Skip to content
Sentynel

#opinion3 min readBy Sentynel

The answer your agent needs lives in one person's head

Most companies have one person who knows where the real answer is. An agent cannot ask them. Writing that knowledge down is the first job of any operation redesign.

Share

Every company has someone who knows which spreadsheet is the real one, which client always pays late, and which number in the report should not be trusted. Ask an agent to do that person's job and it will fail in the same place every time. The knowledge an agent needs most is the knowledge nobody wrote down, and no model upgrade will supply it.

The shortcut nobody documented

Ask how a recurring report gets made and you will hear a process. Watch it being made and you will see a different one: open this export, ignore the first tab, correct that column by hand, check the figure with someone in the other office.

None of those steps is in a manual. They live in one person's habits, and they are the reason the report is right. A new hire learns them over months. An agent gets only what is written, so it produces a clean, confident report that quietly leaves them out.

Why this is an operations problem, not a model problem

When an agent gets it wrong, the usual reaction is to change the prompt or the model. That treats a missing procedure as a missing capability.

Three signs the gap is in the operation:

  • Two experienced people describe the same task differently.
  • The honest answer to "how do you know it is right?" is "I just know".
  • The task works only when one particular person is not on leave.

An agent hired into that situation inherits the ambiguity and runs it faster. A better model does not remove it.

Write the judgment down, not just the steps

A step list says what to do. The judgment says when not to. Capture both, with the person who holds the knowledge:

  1. Have them do the task while you record each decision, especially the ones that begin "usually" or "unless".
  2. For each decision, write the rule and one real example where it applied.
  3. Name the sources of truth, and say which wins when two disagree.
  4. Mark what the agent must not decide alone, and who gets the case instead.

This is dull work. It is also the work that turns a one-person habit into a shared operation that any agent, or any new colleague, can follow.

What changes when the evidence is connected

In the OCESA case, ticket sales, sponsorship results, ratings and media evidence sat in separate places, and one reporting operation was built to connect them. Sentynel's project case record reports a −70% reported operating cost, and the reporting cycle moving from weeks to minutes. The cost comparison is normalized, so treat it as an indexed comparison, not an absolute saving.

The lesson is not the percentage. Connecting the evidence to the question meant deciding, explicitly, where each answer comes from. That decision used to live in the heads of the people who assembled the report by hand.

OCESA · reporting cycle
WeeksEvidence spread across separate sources
MinutesSources connected into one reporting operation

Source: Source: Sentynel project case record. Results are reported project outcomes; the cost comparison is normalized.

From one person's habit to an operation
  1. 01ObserveWatch the real task
  2. 02CaptureRules plus real examples
  3. 03Name sourcesWhich one wins
  4. 04Set limitsWhat needs a person
  5. 05Hand to agent

Only after the fourth step does an agent have something safe to work inside.

A test you can run this week

Pick the report or process your team trusts most, and ask the person who runs it to tell you three things they do that are not written anywhere. If they find them quickly, you have a documentation backlog. If they struggle, the knowledge is deeper than they realize, which is where agents fail.

Which of your processes would break tomorrow if one person stopped answering messages?

Sources

  1. https://sentynel.ai/#cases
Share
← All sygnals