By Enterprises are paying a compounding cost for defaulting to frontier AI models on every workload. Token spend balloons, hallucinations generate correction loops that consume even more tokens, and the total cost of AI operation grows without a proportional increase in business value. The assumption that bigger models equal better outcomes is not holding up under production conditions.
In this interview on TFiR, Tony Thomas, Co-founder at Zoho Corporation, covers how ManageEngine and Zoho are building purpose-optimized internal models, integrating open-weight models, and applying workload-specific sizing to reduce cost without compromising output quality.
Guest: Tony Thomas, Co-founder at Zoho Corporation
Show: The Agentic Enterprise
Watch the discussion on YouTube
Here is what every platform engineer and enterprise IT leader needs to know.
Technical Deep Dive
Q: What is the actual business value of AI in the enterprise beyond feature improvements?
Tony Thomas, Co-founder at Zoho Corporation, draws a clear line between AI as a product feature and AI as a driver of business outcomes. The first wave of enterprise AI showed up as incremental product improvements, but the real unlock is using AI to solve larger problems more efficiently, faster, and with better service delivery. Thomas argues that outcome impact, not feature count, is the correct measure of enterprise AI value.
“The real value of AI getting unlocked in business is based on improving outcomes, solving bigger problems more efficiently and faster, and providing a better business service.”
Tony Thomas, Co-founder, Zoho Corporation
Q: Is token consumption a useful metric for measuring enterprise AI success?
Thomas treats token consumption as a vanity metric comparable to measuring a car’s value by how much fuel it burns per mile. The productive inversion is outcome per token: how much business value is generated per unit of AI spend. As tokenomics becomes a board-level topic, organizations that optimized for consumption volume rather than value density are finding the model unsustainable.
“It should be less about how much tokens you are burning. It should be more about how much value you are getting by spending less tokens.”
Tony Thomas, Co-founder, Zoho Corporation
Q: How does Zoho approach AI affordability and sustainable cost structures?
Affordability is a stated design constraint at Zoho, not a secondary concern. Thomas describes it as the central ethos: delivering value to customers without breaking the bank. That constraint shapes how AI is selected and sized across both Zoho and ManageEngine product lines, pushing the team toward models that fit the job rather than defaulting to the most capable and most expensive option available.
“Affordability has always been at the center of our strategy. Zoho looks at sustainability, being able to provide value to customers without breaking the bank.”
Tony Thomas, Co-founder, Zoho Corporation
Q: Can vibe coding or AI-generated code be used safely in enterprise security and compliance contexts?
Thomas is explicit that security and compliance requirements place hard limits on where AI-assisted or vibe-coded development is acceptable. Infrastructure products must be built on a solid, auditable foundation. The parts of the product that can be accelerated through AI-assisted development are separated from the parts where security controls and compliance requirements override development speed. Security trumps velocity for core infrastructure.
“Security is essential. We can’t vibe code these products. From compliance to being able to ensure that the infrastructure is secure, that trumps everything else in terms of priority.”
Tony Thomas, Co-founder, Zoho Corporation
Q: Why do large frontier AI models cause problems in production enterprise environments?
Large frontier models are not scoped to specific tasks, which means they hallucinate, pursue tangential reasoning paths, and consume tokens on content that does not answer the query. When a model makes a mistake, the correction itself costs additional tokens, compounding the error overhead. Smaller, focused models with guardrails applied to specific workloads produce more accurate outputs with less correction overhead and lower total cost.
“If AI makes a mistake, you have to pay to fix the mistake because you have to waste more tokens to correct it.”
Tony Thomas, Co-founder, Zoho Corporation
Q: What is ManageEngine’s approach to right-sizing AI models for specific workloads?
ManageEngine has invested in building its own series of models, progressively scaling them in size and capability while optimizing each for specific workload types. The strategy involves deploying these models internally first, then opening them to customers as a cost-effective alternative to frontier API calls. Open-weight models are also made available to customers as part of the same cost-reduction strategy. The core principle is matching model capability to task requirements rather than applying the most powerful available model uniformly.
“We’ve been putting a lot of effort into smaller size models. We built our own series of models and we are moving up in terms of both size and capability of those models and optimizing those models for specific workloads.”
Tony Thomas, Co-founder, Zoho Corporation
Q: How do guardrails and harnesses improve AI output quality in enterprise deployments?
Applying guardrails and harnesses to focused, smaller models constrains the model’s output space to the domain it is optimized for, reducing hallucination rates and off-topic responses. Thomas frames this as a compounding benefit: a focused model with guardrails produces fewer errors, which means fewer correction loops, which means lower token spend and higher output reliability. The combination of model right-sizing and guardrail application is presented as the practical path to maintaining quality while reducing cost.
“On top of that you can build guardrails, harnesses, proper. So what is ManageEngine’s approach? Small models are very focused.”
