Last Updated: 5 May 2026
Most enterprise support teams adopting AI asked the same question: which tool should we add to the existing stack? Vishal Sharma asked something different. As CTO of SearchUnify, he has spent over a decade working on enterprise search and CRM, and has written and spoken extensively about why AI deployments in customer support so often fall short of their promise.
SearchUnify's response to that problem is its Agentic AI Suite, which the company describes as a network of purpose-built autonomous agents that handle discovery, triage, resolution and assistance in coordination, powered by the organisation's own data.
Sharma is a frequent voice on topics including LLMs, knowledge management and responsible AI deployment in enterprise environments. He was quoted in a recent TechTarget piece arguing that AI backlash inside organisations is rarely a technology problem. It is a leadership one. In this interview, he talks about how autonomous agents actually work at enterprise scale, where compliance and governance fit into that picture, and what it genuinely takes to move beyond a pilot.
1. You have spent over a decade working in enterprise search and CRM. How has the job of helping someone find information actually changed now that AI agents can act on that information, not just retrieve it?
For most of the time I've been in this space, the core question was: can we surface the right document? That was a real problem worth solving. But retrieval was never the end goal. Resolution was. A user opening a support ticket doesn't need a link to an article. They need the problem to be handled.
What changes fundamentally when agents can act on information is that the architecture of a support workflow shifts. An agent can now seamlessly exchange data, coordinate through distributed orchestration, and execute complex, multi-step tasks both individually and collectively. My background in CRM taught me that most of the friction in support wasn't missing information. It was the gap between having the information and doing something useful with it. That gap is what we're now closing.
2. SearchUnify describes its Agentic AI Suite as a kind of central nervous system for enterprise support. That is a bold claim. What does an AI platform actually have to do, technically and operationally, to earn that description?
The central nervous system framing is intentional: the platform seamlessly automates discovery, resolution, triage, and assistance, with purpose-built autonomous AI agents, co-pilots, and self-service, powered by your data. That breadth is the point. A single-use tool, a deflection bot, a summarizer, cannot be a central nervous system. Something earns that label when it coordinates across functions in real time, with shared context.
What makes this work is orchestration. The core need is the capability to achieve end-to-end execution of tasks within enterprise business functions, where agents dynamically negotiate.
3. A lot of companies say they help enterprises get real value from AI. Most of them are still running pilots. Where do enterprises genuinely get stuck going from experimentation to something that works at scale, and what does SearchUnify do differently?
The pilot problem is, more often than not, a sequencing problem. Enterprises stand up a capable model, measure it on accuracy benchmarks, declare it promising, and then discover that production is a completely different environment, with messy data, contradictory sources, unclear ownership, and no governance in place. By the time those issues surface, the project has already been sold internally as a success.
What needs to be prioritized is making the unglamorous infrastructure, connectors, permissions, audit trails, escalation controls, production-ready from day one rather than retrofitted once the pilot ends.
4. In a piece you wrote for G2, you introduced the concept of FRAG, federated retrieval augmented generation, as a way to go beyond standard RAG. Standard RAG pulls information from a knowledge base to improve AI responses. What does FRAG add to that, and why does the difference matter when an enterprise agent needs to find the right answer fast?
Standard RAG is a meaningful step forward. You're grounding a model's responses in your actual content rather than relying on whatever it learned during training. But RAG typically operates against a single index. For most large enterprises, that's not how knowledge lives.
While prevalent approaches like RAG have challenges including context length limitation, limited contextual understanding, and dataset bias, FRAG adds an extra layer. When integrated with your knowledge management engine, it retrieves best-fit snippets from multiple repositories, then passes them to the LLM to provide context-aware, personalized, and factually correct results. FRAG combines three key layers: Federation, Retrieval, and Augmented Generation, working in tandem to optimize performance and generate more contextually appropriate responses. In a fast-moving support interaction, the difference between "one index" and "all relevant sources" can be the difference between a resolved case and an escalation.
