business Leaders Ready to Start with AI

TL;DR

Zoho shipped three AI products in the past year. Most teams turned on one of them, got a slightly better email composer, and stopped there. The other two are sitting unconfigured in your stack right now, and one of them can run your pipeline autonomously while you read this post.

The gap between “AI that helps you write” and “AI that qualifies your leads at 2 AM” is one settings page and a model selection.

This post maps all three products and shows you what each one actually does to your operations. It walks through Salesforce CRM Benchmark data on which models perform best for which CRM tasks. And it gives you a concrete starting point based on how much configuration time you have. Fifteen minutes, one hour, or half a day.

Three Depths, One Platform

Zoho has shipped a significant amount of AI capability in the past twelve months. If you are a Zoho admin or a business leader running your operations on the platform, you have probably noticed the announcements stacking up. Zia Models. Zia Agents. Zoho MCP. Each one sounds like it could be the big unlock. Each one does something meaningfully different from the others.

The challenge is that these three products sit at different depths inside the same platform, and the boundaries between them are obvious only after you have configured all three. An admin who turns on multiple LLMs in Setup and expects their AI to start reasoning about pipeline data is going to be confused. It only helps with email drafting. Meanwhile, the feature that actually reasons about pipeline data, Zia Agents, lives in a completely different part of the stack and most teams have not configured it yet.

This post maps all three in one place. The goal is to give you a clear picture of what each one does and where model choice actually affects outcomes. We will walk through them in order of depth, starting with the most accessible and working toward the broadest.

Figure 1 – Three depths of Zoho AI: from drafting assistance, to autonomous operations across Zoho apps, to external AI orchestration.

Zia Models

Figure 2 – Zia’s assistive AI threaded across CRM, Desk, Books, Analytics, and Mail. Each app gets its own Zia surface tuned to its workflow.

How Zia Works Across Zoho Apps

Zia is Zoho’s AI assistant, and it runs across the entire suite. In CRM, Zia handles lead scoring, deal prediction, and record summarization. In Desk, it powers sentiment analysis and reply suggestions. In Books, it answers natural language financial queries. In Analytics, it runs conversational reporting through Ask Zia. In Mail, it drafts and rewrites messages. Each app has its own Zia surface tuned to that app’s workflow.

The feature that ties these surfaces together is model configuration. Zoho lets admins enable multiple LLM providers in parallel and connect external API keys. Zia’s generative features (drafting, rephrasing, content generation) then run on the model of your choice. In Zoho CRM, this configuration lives at Setup > Zia > Models. In Zoho Workplace (Mail, Calendar, WorkDrive), admins connect the same providers through Workplace settings. The walkthrough below uses CRM as the primary example because it has the most mature implementation and the clearest configuration screen.

Configuring Providers in CRM

In CRM, the recent update lets you run multiple LLM providers at the same time, which is a meaningful change from the previous setup where only one could be active. Today you can enable Zia (Zoho’s native model), OpenAI, Google Gemini, Anthropic Claude, and Cohere, all in parallel. Your users get a dropdown to switch between them. Each external provider requires your own API key, and admins can control who sees what through profile level permissions.

The division of labor between Zia and the external models matters. Zia handles the analytical work. Record summaries, conversation summaries across calls and emails and meetings, notes summaries, email thread summaries. These features read your actual CRM data, the contact names, deal stages, interaction histories, and synthesize it into something you can act on quickly. That is Zia’s territory.

The external models occupy a different surface. They show up inside the note editor and the email composer, where their job is drafting, rephrasing, adjusting tone and structure. If you are composing a follow up email and want it tightened up or shifted to a different register, that is where OpenAI or Anthropic or Gemini earns its keep. The Template Assistant works across all models as well.

The privacy design reinforces why this separation exists. External models only process the content in the composer window. They do not see the full email thread, the recipient list, attachments, or the underlying record data. Sensitive and PHI encrypted fields stay excluded entirely.

For most teams, this is the right starting point. It takes fifteen minutes to configure in CRM. The cost exposure is predictable because you control the API keys. Every user on your CRM gets a better drafting experience without changing how they work. The analytical intelligence from Zia runs from the start. The external models add stylistic range on top of it. Similar configurations exist in Desk, Mail, and other Zoho apps through their own Zia settings.

Figure 3 – Setup > Zia > Models, with multiple providers enabled in parallel.

Where Zia Models Falls Short

Worth calibrating expectations here. Zia’s summaries are functional, but they stay at the surface level. You get a recap of what happened on a record, who said what, what the last activity was. What you do not get is strategic depth. Zia will summarize a deal’s email thread accurately. It will not tell you that the deal is stalling because the champion went silent after pricing was shared, or recommend a specific re engagement approach based on the buyer’s behavior. That kind of reasoning requires the more capable models, and today those sit in a different part of the stack (Zia Agents, covered in the next section). For basic daily acceleration, Zia’s native summaries do the job. For anything that requires judgment, the external models carry the weight.

