7 lessons from the first wave of agentic AI deployment: theCUBE + NYSE Wired’s AI Agent Conference insights

The enterprise AI stack is getting smarter — but is it getting the context it needs to make agentic AI deployment actually work?

The defining problem of the agentic era might never have been building the agents. Instead, evidence is mounting that even the most capable AI systems stall without a clear strategic grounding and the proprietary organizational knowledge needed to support enterprise decisions — a problem that no frontier model solves on its own. The missing ingredient, it turns out, is not better technology but better context, according to Vanessa Liu, chair at Appen Ltd.

“Data is actually incredibly important for companies to be able to take advantage of AI. You need to train an employee when they come into an organization — even if they’re rock stars, you need to make sure you onboard them well,” she said. “Same thing when it comes to AI agents: You have to give them the business context so that they are going to be able to run well.” 

Liu and Steve Hasker, president and chief executive officer of Thomson Reuters Corp., spoke to theCUBE’s Gemma Allen at theCUBE + NYSE Wired: AI Agent Conference event, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. TheCUBE’s coverage featured chief executives, founders and product leaders from across data infrastructure, finance, enterprise modernization and open source AI, discussing the shift from experimenting with agents to becoming accountable for them.

Here are seven insights defining what it actually takes to put AI agents to work:

1. For agentic AI deployment, data is the proprietary moat.

Frontier AI models are only as effective as the business context they are given, and that context lives in human expertise that has rarely been systematically captured, Liu noted. As building agents on top of frontier models becomes more common, the companies that stand out will be those with a clear customer problem to solve and a defensible competitive moat, according to Hasker. For potential acquirers, the key question is not just whether an agent is useful, but whether it is meaningfully advantaged in the market.

Watch the complete interview on theCUBE.

2. Agents are impatient — just as the people using them are.

When a user asks a chatbot a question and sees “searching the web,” a mental clock starts ticking immediately and tolerance for delay has collapsed compared with even two years ago, according to Ariel Shulman, chief product officer at Bright Data Ltd. Bright Data’s scraped web data is now the starting point for chatbot responses, which means the company has had to deliver pages at much higher speeds, often in under one second, with a median response time of 500 milliseconds. That speed matters because the agent still has to turn the retrieved data into a useful answer before the user loses patience.

Hear more from theCUBE’s conversation.

3. Every agent touching money needs a bank account.

If AI agents are going to move money, they need the financial equivalent of identity, authorization and accountability. Catena Labs Inc. is building toward a “know your agent” model that would let banks verify which person or business an agent represents, what it is allowed to do and why it took a given action, according to Sean Neville (pictured), co-founder and chief executive officer of Catena Labs. The aim is not another closed banking platform, but a shared standards layer for agentic finance.

Explore the full conversation on theCUBE.

4. Token lock is the new vendor lock.

Enterprises that stake agentic AI deployment on a single frontier model are quietly surrendering leverage over their own cost structure, and the bill will come due as inference costs compound, explained Woodson Martin, CEO of OutSystems Inc. A platform layer that lets organizations hot swap models at runtime, without rebuilding the underlying systems, is no longer a nice-to-have — it is the only viable path to profit and loss control in an agentic AI deployment strategy.

Check out theCUBE’s complete interview.

5. Giving people AI tools and getting people to use AI tools are not the same.

Eighty percent of executives believe they are providing excellent AI tools to their workforce, while only a fraction of employees agree — a disconnect that sits at the heart of why so many enterprise AI investments are not delivering expected returns, said Tai Carmi, chief information officer of WalkMe Ltd. The answer is not more tools, but contextual nudges that surface AI capabilities inside the exact workflow moment where they are most useful.

Don’t miss the full sit-down on theCUBE.

6. Start with the most capable model, then swap in the cheapest one that matches it.

Builders who treat inference cost as the first constraint are making a strategic mistake. The right sequence is to unlock the fullest capability from frontier models first, then evaluate whether open-source alternatives can reach the same level at a fraction of the price, according to Qingyun Wu, founder and CEO of AG2ai Inc. That path is how serious enterprise deployments eventually escape the tokenomics trap without sacrificing the capability they built toward.

Catch the complete conversation on theCUBE.

7. Pilots are easy. Production is where agents go rogue.

The gap between a promising proof of concept and a trustworthy production agentic AI deployment is where most enterprise initiatives quietly break down, with agents drawing on stale data, skipping reasoning steps, blowing up token budgets or hallucinating outputs that no one caught in testing, according to Barr Moses, co-founder and CEO of Monte Carlo Data Inc. Courts have already ruled that the company behind an agent — not the user who triggered it — bears full accountability for what that agent does in the world.

Watch theCUBE’s exclusive interview.

Here’s the complete video playlist, part of SiliconANGLE’s and theCUBE’s coverage of theCUBE + NYSE Wired: AI Agent Conference event:

Photo: SiliconANGLE