Startups in the AI world gathered at the AI Agent Conference this week in New York are scrambling to carve out a niche to avoid being overshadowed by industry leaders.
“And it’s not even just startups, but for example what Claude has done to Figma and Canva,” Omer Trajman, founder of AskFora, told The New Stack during the conference, held at a hotel in Midtown Manhattan.
“Startups [founders] are trying to figure out, ‘Where can I innovate where I’m not going to get trampled on by one of the models.'”
Startups dodge model giants
AI is moving really quickly, he added: “Startups [founders] are trying to figure out, ‘Where can I innovate where I’m not going to get trampled on by one of the models.”
At around 3,000, this year’s conference is about ten times last year’s size, said Trajman, one of the organizers.
“We think the next wave of technology is going to feel different, so we’re building companies around roles.” Peter Day, General Partner of investment firm super{set} told The New Stack.
“We want to build technology that absorbs tasks from people.”
“We want to build technology that absorbs tasks from people. AI is going to know what their priorities are, know all their things to be done, and start removing these tasks rather than giving them more things to do,” he added.
“We’ve got a couple of companies based on that thesis. One is called Zig.ai, which works in the sales space. Everything from prospecting to meeting follow-ups, scanning badges at conferences, and following up with emails. Another company on that thesis is Kana, which is helping marketing core jobs to be done well.”
Enterprise AI barely adopted
AI agent adoption is still in the very early stages within the enterprise, said Jai Das, co-founder, President, and Partner at Sapphire Ventures, in his keynote. “I think that we are actually at zero or maybe at one [on a scale of ten] of actual adoption of Enterprise AI.”
“While the consumer market for agents will be dominated by a few companies, enterprises are more diverse, and will not be dominated by one or two companies,” Das added.
Some of the companies started over the past four years are “AI native,” he said. “They are built just differently. Look at one of my companies in the defense industry that got sold for $4B. They had basically four engineers.” And they did everything with AI, Das told the audience.
“But when you look at some of the earlier companies coming from the SAAS world, they have a lot more engineers,” he added, and a different cost and pricing structure.
SaaS adds agents
During their sessions, SaaS providers OutSystems, UiPath, and Workato discussed adding AI agents to their existing enterprise products.
AI agents extend the capabilities of their customers’ business process workflows by implementing non-deterministic tasks that complement their products’ deterministic capabilities.
Their existing platforms offer enterprise-level quality of service for the agents deployed on them, including security, governance, scalability, and reliability.
The agents leverage SaaS platform services, including integration, API management, governance, and data access.
Raghu Malpani, Chief Product and Technology Officer at UiPath, for example, said they recommend that their customers focus on the overall business process and define the orchestration to implement it. Fit agents into the process where it makes sense to include a non-deterministic step.
One of the biggest concerns about using AI agents in the enterprise is that they will cause a data breach or store incorrect data. Agentic access to enterprise production data is typically either prohibited or significantly restricted.
Safe data access for agents
“From a technical perspective, what we solve today for our customers is allowing an agent to touch production data,” Ciro Greco, Co-Founder and CEO at Bauplan Labs, told The New Stack.
Bauplan Labs is building data infrastructure, not agents, Greco said. “We create a Git-like experience for an agent to access your data. Our concept is that your agent should have the capability to read and safely modify your production data.”
To make this work, “the agent creates a branch of your data lake, which is a copy of your production data. The agent manipulates the copy of your production data on that branch,” he added.
Greco said that Bauplan Labs provides the infrastructure to safely merge changes back into the original copy. This supports the trial-and-error pattern often used to debug and validate agents, he added.
An overall theme at the conference is that AI is a sea change in the industry, as were the Internet/web revolution and cloud computing.
Consequently, every IT organization and software company has to rethink what it’s doing in light of the revolutionary capabilities AI offers. Some are deciding to go the “AI native” route and rethink how they do things. Others integrate AI or add AI to accelerate what they are already doing.
“AI is not something you adopt. It’s something you implement… It’s not just, I just turn on the switch, and it’s ready to go.”
Conference organizer Ben Lorica, Principal at Gradient Flow, summed it up this way for The New Stack: “AI is not something you adopt. It’s something you implement. In other words, there’s work to be done in implementation. It’s not just, I just turn on the switch, and it’s ready to go.”
The post AI startups are scrambling to survive in big tech’s shadow