Why an Agent Should See the Leave Calendar Before It Books a Meeting

An AI agent books a review for Thursday. Three of the four people are free. The fourth has approved leave that week, which the agent never saw, because leave lives in the HR system and the agent was only given the calendar. The meeting goes out, someone forwards it to the person on holiday, and the agent looks stupid in a way that is very easy for everyone to notice.

This is the clearest illustration of a general problem. An agent is only as sensible as the records it can read, and most companies keep the records an agent needs in two systems that do not talk to each other.

The failure is not the model

It is tempting to read the Thursday mistake as the agent being bad at its job. It was not. Given a calendar with no conflict on Thursday, booking Thursday is the correct answer.

The agent answered the question it was asked using the data it was given. The data was wrong by omission. No amount of a better model fixes a missing dataset, and this is worth being clear about before anyone tries to solve it with prompts. It is also what managers get wrong about agent memory.

Why calendars alone are not enough

The obvious objection is that people should just block their calendars when they take leave. Some do. In practice this fails often enough to matter:

• Leave approved months ahead, before anyone thinks about calendars.
• Half days, which show as a full free afternoon or a full blocked one.
• Public holidays that differ across locations in the same team.
• Leave approved after a meeting was already booked, where the calendar never gets revisited.
• People who genuinely forget, which is most people occasionally.

The leave record is the authoritative version. The calendar is a copy somebody has to remember to make. Asking an agent to work from the copy is asking it to be wrong sometimes, in public.

Scheduling is just the visible case

Scheduling is the example everyone recognises because the error is immediate and social. The same gap causes quieter damage elsewhere.

An agent that cannot see leave will suggest a deadline that lands while the owner is away. It will summarise a project as on track when the person who owns the next step is out for two weeks. It will assign a follow up to someone who will not read it until after it matters.

None of these produce an embarrassing calendar invite. They produce plans that are slightly wrong in ways nobody catches until the date arrives. Unplanned absence is harder again, because sick leave leaves half finished work behind.

What sharing records actually means

Not an integration. Not a nightly export. One set of records that both the project side and the people side write to and read from.

The difference shows up in the edge cases. With an integration you get a lag, and a mapping between two identity systems, and a category of bug where the two disagree and nobody knows which is right. With shared records the leave approval and the project board are describing the same fact, so there is nothing to reconcile. That is what changes when project boards, leave calendars and agents sit on one record.

SicherOne is built on that idea. Project management, HR and AI agents run on one set of records, so an agent looking for a meeting slot sees leave the same way it sees the board. It is not asking another system and hoping the answer is current.

Access is not the same as exposure

Giving an agent the leave calendar does not mean giving it everything in the HR record.

For scheduling, an agent needs to know that a person is unavailable on certain days. It does not need to know why. Reasons for leave sit behind tighter permissions than dates, and an agent should get the same treatment as anyone else: the fields the job requires and no more.

That separation is worth setting up before you switch anything on, not after someone asks an uncomfortable question about what the agent can see.

The rule underneath

Before you hand an agent a task, ask what a competent new employee would need to look up to do it well. Then check whether the agent can reach all of it.

If a new joiner would check the leave calendar before proposing a meeting time, and your agent cannot, the agent will make a mistake a new joiner would not. Not because it is less capable. Because you gave it less to work with.

And keep a person in front of anything that goes out. In SicherOne a human approves agent output before it ships, which is what catches the cases where the records were complete and the situation still had something in it that no record held.