The more news we read about the evolution of corporate AI offerings and implementations, the more I struggle with one thing: The industry is rapidly turning AI agents into some sort of simulated employees.
There seems to be a rush to build the org chart in silicon: Microsoft gives agents identities, sponsors, access rights, and employment-like lifecycle management , and it explicitly governs them in a similar way as human identities. Salesforce even describes agents as “digital workers,” and promotes the idea of a digital workforce that tries to mimic human capabilities. ServiceNow has introduced individual AI passports specifying what each agent may do at runtime.
Apparently, what we are trying to do is reproduce the human organization in software, instead of asking ourselves what a truly machine-native organization should look like. Now companies and managers brag about the number of agents they supervise, as if this were some sort of badge of honor. As usual, and as I say in all my innovation courses, learning is way more difficult than unlearning.
Is it smart to define our architecture with the employee metaphor?
Why are we sticking to that metaphor? Human organizations have roles and hierarchies partly because we, humans, have characteristics such as limited attention, memory, and communication capabilities. However, I don’t think I need to explain to you that software does not have these restrictions, and therefore, does not necessarily need to inherit those constraints.
You can do so if you want or if you don’t know how to do better: After all, identity and permissions are necessary. But bear in mind that such anthropomorphism is, at some point, going to become technical debt. A collection of increasingly autonomous agents can re-create silos, develop conflicting incentives, and even generate coordination overhead, but now at machine speed.
Agents should be replaceable, not institutional
Perhaps what we should do is treat agents as a temporary execution process: instantiated for a certain purpose, constrained, observed . . . and discarded. Among other reasons, because we should try to avoid at all costs letting knowledge, identity, or operational history become trapped inside particular agents. In the same way models should be replaceable components , agents should be too.
I understand that AI agents are fairly new, and managers still look at them as if they’re some sort of “magic thing.” However, this sort of agent mystique should come to an end. An agent is a software process equipped with an objective, a set of tools, access to context and memory, and some capacity to decide what to do next, which ideally should be able to let us generate some sort of feedback or learning loops. But we should stop treating agents as if they’re magic, and even less as humanlike. They are not.
The employee metaphor is potentially useful for adoption and access management, but completely inadequate as the foundational architecture of an intelligent company.
What must survive the agent
When we consider agents as replaceable components rather than allowing them to become new organizational silos, all the things that the company learns must exist independently of them. The company’s formal representation of itself—from its customers to its products, employees, suppliers, contracts, and processes, along with all the relationships among them—should persist. This is what I usually call “the nouns” and “the verbs,” and Palantir pompously calls “the ontology.”
But the operational state of the company (not just what exists, but also what is happening, what has changed, and what remains unresolved) should be preserved, too, along with the objectives and trade-offs defined by leadership , the permissions governing every action, and the rules determining when a decision must be escalated.
Exactly the same thing applies to institutional memory: Every relevant decision, action, and outcome should be recorded in a journal that no agent or model can silently rewrite, creating a reliable account of how the organization arrived at its present state. Maintaining rigorous control of the integrity of such a journal is essential for future accountability, and as we now know well, even for cybersecurity .
That record must then feed back into future behavior, allowing the company to learn from successes, failures, and unintended consequences. Models may be upgraded or replaced, agents may be created or retired, and entire workflows may be redesigned, but the organization’s accumulated experience must remain. Otherwise, the company is not learning: It is repeatedly hiring intelligent amnesiacs.
Individual intelligence vs. organizational intelligence
In corporate AI, we need to understand that the important unit of analysis should be the organizational loop, not some sort of “synthetic worker.” Better models may improve individual cognitive performance, more agents may increase available activity and make their supervisors feel more important, but none of these automatically makes the company learn in a coherent way.
The next enterprise platform will not look like a workforce or management system for bots. That’s clearly not the point. It will be the persistent substrate in which models and agents can be exchanged while the company’s accumulated intelligence remains safe and under control.
The still missing product category is the system that preserves state, authority, evidence, and learning across models and agents. But don’t worry: It will not be missing forever.
Stay tuned.
