A deadline reminder is easy to automate. A useful follow-up workflow is harder. It has to know what the date means, whether the owner agreed to it, what changed since the reminder was scheduled, and which exceptions still need a person.
The safe design is not an agent that keeps sending messages until someone answers. It is a work record, a small set of explicit policies, and an agent that applies those policies to current evidence.
The design principle is simple: automate the follow-up, not the judgment.
Start with what the date means
Two tasks can both say Friday and carry different consequences.
One date might come from a contract, a client request, a filing requirement or a sprint boundary. Another might be an internal planning target. Those are examples, not universal categories. Each organization needs to define which sources make a date firm and which allow replanning.
Store that meaning with the date. Do not make the agent infer it from the task title, the tone of an email or the number of reminder messages.
A useful work record can answer four questions:
- Who owns the work?
- What is the due date?
- Where did that date come from?
- What policy applies if it is missed?
Without those fields, the model is being asked to invent operating policy one task at a time.
Make agreement explicit
A proposed date is not automatically a commitment.
The owner needs a clear way to accept it or object with a reason. That reason may expose a full sprint, a missing dependency or a conflict with another commitment. Capturing the objection early gives a manager a chance to change scope, sequence the work differently or resolve the dependency.
The alternative is a red overdue badge after the deadline. That proves the system noticed the miss, but it does not show whether the date was ever feasible or agreed.
This boundary matters when AI proposes the date. The agent can suggest timing, preserve the response and track the resulting commitment. It should not turn its own suggestion into evidence that a person agreed.
Recheck the source before chasing
A reminder should be based on the current work record, not only on the event that scheduled it.
KZN uses this principle in its briefing workflow. Before repeating a queued finding, the process re-queries the referenced live record. A meeting acknowledgment, out-of-office reply or operational warning may have been relevant when collected and stale by the time a person sees it.
Deadline follow-up needs the same check. Before drafting or sending a nudge, read the current status, owner, date, blocker and linked evidence. If the work is complete, legitimately replanned or waiting on a dependency, the old reminder should not speak as though nothing changed.
The test is straightforward. Schedule a follow-up, then complete or replan the task before that time. A well-built workflow follows the live record. A notification pipeline simply replays the original event.
Put escalation rules in policy
Managers do not need a copy of every automated nudge. They need the exceptions that change delivery risk.
Define those exceptions before the agent runs. For example, an organization might allow one direct follow-up on an internal target, while routing a missed client or regulatory date to a manager immediately. Another team may use different categories. The important part is that the rule is configured and reviewable, not improvised by the model.
The escalation should open the same work record. KZN's phone-alert mechanism requires a work item, so an interruption leads to the status, ownership and evidence behind it. The manager can act on the record rather than reconstructing the story from notification history.
Avoid made-up progress signals. A changed status, a delivered artifact or a recorded blocker is evidence. Silence is not a percentage complete, and time reserved on a calendar is not proof that the work finished.
Review the workflow before automating it
If commitments are spread across email, tickets and project plans, start with the record the follow-up should open. Then write down the policy:
- Which system owns the status?
- Who can set or change the date?
- How is the source of the date recorded?
- How does the owner accept or object?
- What evidence closes the work?
- What conditions suppress an old reminder?
- Which miss reaches a manager?
- What record does the escalation open?
KZN's AI Workflow Review is built for this kind of question. A team brings up to three workflows it already suspects. Over two weeks, the review maps the workflow, data and sign-off, then returns a written assessment of what to build, what would block it and which workflow should go first.
Deadline chasing is a good candidate because the automation is not the message. It is the policy and evidence that decide whether a message is still true.
The related attention pattern is in How we cut thirteen alerting agents down to one.