Notes
Why agents do not onboard like people
People become useful by seeing work, acting, being corrected, and updating their company map. Agents need that loop made durable.
A new employee is not useful because they received enough documents.
They become useful because the company becomes legible to them.
That difference explains a lot about why AI agents still struggle inside real companies.
A person joins, reads the docs, sits in meetings, asks awkward questions, makes a few small mistakes, and gets corrected. Over time, those corrections change their internal map of the company. They learn what matters, who owns what, which customer is sensitive, which product promise is dangerous, and which source is trusted when systems disagree.
Agents usually do not learn that way.
They receive context, call tools, produce an answer, and move on. The answer may be useful. But unless the correction updates durable company state, the agent does not leave with a better map of how the company works.
Documents are only the start
Companies often treat agent onboarding as a document problem.
Give the agent the handbook. Give it the Notion space. Give it CRM access. Give it the Slack export. Let it search everything.
That helps, but it is not how people actually become useful.
A good employee learns through exposure to the work. They hear disagreements. They notice what gets escalated and what gets ignored. They see one customer treated differently because of a renewal, another because of legal risk, another because the founder knows the buyer personally.
They learn the things that rarely make it into a process doc:
- “we can probably do that” means one thing from sales and another thing from engineering;
- the named account owner is not always the person with the real context;
- a small promise can become a delivery problem three weeks later;
- the official source of truth may be less true than yesterday’s call.
No one calls this a company map.
They call it experience.
The useful loop is correction
People become useful through a simple loop:
- exposure to real work;
- action;
- correction;
- updated map;
- better next action.
The correction is the important part.
A person hears, “No, do not send that follow-up yet. Product has not confirmed the date.” Next time, they recognize the same pattern. A Q3 promise is not just a task. It is a promise risk. It touches product, sales, delivery, renewal value, and what can safely be put in writing.
The map changed.
For an agent, the same correction often dies in the chat.
It helped once. It did not update the account. It did not mark the promise risk. It did not change the rule for future follow-ups. It did not teach the next agent what kind of situation this was.
The gap is not the correction itself. It is that the correction does not change the system.
Memory alone is too loose
The obvious answer is to give agents memory.
Memory helps. But a loose memory that says “be careful with Q3 promises” is not enough.
Careful with which promises? For which product area? Which customers? Who approves the date? What changes downstream if the date moves? Which system should be updated? Which similar cases should trigger the same behavior?
The useful correction has structure.
It should update the objects and relationships around the work: the customer, opportunity, product limit, owner, risk, approval rule, source evidence, and next action.
Permissions are also not enough. Permissions say what an agent is allowed to do. They do not tell the agent what a good action is.
Evals are not enough either. Evals can tell you whether an output passed a test. They do not make the company easier to understand.
The missing middle is the mapped job: the part of the company where the agent works, with the relevant objects, state, history, rules, approvals, and corrections made explicit.
Agents need the map made explicit
This does not mean agents should replace the human onboarding process.
It means we should study why onboarding works for humans.
Humans become useful because they convert messy exposure into a working model of the company. They do not just store facts. They learn relationships: this customer matters because of this renewal; this promise is risky because of this product limit; this follow-up needs approval because of this buyer, deal stage, and prior conversation.
Agents need that model outside someone’s head.
Start with one job an agent should own. Turning calls into follow-up work. Keeping an account plan alive. Watching delivery risk. Updating customer context after a meeting. Finding the right accounts in a regulated market.
Then map the work around that job.
What should the agent read? What can it change? Which objects matter? What counts as evidence? What needs approval? What should happen when a human corrects it?
The correction should not disappear after one task. It should improve the system the next task runs on.
Agents get better inside a company when corrections change the map they work from.
Related: access is not a map and what a company ontology is.
