When Shripati Acharya sat down with Umesh Sachdev, Co-founder and CEO of Uniphore, the brief was clear: go beyond the GenAI buzz and unpack what it actually takes to build a durable enterprise AI business.
Uniphore, “one of the world’s largest AI native enterprise SaaS companies,” was started in 2008, and today “powers multimodal AI across voice, video and text for 2000 plus global businesses.” Umesh has seen multiple hype cycles in AI, this conversation is about what’s working now, and what founders often underestimate.
From PoCs to “industrialising” AI
The turning point for Uniphore came from a simple but uncomfortable question large customers began asking: How do we industrialise AI inside the enterprise?
Umesh contrasts the early PoC era with what CIOs now expect:
“They were taking six to eight months to do a piece of PoC or a demo. But some of the more forward leaning enterprises were already thinking about industrialization.”
Industrialisation, as he defines it, means every business function being “self-sufficient to bring their own data, fine tune a model and then build these AI agents and put them to use instead of that headcount, design this AI agent for your task and put them to use.”
That shift led Uniphore to a key insight: in SaaS, vertical products created velocity; in AI, velocity will come from a horizontal platform that centralises choices around LLMs, security, guardrails and data access, while letting every function innovate on top.
A sovereign, open platform – not another foundation model
Uniphore’s answer is to position itself as “an end-to-end data and AI platform designed for large businesses.” The platform is deliberately opinionated about where it plays and where it does not.
Umesh is clear: “We are not a provider or builder of a foundational model. We are not a provider or builder of a data compute platform. We are not a provider or builder of a vector DB.”
Instead, Uniphore is “open to any choice that the enterprise may” make across those layers, while ensuring the platform can run “on any public or private cloud” and crucially on-prem. That sovereignty, openness and end-to-end coverage is what he believes enterprises now demand when AI starts touching regulated workflows, sensitive data and multiple business functions at once.
Why Uniphore is betting on small language models
While the broader ecosystem obsesses over bigger and bigger LLMs, Uniphore’s architecture is built around domain-specific small language models (SLMs).
“At Uniphore we are big fans and proponents of small language models,” Umesh says. “Most if not all business AI use cases that we come across are best served by focused and fine-tuned small language models as opposed to generic large language models.”
There are three reasons:
- Determinism and control – SLMs can be tightly constrained to one process or domain, making behaviour more predictable.
- Cost at scale – “The difference per query on a 7 to 8 billion parameter model versus a 80 or 100 billion parameter model is about 100 times per query of cost. Try putting this at company wide scale and production rollout. The cost itself won’t add up.”
- Enterprise-specific reality – every large company runs its processes differently. As he points out, even something as standard as CRM is unique in each company. “These small language models have to learn that company’s process, its own data and its own policy until actual usage starts to give back reinforcement learning.”
Uniphore has built what Umesh calls “a factory for domain specific fine tuned small language models,” then surrounds them with observability, explainability and guardrails.
Guardrails, unitary agents and the holy grail of determinism
If there’s one phrase Umesh returns to again and again, it is determinism. “Making these AI agents deterministic is really now the holy grail of business AI adoption,” he says.
Give a reasoning agent the same task ten times and “don’t be surprised, you get 10 different outcomes.” For regulated customers in financial services or healthcare, that’s unacceptable.
Uniphore’s “business AI cloud” is built as a four-layer architecture from data agents that “automate the task of discovering data and then preparing data for LLM usages” to a model garden with “very stringent guardrails evaluations, observability and explainability frameworks.”
He illustrates guardrails with a simple example: a model trained on purchase orders and job offers, while company policy says “there can be no conversation ever about gun or violence.” A guardrail system acts like a firewall: “The minute it sees the word gun it won’t even let the query hit the LLM.”
On top of this, Uniphore prefers “unitary agents” with singular, narrow tasks, then orchestrates them using BPMN-style workflows so “even though each agent reasons, the output becomes more deterministic, more manageable.”
A billion-dollar “rounding error” and what adoption looks like today
To ground all this theory, Umesh shares a live deployment with a Big Four consulting firm and a Fortune-10 oil and gas company.