5. Most large companies have their knowledge scattered across dozens of systems. Sales tools, ticketing platforms, internal documentation, community forums. When an AI agent needs to find an answer, how does SearchUnify decide where to look first, and what determines the order?
Prioritization isn’t static, it’s driven by a combination of user context, query intent, and historical signal. The platform analyzes user and historical data to understand intent and personalize the experience accordingly, ensuring accurate results without superfluous distractions. The agent isn't performing a brute-force sweep across every connected source. It's ranking sources by relevance to the specific context before it even starts retrieving.
Connectors unify enterprise applications by integrating various platforms and systems into a single search platform, providing quick, role-based access to relevant information. Native connectors respect role-based and access-based permissions, ensuring users only see content they are authorized to access. The question of where to look is always happening in parallel with the question of what this user is permitted to see. Those two things can't be separated cleanly in enterprise environments, and any architecture that treats them as independent creates problems.
6. Enterprise knowledge bases are rarely clean. Information goes stale, different documents contradict each other, quality varies enormously. What happens when a SearchUnify agent hits a knowledge gap or a conflict? How does it decide what to do next?
This is one of the more honest problems in knowledge management, and worth not glossing over. Our position is that a static knowledge base is always going to degrade. The fix isn't a better one-time cleanup project. It's treating knowledge maintenance as an ongoing, automated function.
The AI Knowledge Agent smartly tracks and detects knowledge gaps, identifies creation or revision opportunities, and keeps the knowledge base updated. It automatically creates knowledge articles using templates and contextual data from support tickets, chat logs, or community forums, and performs quality checks using the KCS content standard checklist. When an agent encounters a gap or a conflict in the moment, it flags it, routes to a human when appropriate, and feeds that signal back into the knowledge lifecycle. The goal is for every unresolved gap to become an input into fixing the gap, not just a dropped ticket.
7. Showing users where an AI-generated answer actually came from, through source citations, is increasingly seen as a trust mechanism in enterprise AI. What makes a citation genuinely useful in that context, as opposed to just a technical feature that looks good on paper?
A citation that points to a document a user can't access, or to a source that's six versions out of date, undermines trust. The bar for a citation being genuinely useful is higher than most implementations acknowledge.
Advanced reporting and citation tracking ensure continuous improvement and data-backed decision-making. A citation must be actionable. The user needs to be able to trace the reasoning, verify the source, and understand why the system surfaced it. When citations serve that function, they're part of how the system earns ongoing trust rather than simply requesting it upfront.
8. SearchUnify emphasises domain specialisation, meaning its agents are built to understand the specific language, products and workflows of a given industry. How do you train an agent to be specific enough to be useful without making it so narrow that it breaks the moment something changes?
The answer is in the architecture more than the training. If an agent's domain knowledge is baked into the model itself, every product change or workflow update requires a retraining cycle. That's brittle and expensive.
Our approach keeps domain specificity in the retrieval layer, the knowledge base, the connectors, the indexed content, rather than encoded rigidly in the model. Modular architecture allows enterprises to integrate new AI agents or refine workflows without disrupting existing operations. When something changes, a new product line, a revised policy, you update the knowledge rather than the model.
9. The Agentic AI Suite handles discovery, triage, resolution and assistance at the same time, with different agents working in parallel. How do you prevent those agents from producing conflicting outputs or duplicating effort, and what does that coordination actually look like in practice?
Coordination is the hard part of multi-agent systems that vendor demos tend to skip. Shared context and clear ownership are vital. Every agent in the suite needs to operate from the same understanding of the case at any given moment, and there needs to be unambiguous logic for which agent owns which decision. Agents seamlessly exchange data and coordinate through distributed orchestration, dynamically negotiating via MCP protocols to enable complex, multi-step execution both individually and collectively. Beyond the protocol layer, each agent in the suite has a defined scope. Preventing conflicts starts with clear functional separation at the design level, not just runtime coordination.
10. Autonomous AI sounds appealing until something goes wrong with a real customer. Where does SearchUnify draw the line between letting an agent resolve a case on its own and handing it to a human, and who ultimately gets to set that boundary?
