Enterprise workflows are outdated and breaking under pressure.
IT teams are drowning in recurring incidents. Support agents are buried in redundant tickets. Employees waste hours searching for information, only to hit dead ends. Automation promised efficiency but left enterprises tangled in rigid rule-based systems that crumble under real-world complexity.
But now, AI agents are rewriting the rules. Instead of waiting for problems to escalate, AI predicts and prevents them. Instead of forcing employees to sift through fragmented data, AI surfaces the right insights instantly. Instead of relying on static logic, AI workflows adapt dynamically.
This isn’t the distant future. The shift is already happening. Enterprises that don’t embrace Agentic AI will find themselves buried under inefficiencies, struggling to compete in a world where agility and automation define success.
So, how do enterprises deploy AI agents that don’t just assist—but fundamentally transform workflows?
That’s the question at the heart of this conversation between Dheeraj Pandey, Founder and CEO of DevRev, and Amit Prakash, Co-founder and CTO of ThoughtSpot. Both have spent decades building enterprise software that powers some of the world’s largest businesses. In this discussion, they lay out a no-nonsense roadmap for AI-driven workflows—what works, what doesn’t, and what comes next.
What AI agents really are—and what they aren’t
I don’t like the word ‘agentic’ too much. I feel like it’s functionality, just like the word ‘functionality’ never existed in the English dictionary. But then Silicon Valley added it—like, oh, we’ll use the word ‘functionality.’ I think ‘agentic’ is one such word.
The term “AI agents” has become a buzzword, but what does it actually mean? Is Silicon Valley just rebranding automation with a fancy new label?
According to Dheeraj, AI agents aren’t just about automation—they represent a new operating system for enterprise workflows. "If anything, we call it AgentOS because it’s a platform—our version of Windows,” he says, referring to DevRev’s AgentOS platform.
Adding to Dheeraj’s definition, Amit referred to the 2 systems of thinking as described by Nobel Prize laureate and psychologist Daniel Kahneman in his book Thinking, Fast and Slow.
- System 1 Thinking is quick, unconscious, and requires little to no effort, which allows us to make quick decisions and judgments based on patterns and experiences.
- System 2 Thinking, on the other hand, is slow, conscious, and requires intentional effort, helping us with complex problem-solving and analytical tasks.
AI agents operate within a dual-processing model, according to Amit, where AI handles rapid intuition (System 1) while deterministic algorithms execute deliberate, rules-based decisions (System 2).
Agents are this idea of LLMs provided us the System 1. And now, as an industry, collectively, we’re trying to figure out the design for System 2 so that we can get closer to the dream of AI.
How AI agents are the next big upgrade in enterprise workflows
Enterprise workflows have undergone a massive transformation over the past two decades. Dheeraj breaks down how enterprise workflows have moved through distinct phases over the past two decades:
- On-prem systems (2000s): Workflows were hard-coded into enterprise software, requiring extensive development effort to modify.
- Cloud-based automation (2010s): Companies like ServiceNow introduced cloud-based workflow automation, making processes more accessible.
- Low-code/no-code platforms (late 2010s–early 2020s): The rise of platforms like Zapier enabled non-technical users to build automated workflows using drag-and-drop and WYSIWYG (What You See Is What You Get) interfaces.
While low-code/no-code democratized automation, it still relied on static logic. If a situation fell outside predefined rules, humans had to step in.
AI agents remove this bottleneck by allowing workflows to evolve based on context and past interactions. Instead of executing a fixed sequence, workflows can now adjust dynamically depending on the complexity of the task.
At the core of AI-enhanced workflows are AI nodes—specific decision-making points where LLMs and machine learning models assist in classification, clustering, and automated routing. “Essentially, agentic means you’re doing multi-step things with LLMs, like you’re combining workflows and LLMs,” Amit highlights.
Dheeraj adds to this observation that AI agents introduce probabilistic reasoning into these workflows, allowing systems to handle ambiguity, infer intent, and adapt dynamically.
To make this concrete, let’s look at some use cases for enterprises where AI agents transform enterprise workflows and deliver real value:
- Enterprise search: Retrieves contextual answers dynamically instead of relying on keyword matches. Helps reduce internal search times substantially.