Figure 4 – Zia’s Record Assistant returns generic CRM process advice when asked for a strategic recommendation, instead of reading the email and tailoring to the conversation.

One clarification that saves confusion later. If you add an OpenAI or Anthropic key here in Zia Models, that key powers drafting inside your editors. A separate configuration in Zia Agents (covered in the next section) connects the same or different models to power autonomous digital employees. The two configurations are independent.

Zia Agents

Zia Models helps your team write faster and summarize better across Zoho apps. Zia Agents is where Zoho’s AI starts to think.

Figure 5 – A pre built SDR agent handles a new lead end to end without a human touching the record. The agent reads, scores, drafts, schedules, and updates the pipeline.

What Zia Agents Does

The difference is fundamental. Zia Models sits inside your editors and helps you compose. Zia Agents deploys autonomous digital employees across 60+ Zoho apps with access to over 700 pre built actions. They qualify leads, enrich data, analyze lost deals, coach reps, route support tickets, screen candidates, and send targeted emails. All without a human working through each step.

Think of it this way. Zia Models is the intern who improves your draft. Zia Agents is the employee who handles the task while you sleep.

Zoho gives you two ways to get started. The Agent Store offers 100+ pre built agents ready to deploy. The SDR Agent nurtures and qualifies new leads through targeted emails, handles objections, and schedules meetings. The Deal Analyzer reviews unsuccessful deals and surfaces the top 3 reasons for each loss. The Customer Support Agent processes incoming requests, reads the context, and either answers directly or routes to a human rep. The Candidate Screener ranks applicants by role requirements, skills, and experience. Each one plugs into your existing Zoho environment and starts working within minutes.

For anything the pre built agents do not cover, Agent Studio lets you build your own. It is a no code environment. You describe the agent’s role in natural language, select which model powers it, assign tools (actions the agent can perform, like “Get Deals” or “Create Task”), connect knowledge sources (PDFs, SOPs, WorkDrive files), and deploy. No developer required. The agent runs inside Zoho Cliq, inside specific Zoho apps, or through Ask Zia.

Every agent operates within your existing user permission structure. Admins control which tools and data sources each agent can access. Every action is logged with a full audit trail. You can deploy agents as rule triggered (fires when a condition is met), button activated (a user clicks to invoke), or fully autonomous (runs on its own schedule).

We covered the full deployment architecture in our post on Zia Agents as digital employees in Zoho CRM.

Choosing Your Model

Here is where most Zoho admins stop reading and miss the most important decision in the entire setup.

Every Zia Agent runs on a language model. You pick that model when you create the agent. In Zia Models, model choice affects writing style. In Zia Agents, model choice affects reasoning quality, scoring accuracy, and the reliability of every autonomous action the agent takes. Different decision. Higher stakes.

Zia Agents gives you two paths.

ZKS (Zoho Key System) requires no API key. Zoho manages everything. Under ZKS you get two categories. Zoho hosted models (Qwen 14B, Qwen 30B MoE, GLM 4.7 Flash, GLM 5) come with a free monthly token allowance. OpenAI models (GPT-4o-mini, GPT-4.1-mini, GPT-4o, GPT-4.1, GPT-5.1) consume AI credits at exact vendor rates from the first token.

BYOK (Bring Your Own Key) lets you connect your own API key from OpenAI, Anthropic (Claude 3 Haiku, Claude Sonnet 4, Claude Opus 4), or Google Gemini (2.0 Flash, 2.0 Flash Lite, 2.5 Pro, 2.5 Flash). Zero cost on the Zia Agents side. Your vendor bills you directly.

What Zia Agents costs (from Zoho’s Platform Details page)

The platform itself is free. Agent builder, deployment infrastructure, and BYOK integration all come at zero cost. Zoho hosted Standard models (Qwen 14B, Qwen 30B MoE, GLM 4.7 Flash) include 30 million tokens per month free. The Pro model GLM 5 includes 20 million tokens per month free. After the free allowance, Standard runs $1.00 per 1M tokens and Pro runs $3.00 per 1M tokens. OpenAI through ZKS (Zoho managed keys) carries no free allowance and consumes AI credits from the first token at exact vendor rates. BYOK costs zero on the Zia Agents side. Purchased credits roll over month to month. The free monthly allowance resets on the 1st and does not roll over.

Which Model Should You Pick

The short answer is GPT 4.1 Mini through ZKS. No API key needed, competitive accuracy across summarization, generation, and agentic tasks, fast, cheap, and 89% Trust and Safety on the Salesforce CRM Benchmark. It is the safest default for your first agent.

