As the case shows, rolling out an AI sales agent comes with a few concerns: hallucinations, lack of model control, data-privacy fears from enterprise clients, and potential damage to trusted human relationships. Internal teams (especially sales and customer service) may also worry about job loss, eroding morale.
But, waiting passively comes with its own risks. Competitors who adopt AI early can rapidly scale service, personalize interactions, and improve efficiency. Delaying too long may leave companies scrambling to catch up.
PulsePoint should first (and immediately) adopt AI internally for tooling and as a co-pilot for human employees, while being cognizant of LLM limitations especially for its own specific use cases. This internal adoption will give time for PulsePpoint to figure out better integration of AI with existing workflows, as well as model performance against custom benchmarks, before rolling out external agents for sales.
_____________
What should leaders consider when deciding?
1. Start with the strategy, not the technology.
What core business problem are you solving? Profitability? Customer experience? Sales efficiency? In the case, PulsePoint’s real need is margin improvement, not simply “innovating with AI.”
2. Consider internal vs. external deployment.
Consider deploying agents for internal tooling first: AI-assisted sales prospecting, automated proposal and collateral drafting, copilots for human employees. This will boost productivity and build internal trust/morale while avoiding immediate client backlash.
3. Manage customer trust and expectations.
Clients like Orion may reject AI interactions outright. Hybrid or opt-in pathways protect relationships while still advancing innovation.
4. Understand LLM limitations
Today’s agents are bottlenecked by memory and personalization. They consume millions of tokens per user, but can’t learn from or remember past interactions. Models degrade over long sessions, lose critical info between conversations, and stay generic despite long-term usage.
Meanwhile, we are increasingly context hungry. As usage trends more and more towards deep-research, our hunger for context — and this memory gap — will only grow. 2.5 years ago, the standard context window for most large language models was only 4096 tokens (around 3,000 words). Today, the ceiling is 1 million (albeit with significant context rot).
Will today’s agents have the memory and personalization needs that PulsePoint needs to sound truly human? Will voice agents be able to cross the uncanny valley? Rather than giving into the hype, PulsePoint should carefully evaluate model performance for their needs, perhaps even developing their own internal evals and benchmarks for their specific use cases.
