From Okta to Homeward: What Two Years Building a Failed AI Assistant Taught About Enterprise Needs

From Okta to Homeward: What Two Years Building a Failed AI Assistant Taught About Enterprise Needs

The most valuable product education often comes from what doesn’t work. Amar Kendale spent two years at Okta building an AI assistant with a talented team, significant resources, and clear enterprise needs to solve. The result? “I spent two years at Okta building an AI assistant, and it couldn’t even book a conference room,” he recalls.

In a recent episode of Category Visionaries, Amar explained how that failure became the foundation for Homeward‘s architecture, positioning, and go-to-market strategy. The lessons weren’t about doing AI assistants better—they were about recognizing that AI assistants solve the wrong problem entirely. Enterprises don’t need help finding information. They need AI that completes actual work.

The Conference Room Revelation

On the surface, booking a conference room seems trivial. Check availability, reserve a time slot, send calendar invites. Any competent software engineer could build this in an afternoon. So why couldn’t an AI assistant with two years of development handle it?

The answer reveals everything wrong with the traditional approach to enterprise AI. AI assistants are built to retrieve and present information, not to take action. They can tell you which conference rooms are available. They can even suggest optimal times based on participant calendars. But actually booking the room—writing to the calendar system, sending invites, confirming attendance—requires a completely different architecture.

This limitation wasn’t unique to conference room booking. It manifested across every workflow where the AI needed to do something rather than simply report information. The assistant could identify which support tickets needed attention but couldn’t resolve them. It could analyze sales data but couldn’t update the CRM. It could flag compliance issues but couldn’t remediate them.

The pattern was consistent: the AI could help humans work faster, but it couldn’t replace the work itself. For enterprises drowning in repetitive workflows, this was the wrong value proposition. They didn’t need better information access—they needed automated task completion.

The Architecture Lesson: Authentication Comes First

Working at Okta—a company literally built around identity and access management—gave Amar a crucial insight about what enterprise AI actually requires. The barrier to AI completing work isn’t intelligence or natural language understanding. It’s system access.

“We have built deep integration with authentication systems. So with Okta and Ping and Azure AD, we’re able to essentially authenticate Homeward as if Homeward is a real employee,” Amar explains. This became the foundational architectural decision for Homeward, and it came directly from understanding Okta’s enterprise challenges.

Traditional enterprise AI treats authentication as an afterthought. Build the intelligence layer first, then figure out integrations later. This creates systems that can read from various sources but struggle to write to them. They can query databases but can’t update records. They can see calendar events but can’t create them.

Homeward inverted this priority. By solving authentication first and comprehensively, they created AI that could operate with the same permissions as a human employee. This wasn’t just about technical capability—it was about fundamentally changing what AI could accomplish in enterprise environments.

The Okta experience taught that enterprises already have sophisticated authentication infrastructure. They don’t need another integration pattern. They need AI that works with their existing identity management systems. This insight eliminated years of potential technical debt and enabled deployment speeds that competitors couldn’t match.

The Positioning Lesson: Employees, Not Assistants

The conference room failure also crystallized a positioning insight that would define Homeward’s entire go-to-market strategy. “We actually think of Homeward as an AI employee. It’s not a chatbot that you’re asking questions to, but it’s actually an employee that’s getting work done,” Amar notes.

This wasn’t marketing copy invented after building the product. It was a direct response to watching an AI assistant fail at basic work completion for two years. The assistant framing creates wrong expectations: something that helps you but doesn’t replace you. The employee framing creates different expectations: something that completes entire workflows autonomously.

At Okta, Amar watched customers react to an AI assistant with questions like “How accurate are the answers?” and “How quickly can it find information?” These are the wrong questions for enterprises drowning in operational work. The right questions are “What workflows can it handle end-to-end?” and “How reliably can it complete tasks without supervision?”

The employee positioning also changed how customers thought about ROI. An assistant that makes employees 10% more efficient delivers marginal productivity gains. An employee that automates entire workflows delivers labor replacement value. The economic case is completely different, as is the buying conversation and the urgency of adoption.

The Product Lesson: Accuracy Enables Autonomy

Building at Okta also taught Amar about the accuracy threshold required for enterprise AI. An assistant that gives wrong answers occasionally is annoying. An employee that takes wrong actions occasionally is unusable.