- Workflow orchestration: Automates multi-step approvals, ticket resolutions, and escalations. Can cut down manual interventions.
- Intelligent automation: AI-based classification, clustering, and deduplication streamline IT and support workflows. This helps enterprises eliminate redundant tickets and reduce incident backlogs.
Why enterprise search is broken—and how AI is fixing it
‘Who broke the build?’ It drives me nuts that people actually go ask this question on a Slack channel with 100 people because some of these things we have to use AI for. This is a simple workflow of saying ‘who broke the build’—it’s probably kept in a build system somewhere or a test system somewhere or a CI/CD system somewhere.
Every enterprise employee has experienced this frustration: an employee asks a question in Slack, gets no response, then asks again, and eventually, someone provides an answer that was already shared a dozen times before. The problem isn’t a lack of knowledge, but poor knowledge retrieval.
Slack and Microsoft Teams helped enterprises move away from endless email threads. But they introduced a new problem: information fragmentation, which poses three barriers for seamless collaboration among different teams:
- Knowledge is siloed in different channels. If an employee wasn’t part of a discussion, they won’t know it happened.
- Search is keyword-based, not intent-based. Employees must guess the right keywords to find past discussions.
- Repetitive questions waste time. The same issues are discussed over and over because there’s no intelligent surfacing of answers.
Slack and Teams were built for real-time communication, not long-term knowledge retention. Conversations are buried in hundreds of channels, and finding relevant context becomes a scavenger hunt. This is where vector databases are reshaping enterprise collaboration.
What is a vector database?
A vector database is a type of database designed to store and retrieve high-dimensional data, such as AI-generated embeddings. Unlike traditional databases, which retrieve data based on exact matches (structured queries, keyword search), vector databases store high-dimensional embeddings that capture meaning, context, and relationships between different pieces of information.
Instead of simply matching words, vector databases allow AI models to “understand” the meaning behind a query and retrieve the most relevant information, even if the exact phrasing differs.
This capability is critical for AI-powered knowledge management. Employees don’t always know the precise keywords to search for, but AI—leveraging vector search—can infer relevance based on semantic meaning, historical context, and real-time interactions.
Large language models (LLMs) are particularly effective at working with vector databases because they generate embeddings that represent complex ideas, which can then be indexed and retrieved in a way that mimics human intuition.
“Now you can collaborate directly with the LLM itself—with the vector database itself—as opposed to having to do this on a Slack channel,” Dheeraj explains. “We need people to interact and collaborate with the LLMs probably way more than we collaborate with 100 people in the channel because it’s a total waste of time and actually quite disruptive for these things to happen.”
This shift eliminates the need for employees to sift through channels manually. Instead, AI agents proactively retrieve and surface relevant conversations based on:
- Context: What the user is working on.
- Intent: What problem they are trying to solve.
- Similarity: Past discussions that closely match the current query.
The messy reality of integrating AI agents into enterprise workflows
The biggest challenge in my mind in the agent space is that there’s a wide spectrum of how conservative or aggressive you get. You can get very aggressive and claim things like autonomous software, like what Devin demoed and what some of the other people have done. But then, later on, you hear that it was a very pretty demo—not quite works when rubber meets the road.
Enterprise AI sounds like a dream—automating workflows, reducing manual effort, and scaling intelligence across an organization. At the same time, hype often overshadows functionality, as Amit refers to the hype surrounding Devin AI, purported to be the first AI software engineer, following its demo.
But in reality, AI adoption is hitting major roadblocks.
Unlike consumer AI applications, where a chatbot can retrieve a simple fact, enterprise AI needs deep contextual awareness. It must understand security permissions, user roles, and historical interactions—otherwise, it creates more problems than it solves.
If AI agents don’t get context right, they become useless. Imagine a financial analyst using an AI copilot—if it retrieves confidential data not meant for them, it creates a security risk.
At the same time, AI must segment knowledge correctly. For instance:
- An AI copilot assisting customer support must prioritize real-time query resolution.
- An AI assistant for IT must pull from technical documentation, not marketing materials.
- A sales-focused AI agent needs access to CRM data but shouldn’t surface internal financials.
