AI agents are already useful for 24/7 support and repetitive work in areas such as healthcare, insurance, construction, and travel.
2
Multi-agent systems can make work easier to adopt by assigning narrow tasks to separate agents with different levels of accuracy and guardrails.
3
Falling token prices do not automatically create viable businesses because agentic systems can call models more often and raise the cost of each completed task.
Summary
This panel examines where AI agents are useful today and what investors worry about when backing companies built around them. Meera Clark describes agents that answer calls, schedule appointments, and handle administrative work outside normal business hours. Sandeep Bakshi focuses on changing customer expectations, especially when people are used to trying to reach a human. George Robson discusses agents that can assemble marketplace-like experiences and automate work in poorly digitized industries. The panel also considers why several agents may be better than one general assistant, how investors assess accuracy and customer validation, and why proprietary industry data matters. The business discussion is cautious. Token prices have fallen sharply, but agentic products may use more calls and more tokens per task. The panel expects pricing, distribution, data access, and the cost of mistakes to matter as much as model performance. It closes with predictions about agent-to-agent payments, new interaction formats, and personal shopping assistants.
Agents can extend service beyond normal business hours
Meera Clark is most excited by agents that provide support around the clock. Companies in the panel's portfolio use them to answer phone calls and schedule appointments in healthcare and property and casualty insurance. She frames this as support for customer-facing staff rather than a simple labor replacement story. Agents can handle repetitive questions and administrative tasks, while people spend more time understanding customer problems, building trust, and helping with decisions. She gives examples from dental offices and hardware stores, where always-on assistance could change the customer experience.
Customer behavior has to change before online agents become useful
Sandeep Bakshi says customer support agents are not new. When he books travel online, he often assumes the agent is limited and tries to find a way to reach a person, even late at night. The product challenge is to make customers believe that better prompting will get them better help. That requires more than improving the model. Companies also have to change a habit that has developed through years of frustrating automated support. The value of an agent depends partly on whether customers trust it enough to keep using it.
Agents can lower the cost of assembling marketplace-like services
George Robson connects agents with industries that have resisted digitization. Agents can request quotes from many providers, in multiple languages, and combine the responses into something resembling a marketplace experience. This could digitize supply chains that have remained heavily brokered. The opportunity is especially large where businesses are still poorly automated or where a service has been too expensive to coordinate manually. The panel treats this as a way to open markets that earlier marketplace companies could not reach.
Separate agents make complex automation easier to adopt
Meera Clark says multiple agents are useful because current models are imperfect and can benefit from specialization and tighter guardrails. There is also a human reason to divide the work. Companies may name agents and assign them distinct jobs, much as people think of different employees handling different tasks. In construction, one agent might answer questions about room specifications while another updates schedules. This lets a company introduce capabilities one at a time instead of handing users a general assistant with every tool and no clear starting point.
Sandeep Bakshi says investors have to ask how much inaccuracy a customer can tolerate. A system that is 90 percent accurate may be acceptable in one setting, while a lower error rate may be required in another. The relevant comparison is often the existing process, including its accuracy, cost, and consequences when something goes wrong. The panel does not treat model accuracy as a universal threshold. It has to be judged against the industry being served and the cost of an incorrect action.
Customer validation and proprietary data matter more than a compelling demo
George Robson looks for evidence that a product fits an existing process and its performance expectations. He advises speaking with customers, finding the innovation budgets that can support an early deployment, and getting validation quickly. Enterprise procurement provides stronger evidence than a design-partner conversation because it tests regulation, compliance, need, and willingness to pay. He also says companies should speak with people building foundational infrastructure, since each additional performance improvement can take much longer and cost much more. Sandeep adds that sectors such as healthcare and construction may offer room for improvement because their valuable data is not widely available for model training.
Lower token prices do not guarantee better unit economics
Meera Clark says many AI businesses are not yet viable at current cost structures. Sandeep Bakshi agrees that some companies depend on models becoming cheaper, although proprietary data and difficult-to-enter industries can give them room to grow. Paul describes a sharp fall in the cost per million tokens from GPT-3 DaVinci to GPT-4o, then explains that agentic systems make more model calls and use more tokens. In his example, the cost per answered question increased even as the token price fell. The panel therefore treats completed tasks, rather than token prices alone, as the relevant economic unit.
AI products may need outcome-based pricing and a clear route to distribution
George Robson argues that AI does not provide distribution by itself. A company still has to reach customers and compete with both legacy incumbents and newer software businesses. He sees an opportunity to sell outcomes rather than a system of record, with pricing tied to the value created by automation. Sandeep Bakshi says the comparison with incumbent services businesses can make an AI product attractive even when it carries model costs. Both investors focus on the customer's existing budget, buying habits, alternative solutions, and acquisition path instead of forcing every company into a fixed SaaS or infrastructure category.
"What I think is really interesting with these agents is this 24/7 support that they're able to offer."Meera Clark05:11
Who should watch
You are deciding whether an agent should automate a customer-facing or operational workflow and need to think through accuracy, escalation, and customer trust.
You are evaluating an AI startup whose costs depend on model calls and want a practical discussion of unit economics, distribution, and pricing.
You invest in or build enterprise AI products and want examples of the evidence investors seek beyond a working demo.
