Most missed deadlines give no warning. The owner may have known the date was unrealistic, but nobody asked, and the first sign of trouble is that the date has passed. By then, the manager can no longer change the scope or clear the blocker.
Simple automated reminders do not fix this, and often make it worse. They remind people about work that is already done. They treat a date from a contract the same as an internal target. They copy managers on every message, so the one miss that matters gets lost among the rest.
Sending a reminder on a date is easy to automate. Useful follow-up is harder. The system has to know where the date came from, whether the owner agreed to it, what has changed since the reminder was set, and which problems still need a person to decide.
An agent that keeps sending messages until someone answers cannot do this. What works is a record of the work, a short list of written rules, and an agent that applies those rules to the current state of the work.
The agent sends the follow-up. People still make the decisions.
Record where each date came from
Two tasks can both be due on Friday and matter in very different ways.
One date may come from a contract, a client request, a filing requirement or the end of a sprint. Another may be an internal planning target. Each organization decides for itself which kinds of date are fixed and which can move.
Store that with the date. Do not ask the agent to guess it from the task title, the tone of an email or the number of reminders already sent.
A useful record of the work answers four questions:
- Who owns the work?
- When is it due?
- Where did that date come from?
- What happens if the date is missed?
Without those answers, the model has to make up the rules one task at a time.
Ask the owner to agree to the date
A proposed date is not yet a commitment.
The owner needs a simple way to accept the date, or to object and say why. The reason may be a full sprint, a missing dependency or a clash with other work. When the objection comes early, a manager can still change the scope, reorder the work or clear the dependency.
Without this step, the first signal is a red "overdue" label after the date has passed. That shows the system noticed the miss. It does not show whether the date was ever realistic, or whether anyone agreed to it.
This matters most when the AI proposes the date. The agent can suggest a date, record the owner's answer and track the date they accept. Its own suggestion is not evidence that a person agreed.
Check the current status before sending a reminder
Base each reminder on the work record as it is now. The event that scheduled the reminder may be out of date.
We do this in our own daily briefing. Before the agent repeats an item from the queue, it looks up the record the item refers to again. A meeting acceptance, an out-of-office reply or a warning may have mattered when it arrived and be out of date by the time a person reads the briefing.
Deadline follow-up needs the same check. Before drafting or sending a reminder, read the current status, owner, date, blocker and linked documents. If the work is done, has a new agreed date or is waiting on someone else, the old reminder should not go out as if nothing had changed.
You can test this. Schedule a follow-up, then finish the task or change its date before the follow-up is due. A well-built workflow reads the current record and stays quiet. A simple notification system sends the original reminder anyway.
Write down when a manager hears about it
Managers do not need a copy of every automatic reminder. They need to hear about the missed dates that put delivery at risk.
Decide those cases before the agent runs. For example, one team might allow one direct reminder for a missed internal target, and send a missed client or regulatory date to a manager straight away. Another team may draw the line somewhere else. What matters is that the rule is written down where people can review it, and the model does not make it up as it goes.
The message to the manager should open the same work record. In our own system, an alert cannot go to a phone unless it is attached to a work item, so the person who gets it goes straight to the status, the owner and the documents behind it. The manager can act on the record instead of piecing the story together from old notifications.
Count only real signs of progress. A changed status, a delivered document or a recorded blocker is evidence. Silence tells you nothing, and time booked on a calendar does not prove the work is done.
Review the workflow before you automate it
If commitments are spread across email, tickets and project plans, first decide which record the follow-up should open. Then write down the rules:
- Which system holds the status?
- Who can set or change the date?
- Where is the source of the date recorded?
- How does the owner accept the date or object to it?
- What evidence shows the work is done?
- What stops an out-of-date reminder from going out?
- Which missed dates go to a manager?
- Which record does the manager's alert open?
Our AI Workflow Review is built for these questions. A team brings up to three workflows it wants to automate. Over two weeks, we map each workflow, its data and who signs off, then deliver a written assessment: what to build, what would stand in the way, and which workflow to start with.
Deadline follow-up is a good place to start. Sending the message is the easy part. Most of the work is in the rules and records that decide whether the reminder is still true.
For how we decide which alerts reach a person at all, read How we cut ten alerting agents down to one. Our own follow-up rules are on How we run Kaizen on AI.