The line should be drawn by the customer, not by us, and not by the model. Our role is to make those escalation controls precise and configurable, not to prescribe where the threshold sits.
For complex or unresolved issues, the AI Support Agent initiates smart handoffs to human agents with full context, ensuring seamless and fast case resolution. "Full context" is the key phrase—a handoff without context is almost worse than having no handoff at all.
But ultimately, agentic AI is not about replacing people. It should elevate them. The handoff decision is one of the clearest places that principle shows up in practice.
11. Sentiment analysis in support workflows, where AI reads emotional signals in a customer interaction to influence routing and prioritisation, is one of the more sensitive applications of enterprise AI. How do you build that responsibly? Inferring how a person feels based on their words is a significant step for any automated system.
Reading emotional signals from text and acting on them autonomously is a significant capability, and the design choices around how it's used matter considerably.
Sentiment analysis is a routing and prioritization signal. It raises flags, surfaces urgency, and routes cases faster. Our AI Agent Partner summarizes history, sentiment, and timeline so agents resolve faster with complete clarity, but the human agent is always the decision-maker for emotionally charged cases. AI should equip them to respond, not substitute their judgement.
12. Many consumers still prefer speaking to a human over interacting with a chatbot. What is the honest ceiling for AI self-service, and are there situations where a handoff to a human should always happen?
There are case types, anything requiring negotiation, genuine empathy, or nuanced judgment under ambiguity, where a capable human will outperform any agent we can build today. Only a narrow set of tasks can currently be executed entirely by AI agents, including order tracking, password resets, and knowledge base lookups. For anything involving multiple steps, judgment, or empathy, the agent usually hands off to a human.
What AI self-service does well, it does very well: speed, consistency, availability, and scale across routine interactions. The boundary between what AI should own and what humans should own will keep moving as both the technology and our understanding of appropriate use cases mature.
13. SearchUnify has positioned agentic AI for highly regulated sectors such as financial services, with features like real-time data masking, audit trails and escalation controls. Being certified for compliance and being genuinely compliant in practice are two different things. What is the difference, and how does SearchUnify close that gap?
Certification answers the question: does this system have the right controls? Genuine compliance answers a different question: are those controls actually working in every interaction, at scale, in production? Auditors check the former. The latter only surfaces when something goes wrong.
SearchUnify holds SOC 2, ISO 27001, GDPR, and HIPAA certifications, with InfoSec review support available at every stage of enterprise procurement. The architectural decisions matter more than the certificates, though. PII and PCI data is masked in real time before any LLM interaction, with full alignment to regulatory requirements, and every agentic action generates a full, explainable audit trail.
The distinction I'd draw is between compliance as a posture and compliance as an operational discipline. The audit trail isn't there for the auditors. It's there so every decision the system makes can be reviewed, questioned, and corrected.
14. When an AI agent retrieves information to answer a question, it is potentially accessing content from across the entire enterprise. How does SearchUnify make sure an agent only sees, uses and cites information that the specific user in front of it is actually allowed to access?
This is a foundational design question, and the answer has to be architectural. You can't bolt permissions on after the fact and get a reliable system.
Native connectors respect role-based and access-based permissions to ensure that users only see content they are authorized to access, preventing any information leakage. The permissions enforcement happens at the retrieval layer, before content reaches the model. In practice, an agent responding to a support query will never surface, summarize, or cite a document that the requesting user doesn't have permission to see in the source system. The access logic of the original system is respected throughout.
15. Guardrails is a term that gets used a lot in enterprise AI. What do SearchUnify's guardrails actually do in practice, and how much say do customers have over things like how cautious the model is, how it communicates, and where it refuses to act autonomously?
"Guardrails" has become one of those terms that can signal safety-consciousness without necessarily describing anything specific. What I'd point to in our case is the configuration depth. The platform fosters user confidence with built-in governance and guardrails that steer AI ethically, which covers controlling the model's tone and communication style, setting escalation thresholds, defining where the system will and won't act autonomously, and configuring how cautious the agent is when operating near the edge of its confidence range. Customers set those parameters, and the platform continuously refines system performance through built-in audit trails and human-in-the-loop feedback mechanisms. The guardrails are also inputs into an ongoing feedback loop that adjusts system behavior based on what's actually happening in production, so they're not static configuration.
