AI Agents Are the New Employees
And the costs human patience used to hide can now be metered task by task
Disclosure: I build IndustryVault, whose architecture reflects the governed data architecture argued for here.
AI’s New Economics argues that repeatable cognitive labor is becoming abundant. But abundant labor still has to be onboarded. AI agents are new employees in one operational sense: they need company context before they can do useful work. This essay is about the recurring cost of supplying it.
A capable new hire spends their first weeks learning the company — the people, processes, vocabulary, unwritten rules, and which systems to trust. Then they reuse that internal map from task to task. The map is imperfect and needs upkeep, but the employee does not reconstruct it from scratch each morning.
An agent faces the same onboarding problem with a crucial difference: the context it needs has to be represented, retrieved, and processed for the task at hand. Persistent workspaces can preserve some state, but the model does not arrive with an authoritative map of your company. Someone has to supply that map, or the agent has to go fetch it.
For humans, that context-fetching was never free; it was mostly invisible at the task level. We built the workplace for people who remember where things live. Knowledge could scatter across teams, tools, and people’s heads because an employee would learn whom to ask and which system to open. Each trip consumed paid time and throughput, but the cost was absorbed into salary and rarely allocated to the task that incurred it.
That is what changed. When an agent performs the same navigation — opening another system, pulling another team’s spreadsheet, reconciling two numbers that do not match — the cost can be measured far more directly in tokens, latency, tool calls, compute, and dollars, task by task. The inefficiency did not appear with AI. It was financed quietly by human patience; AI made it visible in the system’s telemetry and unit economics.
Make it concrete. Ask an agent a routine question — is this loan eligible for the current pool? — in a fragmented shop. It visits the servicing system for the balance, the document store for the note, the pricing sheet for the pool rules, and the warehouse for payment history. It drags each result into working context and reconciles four vocabularies before the actual work starts.
Much of that effort is navigation. Put everything needed to understand the loan behind one governed front door and the same question can become one governed request, with reconciliation performed upstream as shared work instead of repeated inside every task. The reconciliation remains ongoing as records and rules change, but its cost is spread across every question that reuses it.
Cost is the visible problem. Correctness is the dangerous one.
A person who cannot tell whether a number is current can stop and ask. An agent that has not been designed and evaluated to abstain may instead fill the gap fluently and return something that looks exactly like a right answer.1 Each additional place it visits is another chance to retrieve a stale copy, miss a relevant rule, lose provenance, or infer the part it could not find.
So questions such as “Is this current? Where did it come from? Can it be trusted?” cannot be left until the end. The answer needs to travel with the work. Better models, tools, retrieval, and evals all help, but none makes conflicting sources agree. A larger context window can hold more fragments; it cannot decide which fragment is authoritative.
The design test is simple:
How many places does someone — agent or person — have to visit to understand everything relevant to one customer?
Driving that number toward one can reduce repeated retrieval cost and opportunities for inconsistency, provided the result is current, governed, provenance-aware, and permissioned.
This is not a call to build one enormous system. A front door is a governed place to ask about the entity — a customer, order, or loan — and receive the relevant facts, rules, history, and evidence together. Source systems may remain where they are. But the front door has to identify what is authoritative, expose freshness and provenance, and make disagreements visible. If it merely hides five conflicting backends behind one API, it has shortened the URL, not the scavenger hunt.
The payoff is not only for agents. A workplace legible enough that an agent can reason over it is also easier for a new hire to learn and less likely to lose critical knowledge when someone leaves.
Making the workplace legible solves the operating problem. Own Your Operating Memory takes up the strategic one: who controls the records, rules, and accumulated judgment agents use and create?
Humans learned to navigate the fragmented workplace. Agents make the company pay for that navigation task by task. You do not solve that with a bigger model alone. You design around it.
Footnotes
OpenAI, “Why Language Models Hallucinate”, September 5, 2025. The analysis argues that common training and evaluation procedures can reward guessing over acknowledging uncertainty. Deployments should explicitly reward abstention, clarification, and evidence-backed answers.↩︎