The longer answer is that model rankings shift depending on the task. The model that wins summarization loses at generation. The model that wins generation loses at agentic reasoning. Claude leads Trust and Safety at 93% but the versions available in Zia Agents BYOK (Sonnet 4, Opus 4) are older than the benchmarked versions. Gemini 2.5 Flash wins on speed for high volume workflows. The right model depends on the specific job your agent handles.

We are publishing a dedicated model selection guide that walks through the Salesforce CRM Benchmark data filtered to Zia available models, with recommendations by agent role (SDR, deal analyzer, support routing). Here is the distilled view.

Zia Agents available models in the benchmark (June 2026):

Claude Sonnet 4, Claude Opus 4, and Claude 3 Haiku are available via BYOK but the benchmark evaluates newer versions. Zoho hosted models (Qwen, GLM) are not in the Salesforce benchmark. Check salesforceairesearch.com/crm-benchmark for the latest.

The follow up post will include the full benchmark screenshots, task by task analysis, and cost per task breakdowns.

Figure 6 – ZKS Zoho hosted models, ZKS external models, and BYOK: three billing structures for the same agent platform.

Zoho MCP

Zia Models powers assistive AI inside your Zoho apps, from email composers to record summaries. On the Zia Agents surface, autonomous digital employees handle tasks across Zoho’s app ecosystem. Zoho MCP goes further. It opens your entire Zoho stack to any external AI agent operating outside the Zoho interface entirely.

MCP stands for Model Context Protocol. It is an open standard that gives AI agents a structured way to connect to business software, read live data, and take action. The analogy that works best is HTTP. HTTP made it possible for any browser to load any website regardless of who built either side. MCP does the same thing for AI agents and business applications. Any MCP compatible agent, Claude, GPT, Cursor, or others, can connect to any MCP compatible application through the same protocol.

Zoho built its implementation at mcp.zoho.com. The platform turns Zoho CRM, Books, Desk, Projects, and 15+ other apps into agent ready endpoints. You create a server, select which apps and actions the agent can access, copy the MCP URL, and paste it into your AI client. The setup requires no coding and takes under an hour for a custom server, less for a preconfigured one. The agent operates within the connected account’s existing permission structure.

The difference is where the AI lives. Zia Agents run inside the Zoho environment, deployed per task with defined roles and permissions. An MCP connected agent sits outside Zoho entirely and orchestrates a multi step workflow across apps in a single conversation. Pull a deal from CRM, generate an invoice in Books, open a follow up task in Projects, send a notification in Cliq. One prompt, multiple apps, zero human routing between steps.

We tested this against a real Zoho CRM account with 183 stalled leads. Claude, connected via Zoho MCP, received one prompt. It pulled every record in the Not Responding stage and read the interaction histories. It built its own scoring methodology on the fly and returned a tiered priority list with re engagement recommendations. What would have taken a sales manager the better part of a working day took minutes. We wrote up the full setup, the test, and where it falls short in our Zoho MCP and Claude integration guide.

For teams that want an external AI client like Claude or GPT driving workflows across Zoho apps in a single conversational session, MCP is where that capability lives. For teams that prefer a workflow automation approach with multiple specialized agents coordinating through n8n, we covered that architecture separately. See our post on building an AI agent mesh for Zoho apps.

Figure 7 – One prompt triggers actions across CRM, Books, Desk, Projects, and Cliq through a single MCP connection.

Choosing the Right Depth

The decision depends on what your team needs AI to do and how much configuration time you can commit.

Figure 8 – Match your team’s readiness and use case to the right entry point.

Where to Start

If you have not touched any of this yet, start with Zia Models. Open Setup > Zia > Models, enable the Zoho Hosted LLM, and add one external provider. OpenAI is the most common starting point. Set your profile permissions so the right roles see the right providers. Let your team use it for a week before you evaluate whether to add more providers. The whole process takes fifteen minutes and the value shows up the same day in faster email drafting and cleaner notes.

Once your team is comfortable with that, the next move is Zia Agents. Start with a read only agent. Something that pulls records, scores them, or surfaces information your team would otherwise spend time hunting for manually. The 183 lead test we ran through MCP followed exactly this principle. Reading data and reasoning about it carries low risk. Writing data, creating records, updating fields, sending emails, is where you want a sandbox first. Pick your model deliberately. If your agent handles structured data work, you may find that a smaller, faster model gives you what you need at a fraction of the cost of a frontier model. If your agent handles complex reasoning across multiple related records, that is where the more capable models earn their price.

Zoho MCP is for when your workflow spans Books and Desk and Projects and Cliq, and you want one AI agent coordinating across all of them. The setup at mcp.zoho.com works for anyone who can follow a configuration screen. The scoping decisions around which actions to expose and which account permissions to connect require the same care you would give to onboarding a new employee. Give the agent access to what it needs. Scope tightly.

The common thread across all three is that Zoho has built the infrastructure for AI at every depth of the stack. The question for your team is which depth matches what you need today, and which one you are building toward.