16. A poorly maintained knowledge base, a fragmented tech stack, no clear data governance. For an enterprise in that position, is deploying autonomous AI agents a risk rather than a solution? And would you tell them to wait?
I'd tell them to be specific about what they're deploying before making that call. A company in that position shouldn't deploy a fully autonomous agent suite that makes consequential decisions without human review. Saying "wait until everything is clean" is also a mistake, because the knowledge base is never going to get clean without pressure from a system that depends on it being clean.
Data readiness is critical. Breaking data silos and unifying pipelines is imperative to enable agents to work. Operating on fragmented data will amplify errors, not reduce them. I'd advise starting with the agents that generate the most pressure to fix the data, knowledge gap detection, content quality auditing, rather than starting with the agents that depend on the data being reliable. Use AI to expose the problem before you use it to solve the problem.
17. There is a strong case that governance frameworks need to be in place before AI deployment, not bolted on after the fact. How many enterprises actually do that, and what tends to go wrong for the ones that do not?
Governance before deployment requires a level of organizational discipline that's genuinely difficult to sustain when there's executive pressure to ship fast and show results. The AI initiative tends to move ahead of the governance team's ability to evaluate it.
Governance should always be defined before forward movement, including clear expectations for oversight, review processes, and acceptable use. Mechanisms introduced after an incident feel reactive rather than intentional — which isn’t what we want.
What goes wrong for organizations that skip this isn't usually a catastrophic failure. It's a slow erosion of trust: an unexplained decision here, a bias complaint there, an audit that can't be answered cleanly.
18. In a recent TechTarget article, you argued that AI backlash is often less about model performance and more about how AI changes work, accountability and trust. What should leaders actually address before rolling out AI agents into support workflows?
In this article, I stated that people are reacting to how AI changes their work and who’s ultimately accountable if things go wrong. Before rolling our agents, I advise leadership teams prioritize accountability clarity and maintain transparency about how the system is evolving. AI doesn’t create mistrust on its own.
Support teams are watching closely to see whether their expertise is being replaced or extended. That question deserves a direct, honest answer before the system goes live, not a communication memo sent the week after deployment.
19. SearchUnify is built around delivering measurable business outcomes. Which metrics actually tell you whether enterprise AI is working: resolution quality, case deflection, knowledge health, agent productivity, customer sentiment, or something else entirely?
All of those metrics matter, but they don't carry equal weight at every stage. In early deployment, case deflection is the most legible signal. It tells you whether the system is handling what it was supposed to handle. Deflection alone is a trap, though, if you're not also measuring resolution quality. A deflected case that generates a callback or a second ticket is not a win.
The platform monitors agent performance, identifies knowledge gaps, tracks customer satisfaction trends, and optimizes workflows with comprehensive dashboards and real-time reporting. The metric I pay closest attention to, and the one most organizations underinvest in tracking, is knowledge health. Without accurate knowledge, the rest of the suite cannot function at scale. If the knowledge base is degrading, every other metric will eventually follow. It's the leading indicator that most teams don't watch until the lagging indicators start moving in the wrong direction.
20. There is no shortage of obstacles between an enterprise's first AI pilot and a system that runs reliably at scale. Data quality, orchestration, governance, user trust, proving return on investment. In your experience, which of those derails the most deployments, and what does fixing it actually require?
Data quality is the most common answer, but I'd push a level deeper. The root cause behind most data quality failures is governance: nobody with clear ownership of what "good" data means for this system, no ongoing process for maintaining it, no feedback loop between what the AI surfaces and what gets fixed. Governance and accountability require clearly defined escalation protocols and full audit trails for every agent action.
Fixing it requires making data quality someone's job, not everyone's aspiration. That usually means assigning ownership at the content level, instrumenting the system to surface degradation automatically, and creating a feedback mechanism where agent behavior generates knowledge improvement tasks.