Without intelligent segmentation, AI models return generic, irrelevant, or even dangerous results.
How do we establish intent? Is this a ‘how-to’ problem, or is it a ‘where-is’ problem? And these are two different things. The ‘how-to’ problem is something that’s sitting in some sort of a knowledge base and you need to do a really good job of search and deflection. Or if it’s a ‘where-is’ problem, then you’re really talking about a workflow.
Why Microsoft Copilot isn’t the success it seems
Given Microsoft’s reputation as an enterprise software giant, businesses eagerly rushed to deploy Microsoft Copilot. The promise was enticing: an intelligent assistant embedded within Microsoft 365, capable of generating documents, summarizing emails, and automating routine tasks with human-like efficiency.
But the reality has been underwhelming for many enterprises, Amit points out. The reasons?
- Lack of deep workflow integration: AI suggestions feel disconnected from enterprise processes.
- Context limitations: Copilot struggles with role-based access and personalization.
- High cost, low value: Companies pay millions for AI licenses, yet employees default back to manual workarounds.
Microsoft isn’t alone—the enterprise AI space is littered with lofty claims that fail at execution. What’s worse, Amit notes that enterprises come to the conclusion that AI isn’t for them because of having such underwhelming experiences with products like Copilot.
“It just annoys me to hell that people think this is a safe option and go to Microsoft and then end up with a crap product. And then they have this impression that AI doesn’t work because they tried it with Microsoft," he observes with exasperation.
Enterprises assume AI is plug-and-play. But without deep workflow integration, AI agents become just another tool in the already-bloated enterprise tech stack.
AI needs a supervisor: Why enterprises need humans in the loop
A lot of these conversational agents, when they have to hand it off to humans, you have to acknowledge that there will be a handoff. If you’re not, and you’re saying ‘no human in the loop at all,’ that’s where this industry will flatter to deceive. Being able to really hand it off to systems of record is going to be the opportunity and challenge.
One of the biggest misconceptions about AI agents is that they will fully replace human workflows. In reality, AI needs a structured framework for escalation and oversight.
Enterprises are rightly skeptical about AI making autonomous decisions in critical workflows. A misclassification in finance can lead to compliance violations. A wrong action in IT can shut down critical systems.
That’s why, Dheeraj emphasizes that AI-driven workflows must be designed for progressive autonomy:
- L1 AI Agents: Retrieve information, suggest solutions, and automate lookups.
- L2 AI Agents: Execute basic workflows (ticket resolutions, data extractions).
- L3 AI Agents: Handle multi-step processes but escalate complex cases.
- L4 AI Agents: Automate full-cycle workflows with human review checkpoints.
Another instance where AI agents require human interventions is setting performance benchmarks and stress-testing the AI agents to ensure its effectiveness. Vendors sell AI based on demos—but without real benchmarks, companies struggle to separate hype from reality. “I think it behooves us companies to actually build the best stress tests in this as well and put it in a way that’s one-click automated in a POC environment,” Dheeraj urges.
What should enterprises demand?
- AI stress testing: How does the AI perform under scale, load, and edge cases?
- Error rate benchmarks: What’s the AI’s accuracy in real-world tasks?
- User adoption tracking: Are employees actually using AI workflows or reverting to manual processes?
Ultimately, the companies that get AI right are those that structure their AI agents to work alongside humans—not replace them outright. At the moment, enterprises know they need AI, but they struggle to deploy it meaningfully. So, those enterprises that invest in practical execution of AI agents will define the future.
Extensibility is key to AI’s survival in complex enterprises
Anybody who actually comes out with an agent, a platform, but doesn’t let people build their own, will actually see a huge back pressure on their own developers and engineers. And then you become going from what you thought you’d be a product company to being more of a consulting and a services company.
The future of AI adoption hinges on extensibility—the ability to customize, scale, and integrate AI into existing enterprise workflows. Without extensibility, enterprises face one-size-fits-all AI solutions that fail to deliver true business value.
In other words, the most successful AI platforms won’t just deliver AI models—they’ll provide the tools for enterprises to build their own AI solutions.
Enterprise software has always followed a BlackBerry vs. iPhone trajectory—a battle between rigid control and open extensibility.
