Neo Lee is co-founder of Imagine AI.
Companies are increasingly using AI agents to interact with each other. It’s happening across all functions and roles—marketing, partnerships, even sales—with agents drafting everything from emails and proposals to pitch decks. It’s effectively AI talking to AI, with people involved only at the end of the process.
As CTO of an AI company whose agents plan and produce work on behalf of leadership teams at dozens of B2B companies, I’ve seen a significant shift in Silicon Valley recently. Even last summer, maybe 40% of the interaction was agents; now it’s 95% agents and 5% human. Right now, the main thing keeping humans in control is that a person still has to click “approve” before anything is finalized.
That gate is the last human step in the loop, and its days are numbered. The writing is already on the wall, with new technical standards like Google’s Agent Payments Protocol, now being developed through the FIDO Alliance, giving agents limited permission to act on a company’s behalf—up to and including spending money—without a person having to approve each transaction.
At this rate, it’s easy to envision reaching a point where the 5% human part shrinks even more. That would usher in a tipping point, what I call a “closed-loop production web.” Agents from different companies would start, negotiate and finalize work with each other directly across company boundaries with no person in between. The human role would then shift from gatekeeper to overseer—from approving each deal to setting the rules beforehand and reviewing what happened afterward. Researchers have warned that this kind of handoff can gradually eliminate human influence over the systems we depend on, which is why the rules we set beforehand are so crucial.
Set Limits First
Even though this transition will likely take a few years to reach equilibrium, many leaders are not prepared for it. The first thing CTOs should do is accept that it is inevitable and begin a structured, phased transition to minimize turmoil. I’m seeing this already in Silicon Valley, where some of the more fast-moving startups are in the experimenting and evaluation phases.
Be aware that agents communicate with each other quite differently from how humans do. A recent study testing LLM agents in negotiation and transaction settings found that behavioral anomalies between agents led to losses on both sides, including overspending or accepting unreasonable deals. Different agents also achieved significantly different outcomes for their users, meaning that which agent you deploy matters. Another study of more than 180,000 agent negotiations found that, even though some human negotiation principles still held, dynamics that were unique to AI also emerged, including using prompt injection as a negotiating tactic. Dominant agents were better at claiming value.
That’s why scoped mandates for agents are critical. Companies should set strict limits and rules for what agents can and cannot do in negotiations, especially during the training phases. Agents don’t start with the knowledge a person accumulates from years of experience, so codifying this institutional knowledge for training purposes would help give agents the same context as seasoned employees.
During the experimentation period of phasing out humans, it’s important to track when and how often a person has to intervene in an agent’s processes. Override logging is a good readiness metric for evaluating how effective agents actually are at automating cross-boundary decisions, with the long-term goal being to drive the override rate as low as possible.
Leaders should envision a world where the human bottleneck at cross-company boundaries is lifted and imagine how much faster and more efficient their organizations could be. I believe organizations could move five to 10 times faster once this friction is eliminated, both internally and in interactions with other businesses. The scoped mandates and override logs you build during the transition will become the rules your agents operate under once the human gate is gone, so start preparing now instead of scrambling after the shift arrives.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?