Every enterprise is under pressure to drive revenue and profitability, and agentic AI shows tremendous potential to be a large driver of that growth. Most large enterprises we talk to have AI in production. As they scale it, their data and AI leaders face hard decisions that shape how fast they can move and what they build next.
In the last week of October, Venture Guides hosted a dinner in New York for more than a dozen Chief Data and AI Officers from some of the largest banks, insurers, and asset managers, along with the founders of Hoop.dev and Ray Security. We kept the guest list small to maintain an intimate setting where people could speak openly and built the evening around the two questions we hear most often from these leaders:
What type of LLM should we use? Frontier models have both the most advanced capabilities and the most drawbacks. Open-weight models can be run in an enterprise’s own environment and customized. This summer, more than 270 companies, including NVIDIA, Microsoft, and Google, signed the “Open Weights and American AI Leadership” letter asking U.S. policymakers to keep open-weight models available. Dinner guests discussed why they, like tech’s biggest names, want access to both: the right model depends on whether speed, quality, or control over data matters most.
How can we deploy agents safely? Boards want AI agents deployed. Employees are already building their own. Security and data teams now have to decide what these agents can access and how to review systems that move faster than any human. Should an AI agent perform the review or a person? And how does a small team keep up with thousands of agents?
Neither question has one right answer, which made for a lively exchange. Leaders shared what they had tried and pushed back on each other’s views. Here is what we took away:
Model choice depends on the task
Leaders have many model options, from frontier models to open-weight and small language models, but each comes with tradeoffs. The consensus from the evening was that the right choice depends on the task.
Most teams start with frontier models but then move some work to their own open-weight models. Cost rarely drives the move, since token spend is small compared with the engineering time required to run their own models. The drivers are performance and control over data: when a narrow task needs higher accuracy, when speed matters most, or when the data is too sensitive to send to model providers.
Many end up using both kinds of models in the same process, with a small, fine-tuned model sorting requests before a frontier model sees them. Running their own models also takes GPUs, which leaders expect to stay scarce for the next 12 to 24 months, so most expect frontier models to remain the default until more hardware is available.
Access rules built for humans don’t work for agents
Leaders agreed that giving agents enough access to be useful without creating new exposure is one of the hardest problems in enterprise AI adoption. Access rules were built for people who use judgment about what to open. An agent will open every file, application, and database its permissions allow.
The two portfolio companies at the table each solve a different part of this problem. Hoop.dev enables safe use of agents by controlling what an agent can do and access while it runs, hiding sensitive data and blocking changes that would break systems. Ray Security reduces the attack surface agents can reach, removing access to sensitive data that people and agents shouldn’t use and blocking unusual activity.
Human review doesn’t scale and agents cut corners
Keeping people in charge of agents came up again and again. Two problems stood out: reviewing agents breaks down at scale, and agents don’t always follow the rules.
Most firms keep a person in the loop for high-stakes agent decisions. The problem is that as agents multiply, reviewers see more than they can read and start approving everything. To solve that, some firms are moving their checks earlier. Leaders described a data use council that decides which data each AI project can use before it starts. The council slows projects down, but leaders in regulated industries see the tradeoff as worthwhile. The hard part is keeping those approvals current, since every new project can change the rules for the data, people, and agents involved.
Several leaders told stories of agents taking shortcuts. Agents finish tasks any way they can, even if that means deleting work or skipping steps they were told to follow. When an agent writes tests for its own code, the tests almost always pass, so they prove very little. The view in the room was that testing teams will matter more, this time to find where agents fail.
Leaders left with practical ideas on matching models to tasks, giving agents access safely, and deciding where human checks belong. If you are building in AI infrastructure, supply chain intelligence, or cybersecurity, or would like to join us at a future dinner, please reach out.