BlackBerry, once the gold standard for corporate mobile devices, prioritized security and control but failed to build an ecosystem that allowed for expansion and customization. Apple, on the other hand, provided an infinitely customizable platform by launching the App Store and enabling third-party developers to extend the iPhone’s functionality.
The result? BlackBerry faded into irrelevance, while the iPhone became the dominant device across both consumer and business markets.
The same pattern is now playing out in enterprise AI.
AI platforms that focus solely on control, security, and prebuilt functionality without extensibility will struggle to keep up. The real winners will be those that provide flexibility at multiple levels, enabling businesses to mold AI around their unique needs.
For AI platforms to succeed, they must offer extensibility in three critical ways:
- Customizable workflows: Enterprises have unique processes. AI must adapt to them, not the other way around.
- Seamless integrations: Open APIs and modular architectures should ensure that AI doesn’t sit in a silo, but becomes the connective tissue across departments and functions.
- Scalability: AI solutions must grow with the business, handling increasing complexity without creating bottlenecks.
Chatbots flattered to deceive. AI agents will deliver
Enterprise AI is at an inflection point. The technology is advancing rapidly, but its adoption hinges on one critical factor: user experience (UX).
The most sophisticated AI agent means nothing if users can’t interact with it effectively, trust its outputs, and integrate it into daily workflows. “The UI for this AI is a very important thing that we should not ignore,” Dheeraj urges.
There’s a fundamental flaw in how most enterprise AI is designed today: It’s built for engineers, not for business users. AI agents must adapt to human workflows rather than force users to adapt to complex AI interfaces.
Nowhere is this clearer than in the failure of chatbots.
For years, companies poured billions into chatbot development, expecting streamlined interactions and cost savings. But what was the result of chatbot adoption? Frustrated users who often ended up escalating their queries to human agents.
The reason lies in its design. Early chatbots were designed to be little more than glorified decision trees, functioning like interactive voice response (IVR) systems—press 1 for support, press 9 to speak to a human. They followed rigid, rule-based logic, responding with a predefined script for customer requests. But if the request deviated slightly? The chatbot failed.
Businesses quickly realized that rigid chatbots couldn’t handle real-world complexity. Users wanted bots to understand context, intent, and even sentiment—something rule-based systems simply couldn’t do.
“Bots are done, and they struggled for the last 10 years,” Dheeraj states. “They flattered to deceive.”
With chatbots proving inadequate, enterprises are now looking toward AI agents. Instead of rigid if-then-else rules, AI agents leverage:
- Natural language understanding (NLU) to grasp user intent beyond keyword matching.
- Retrieval-augmented generation (RAG) for smarter, real-time knowledge retrieval.
- Multi-step reasoning to handle complex, multi-turn conversations without escalating to humans.
Chatbots were static. AI agents are dynamic. The businesses that cling to outdated chatbot models will fall behind, while those that adopt AI-driven agents will redefine enterprise automation.
The chatbot era is over. The AI agent era is upon us.
Wrapping up: Enterprises can’t afford to ignore AI agents
Enterprise workflows are broken. Traditional automation is rigid, rule-based, and buckles under complexity. IT teams chase down recurring incidents. Support reps drown in duplicate tickets. Employees waste hours searching for the right information—only to come up empty-handed. Scaling operations without AI is a losing battle.
AI agents change the game. They don’t just automate tasks—they make workflows intelligent. Instead of waiting for problems to escalate, AI predicts and prevents them. Instead of forcing employees to sift through siloed data, AI surfaces the right insights instantly. Instead of rigid rules, AI adapts in real-time, combining LLM-driven intuition with deterministic execution.
This isn’t a distant future. The shift is already happening. Enterprises that embrace AI-first workflows will build self-improving systems that evolve with their business. Those that don’t will struggle to keep up.
The question isn’t whether AI belongs in your workflows. It’s whether your business can afford to function without it.
The real breakthroughs in the world of AI—the bold ideas, the hard-won lessons, the frameworks that separate hype from execution—are happening in conversations like this. If you want to keep your edge in all things AI, dive deeper into the discussion on The Effortless Podcast Substack.
