What an Agent Needs to Know About Leave Before It Reassigns Work

An AI agent that reassigns work without seeing leave records will give Tuesday's task to someone who lands in another country on Monday night. It will do this confidently, in well written English, and nobody will notice until Wednesday. The minimum an agent needs before it touches an assignment is simple: who is away, when, and who is covering.

Everything else is refinement. If an agent knows only those three things, it stops being embarrassing.

The failure looks competent

This is what makes it worth writing about. An agent working from a task list alone produces output that looks correct. The task exists, the person exists, the date is in the future. Nothing in the record contradicts the assignment.

The contradiction lives in a different system, which the agent was never given. So the mistake is invisible to it and obvious to every human who reads it, which is exactly the pattern that makes people stop trusting a tool.

The minimum an agent needs

Four things, in order of how much damage their absence causes.

Approved leave dates. Not requested, approved. An agent should treat an approved absence as hard as a public holiday, because to the person it is one.

Working pattern. Someone at three days a week is not available on the other two. An agent that assumes five days will schedule work into days that do not exist. This causes quieter damage than leave because nobody is visibly absent.

Who covers whom. When someone is away, the useful action is not "leave it unassigned" or "give it to anyone free". It is "give it to the named cover, or hold it". Without a cover relationship the agent has no correct move, so it picks a wrong one.

Public holidays and regional differences. Teams split across places do not share a calendar. An agent working from one country's holidays will schedule into another country's long weekend every time.

Two more that make it genuinely useful

Avoiding the obvious mistake is the low bar. To be genuinely helpful, an agent also needs two more.

Current load. Reassigning away from someone on leave is easy. Reassigning to the person who already has the most work is a new problem. If the agent cannot see what everyone is carrying, it will balance by name order or by whoever appears first, which is not balancing.

Skill or role fit, even roughly. Handing a task to whoever is free ignores whether they can do it. A rough role marker is enough to keep the agent from suggesting something silly, and a person will catch the rest.

What the agent should do when it does not know

This matters more than any of the above. An agent that lacks context should say so rather than guess.

The right behaviour looks like this: "Three tasks due next week belong to someone on approved leave from Monday. No cover is recorded. Suggested reassignments below, needs a decision." That is useful. It has found the problem, done the tedious part, and stopped short of acting.

Compare with silently moving the tasks to a random available person. Same information, worse outcome, and now someone has to work out what changed and why. It is the same reason some agent outputs should carry a human signature while others can go straight through.

Why this is a records problem, not a model problem

People tend to treat this as a question of how clever the assistant is. It is not. A very capable model with no leave data will make the same mistake as a simple one. The constraint is what the agent can see.

That is the argument SicherOne is built on: project management, HR and AI agents work from one set of records, so a board knows who is on leave and an agent has full context rather than a fragment. A human approves agent output before it ships, which is the part that keeps a suggestion from becoming an unnoticed change. Where data handling requires it, private models can be self hosted.

Access should be narrow and deliberate

Giving an agent leave data is not the same as giving it HR data. It needs to know that a person is away and until when. It does not need to know why, whether the leave was sick leave, or anything about pay or performance to reassign a task. Where the records are sensitive enough that they should not leave the building, keeping the model on your own hardware is a separate decision worth making early.

Draw that line before switching anything on. Write down what the agent can read and what it can put into output that a wider group will see. An agent that can read a record can quote it, and a status update drafted by an agent read by twelve people is not the place for something that was shared with one manager.

A short checklist

• Approved leave, with dates
• Working pattern per person
• Named cover for each person, kept current
• Correct holiday calendar per location
• Rough current load
• Rough role fit
• Clear instruction to flag rather than act when cover is missing
• A person approving anything that changes an assignment

Six pieces of context and two rules. That is the difference between an assistant that saves a manager an hour and one that creates a small mess every fortnight.