“We spent a lot of time improving accuracy. Our accuracy is actually so good right now that we have several customers where the entire workflow is automated,” Amar shares. This focus on accuracy wasn’t about perfectionism—it was about crossing the threshold where enterprises trust AI with fully autonomous operation.

The accuracy breakthrough came from understanding that context matters more than raw intelligence. “Homeward has access to a lot more data about you and your company compared to any other solution, and therefore Homeward can actually generate insights or complete work with much higher confidence,” Amar explains.

At Okta, the AI assistant had limited context—whatever data it was explicitly given access to. This meant its decisions were based on incomplete information, leading to lower confidence and more errors. Homeward’s architecture, with its comprehensive authentication, means the AI sees everything a human employee would see. More context enables more accurate decisions, which enables true automation.

The Deployment Lesson: Speed Matters More Than Features

Watching enterprise deployments at Okta revealed another crucial insight: the bottleneck isn’t usually functionality—it’s deployment friction. Customers wanted AI capabilities but couldn’t stomach months-long implementation projects.

“We probably on average can deploy an end to end workflow within a week,” Amar notes. This speed comes directly from architectural decisions informed by Okta’s challenges. By solving authentication comprehensively rather than building point integrations, Homeward eliminated the custom implementation work that typically extends deployments.

At Okta, each new AI capability required new integration work, security reviews, and configuration. This meant even simple features took months to deploy across customer environments. Homeward’s universal agent architecture means customers can deploy new workflows without waiting for engineering resources or going through repeated security reviews.

The deployment speed also enabled Homeward’s product-led growth strategy. When customers can go from signup to production in a week, traditional enterprise sales cycles become unnecessary. The product proves itself faster than any sales process could, making the generous free tier economically rational.

The GTM Lesson: Product Velocity Replaces Sales Headcount

Perhaps the most counterintuitive lesson from the Okta experience was about go-to-market strategy. “We actually have just one sales rep,” Amar reveals. This isn’t a constraint or a sign of early-stage resource limitations—it’s a deliberate strategy enabled by product decisions.

At Okta, Amar saw how much enterprise sales effort went into convincing prospects of value that couldn’t be quickly proven. Lengthy demos, proof-of-concept projects, executive presentations—all necessary because the product couldn’t demonstrate value rapidly enough to sell itself.

Homeward inverted this dynamic by making the product experience so compelling and so fast that sales involvement became optional. When prospects can deploy production workflows in a week and see measurable business impact immediately, the traditional sales cycle collapses. The product does the qualification, demonstration, and proof of value—sales only gets involved for expansion conversations with customers who’ve already proven they extract massive value.

The Expansion Lesson: Let Customers Innovate

The final lesson from Okta was about product flexibility. Enterprise AI built for specific use cases struggles when customer needs evolve or differ from assumptions. “A lot of our customers are actually deploying Homeward in ways that we haven’t even built anything for,” Amar explains.

At Okta, new customer use cases meant new development work, new integrations, and new deployment cycles. Homeward’s architecture allows customers to create novel workflows without engineering support. This transforms customers from passive users into active innovators, discovering applications faster than any product team could.

This customer-driven innovation only works with the right foundation. The comprehensive authentication means customers can connect any systems. The autonomous execution capability means they can automate complex workflows. The universal agent architecture means they’re not constrained by pre-built integrations or workflows.

The Compound Effect

Looking back, the two years at Okta weren’t wasted—they were essential education. Every limitation of the AI assistant informed a design decision for Homeward. The inability to take action led to authentication-first architecture. The assistant positioning led to employee framing. The accuracy gaps led to comprehensive data access. The deployment friction led to universal agents. The sales intensity led to product-led growth.

“We have about 400 AI agents deployed right now across all of our customers,” Amar shares. Four hundred autonomous AI employees handling real enterprise workflows, deployed in nine months, with one sales rep. None of this would be possible without the lessons learned from building something that couldn’t book a conference room.

The irony is perfect: spending two years building an AI assistant that failed at simple tasks taught exactly what enterprises need AI to do—and what architecture, positioning, and go-to-market strategy enables it. Sometimes the best way to learn what to build is to spend years building the wrong thing first.