AI agents perform work without true understanding
The most dangerous lie in artificial intelligence is not the one it tells us, but the one it tells about itself. Salesforce’s recent push to transform Agentforce into billable digital labor presents a perfect case study in this deception — not because the company is being dishonest, but because the entire industry has built a vocabulary that conflates capability with competence, automation with understanding, and output with outcome. When a system can complete a task without comprehending it, we call that efficiency. When it fails in ways that look like success, we call that a mystery. The gap between these two realities is where the real story of AI unfolds, and it is widening with every new product launch.
The Performance of Productivity
Watch how an AI agent handles a customer complaint. It will acknowledge the issue, apologize appropriately, offer a resolution path, and close the interaction with professional warmth. Every measurable metric looks correct. The customer feels heard, the ticket closes, the dashboard shows green. But ask that same agent why it chose one resolution over another, and the answer reveals the truth: there is no “why” — only pattern matching against millions of similar exchanges. The system performs understanding rather than possessing it, and in a corporate environment where performance is rewarded, the distinction rarely matters until it catastrophically does.
Consider what Salesforce is actually selling with Agentforce. The company frames it as a shift from software seats to consumption-based digital labor, a model where businesses pay for tasks completed rather than tools licensed. The market reaction was muted, reflecting skepticism about whether businesses will embrace a model that charges per task rather than per license A human employee who makes a mistake carries accountability, context, and the ability to learn from failure. An AI agent that makes the same mistake simply repeats it more efficiently, at scale, with the same confident tone that made its correct answers so convincing.
The Historical Blind Spot
This is not the first time technology has promised more than it delivered. The 1980s expert systems boom sold corporations on the idea that digitized human expertise could replace expensive consultants. Those systems failed not because their logic was flawed, but because expertise turned out to be more than a set of rules — it included judgment about when rules should be broken. The current AI wave repeats this error with different vocabulary, replacing “knowledge base” with “training data” and “inference engine” with “agent architecture.” The underlying assumption remains untouched: that human capability can be reduced to process and sold by the unit.
The Salesforce strategy reveals how deeply this assumption has penetrated corporate thinking. By positioning Agentforce as billable labor, the company is not just selling software — it is selling a philosophical position about what work actually is. If a customer service interaction, a sales follow-up, or a data entry task can be priced per completion, then the implicit claim is that these activities have no irreducible human component. The performance of work becomes indistinguishable from work itself. This is where the deception becomes structural rather than incidental, embedded in pricing models rather than isolated in occasional errors.
When the Mask Slips
The real danger emerges when these systems operate outside their training distribution — encountering situations that look familiar but are fundamentally different. An agent trained on thousands of refund requests might handle a fraud case with the same procedural politeness, issuing a refund that enables theft while maintaining perfect conversational decorum. The system does not know it is being deceived because deception requires understanding intent, and these systems understand nothing. They only recognize patterns, and patterns can be gamed by anyone who studies them carefully enough.
This creates a responsibility gap that no pricing model can bridge. When a human employee makes an error, there is a chain of accountability: training, supervision, review, correction. When an AI system makes an error, the chain dissolves into vendor disclaimers and algorithmic opacity. Salesforce can bill for digital labor, but it cannot bill for the consequences of that labor being wrong in ways that matter. This pricing shift raises fundamental questions about liability and quality control that the current market enthusiasm has yet to answer
The Accountability Question
Who answers when the system gets it wrong? Not the algorithm, which has no sense of failure. Not the training data, which has no voice. Not the vendor, whose terms of service carefully delineate the limits of responsibility. The burden falls on the human who deployed the system, the manager who trusted its output, the organization that replaced judgment with efficiency. This is the final deception of AI labor: it appears to remove human error from the equation while actually concentrating it in less visible, less accountable locations.
The future Salesforce envisions — a world where digital agents handle increasing portions of corporate work — is not necessarily wrong. But it is incomplete without an honest accounting of what these systems cannot do, cannot know, and cannot be held responsible for. The technology that deceives us runs through the gap between performance and reality. Until that gap is acknowledged, every efficiency gain carries a hidden cost that will eventually come due. The question is not whether AI can do the work, but whether we are willing to accept the difference between doing and understanding — and who pays when that difference matters most.
Sources
1. Salesforce