Pre-GenAI, the consulting project contract leakage optimization was scoped for “a little over 1,000 consultants” manually reading contracts and invoices to find revenue leakage. The leakage itself “just happened to be little over a billion dollars, which in itself is a rounding error for the size of the company.”
With Uniphore, the Big Four fine-tuned an oil & gas contract-leakage SLM. Today, “on a weekly basis, about 35,000 invoices are read by these AI agents supervised by four human beings. And we’ve already seen about 50% over six months, 50% reduction in that contract leakage that the end client was receiving.”
Adoption, he says, is “not wide, but it’s very deep.” The expectation is that the next phase is about going enterprise-wide after proving outcomes in a handful of high-value use cases.
For Indian founders: stay close to the customer
When the conversation turns to Indian founders building in enterprise AI, Umesh is candid. The pace of innovation is “so violent” that “the whole concept of product management and philosophy is out of the window.” Weekly innovation cycles, constantly changing models and data platforms, this is the new baseline just to “get a licence to fight or play in the game.”
In that chaos, he returns to one anchor:
“When in doubt, I always go to my customers. So I would say customer intimacy in this phase is worth its weight in gold.”
He cautions against misusing AI where it doesn’t belong: “Not everything is worth the horsepower of LLMs and GPUs, I see some founders also make the mistake of using this very important bazooka to go after small flies.”
His view of Indian founders is generous but demanding: “Most of the Indian founders are very technically deep and when they’re not, they’re capable of very quickly filling those knowledge gaps for themselves. And therefore what’s absolutely necessary then is customer intimacy.”
Winning North America cannot be done at a distance
Umesh’s most hard-earned lesson is about geography and ambition. He recalls advice from his mentor, John Chambers, former CEO of Cisco:
“You cannot win that global game without winning North America as a singular market. It is the largest enterprise market.”
When Umesh suggested appointing a country leader to run the US, Chambers pushed back: “If you really want to win America, the CEO has to be here, the leadership team has to be here.”
Looking back, Umesh says, “Thankfully I listened to him and moved.” His advice to founders is unambiguous: “Whatever be your target market, you need to be in today’s world very intimate with that. If North America is a significant portion of the business plan, I don’t think this market can be run through a distance.”
For founders trying to build durable AI companies, Uniphore’s journey offers a clear message: industrialise, don’t just demo; prioritise SLMs and determinism over hype; stay fanatically close to customers; and if you’re going after the world’s largest enterprise market, be prepared to physically show up.
Timestamps:
00:00 Introduction
03:00 – What Uniphore Actually Does
05:47 – Why Enterprise AI Struggles to Scale Beyond Pilots
10:08 – The Determinism Problem
12:32 – Unitary Agents and Workflow Orchestration for Predictable AI
14:00 – Small Language Models vs Large LLMs for Enterprise Use Cases
15:25 – Why Guardrails and Governance Matter in Real Deployments
16:39 – One Big Agent Fails but Ten Small Agents Work Better
20:40 – Why AI Pilots Fail and Why That Is a Good Thing
22:20 – Converting Experiments into Enterprise-Scale Adoption
24:26 – 35,000 Invoices a Week with Only Four Humans: ROI Case Study
26:25 – What Enterprises Really Look for in AI Systems
28:45 – Lessons from Building Across India and the US
32:51 – Why Founders Must Be Close to Their Biggest Market
34:44 – Structuring Teams Across Geographies for Global Scale
38:10 – The Journey of Building Uniphore Over Seventeen Years
42:36 – Hard Lessons Learned Along the Way
46:45 – Why NVIDIA, Snowflake, AMD and Databricks Invested
51:17 – How the Strategic Round Came Together
52:22 – Closing Note
(Disclaimer: The views and opinions expressed in this article are those of the author and do not necessarily reflect the views of YourStory.)
Original Article
(Disclaimer – This post is auto-fetched from publicly available RSS feeds. Original source: Yourstory. All rights belong to the respective publisher.)