Tony Thomas, Co-founder, Zoho Corporation
Resources & Documentation
- ManageEngine, IT management platform by Zoho Corporation with integrated AI capabilities
- Zoho Corporation, enterprise software suite including Zia AI across Zoho and ManageEngine products
***
👇 Click to Read Full Raw Transcript
Swapnil Bhartiya: And since you opened the Pandora box of AI. So I would. And also when I was looking at this year’s conference, actually I was at Splunk conference just two weeks ago. I mean, every discussion is about AI. 99.9% of my interviews, no matter who I talk to, is about AI. How is AI reshaping Manage Engine or Zoho? Because if you look at the market since you started, market has evolved new technologies. And it’s less about technology, it’s about where users are. You have to be where users are, right? Otherwise you’ll be left behind, as we have seen with a lot of technologies. So talk about how is AI reshaping your whole architect, the foundation of Zoho as well as Manage Engine.
Tony Thomas: It’s a good question, I think. When AI first came out, certainly the personal use of AI was huge, and it’s still growing. When it comes to using AI in the enterprise, the first glimpses of AI came as feature products. So AI is improving some part of the product and it gets built into the product. But the real value of AI getting unlocked in business is based on improving outcomes, solving bigger problems more efficiently and faster, and providing a better business service. So in that context, doing more with AI comes down to impacting how business is done, how problems are solved within enterprises, and that’s a real value of AI. So for Zoho and ManagedEngine, the key changes that are happening now, are we engaging much more with the market, looking at what use cases and outcomes customers want to solve and then looking to implement that for customers and then taking that to service providers and others to extend the reach of what we can do.
Swapnil Bhartiya: AI has done two things from my perspective, and I’m heavy user of AI. I mean, I build my own servers. I mean ton of things. I’m not a developer, I’m a writer, journalist, a filmmaker. But I’m wipe coding. No offenses to those who don’t like wipe coding, but I have done and it’s secure. What it has changed is, first of all, initially it looked magical, right? It can do things. Because AI has been around for so long, LLMs have changed things. So it’s more like you can throw whatever it is, it will do it. Second is what we started seeing in the enterprise space. People used to boast about how many tokens they have consumed. Nvidia would talk about if you’re not consuming that many tokens, which are equal to your salary, you’re not using. I don’t think that boasting on is more like saying, hey, My car consumes, you know, that many gallons of gas to go one mile. Yes, you can get, you know, all those. But that is also changing because token cost is becoming a big tokenomics is becoming a big topic. So I think it is flipping. It should be less about how much tokens you are burning. It should be more about how much value you are getting by spending less tokens. So how do you look at where things are heading? And because this is not sustainable, I’m not even talking about data center. I’ve got a lot of fudder around data center as well. A farm consumes more water than a data center, depends on how you look at it. But GPU availability. So this is not sustainable the way we are heading. So we have to change something. Now we cannot change a lot of things in hardware, but we can change how we consume. So how from your perspective, what organizations are doing or should do or how manage engine is helping them because cost should not mean drop the quality. That’s what happens. Right. So what is your take on that?
Tony Thomas: No, it’s actually a great question because affordability has always been at the center of our strategy. Zoho looks at sustainability, being able to provide value to customers without breaking the bank. That has always been the central ethos for Zoho. And so looking ahead at what AI needs to do, certainly wipe coding is helping with certain things. But if you look at all of our managed engine and Zoho products, security is essential. We can’t wipe code these products. You have to from compliance to being able to ensure that the infrastructure is secure, that trumps everything else in terms of priority. So building on a solid foundation, making sure it is secure and allowing just the parts that can be made, wipe coated and made much faster to develop. That’s how we look at AI and in terms of cost, using the right size model for the right job is part of what we are pushing very hard on. We are building our own internal models to optimize various parts of our workload. And you don’t need to use the frontier models for everything. And so making that a central piece of our solution is part of it.
Swapnil Bhartiya: You also hit on the Neil, everybody is obsessed with the bigger model. I feel the model, I mean I am heavy user myself. That model should fit on your general purpose laptop, you know. And second thing is that if you walk to Aristotle and ask about, you know, a small how to take turn how to reach to the exit, he will tell you about it. He will not. So that is with the models also they hallucinate and they will take a lot of side quest and they will not answer your question and you’ll burn all the tokens. Another problem with the AI is if AI makes mistake, you have to pay to fix the mistake because you have to waste more token to correct it. So it is also important sometimes small models are very focused and then also on top of that you can build guardrails, harnesses proper. So what is manage engine’s approach? I mean you folks, I think use Zia. Right, Zia. So what is your approach to bring the cost down to have guardrails harness focus and also the quality should not be compromised.
Tony Thomas: So we’ve been putting a lot of effort into smaller size models. We built our own series of models and we are moving up in terms of the both size and capability of those models and optimizing those models for specific workloads. So we look at taking our technology, deploying it for our own use, and then slowly opening it up to customers to come in and use it as an alternative to the more expensive model. We’re also taking the open weight models and making them available to customers. And so we are definitely seeing that as an important piece. The right sizing the model for the job is a very important piece going forward.
