Tracking open items: a list from meeting minutes with AI

A list of open items records who does what by when, and it is updated after every meeting. The AI can prepare it from the minutes: it proposes entries with a source reference and compares the next minutes with the list. What is added, changed or closed is decided by the project lead. Sources as of November 2026.

Why a separate list next to the minutes

In minutes, the open items are scattered: one sentence under item 3, a commitment under item 7, a question that has been carried along for three meetings. Anyone who wants to know what is still outstanding reads everything again. A list of open items (Liste offener Punkte, LOP) pulls this together. BauNetz Wissen describes it as an alternative to detailed meeting minutes that records “the most important tasks and responsibilities in compressed form”. Whether your project keeps a list, minutes or both depends on the project organisation and on the contract.

What matters is that the list is maintained. Course slides on project management from HTWK Leipzig (a university of applied sciences) describe the list of open items as a record of all project events that can replace classic project minutes, and they highlight regular updating and care with an exclamation mark. As a disadvantage they name that the list keeps getting longer. In my assessment, entries in a list that nobody updates go stale, and deadlines pass unnoticed.

What an entry contains

The HTWK slides name this structure: serial number, status, date of origin, priority, category, description, measures, responsibility and date (by when done). That is an example and not a standard. For a list built from minutes I would keep the fields leaner and add two: the source reference (minutes, date, item) and the date of the last entry. With the source reference you check every proposal of the AI against the original without reading the whole minutes.

How the AI proposes entries

The input is the approved minutes, not the raw recording. The language model reads them and proposes an entry for every task, every decision with a follow-up task and every open question. The instruction is to take over only what is in the text and to mark missing details as empty.

An entry could look like this (example with invented data): “Task: add access hatches in the corridor and update the drawing. Owner: Example Building Services Ltd. Deadline: 12 March. Source: minutes 14, item 5.2.” If the minutes say only “building services will clarify this”, the deadline stays empty, and the query reads: “By when?”

Unlike typing it up by hand, the AI also reads the subordinate clause that is easily overlooked when writing up (my assessment, not measured). It enters figures, deadlines and names only if the minutes state them. Calculations, such as converting deadlines into calendar days, are done by a fixed program, not by the language model.

The project lead confirms

Every proposal remains a proposal until the project lead confirms it. They add it, change it or reject it. The German Federal Office for Information Security (BSI) describes, in its guide to attacks on language models, human filtering of actions: the person is asked to authorise a critical action. For a list of open items this means that your confirmation is what turns a proposal into a task assigned to someone. The German Federal Chamber of Engineers (Bundesingenieurkammer) states that when AI tools are used, it is always engineers who in the end take personal responsibility. By analogy, that applies here.

Two things belong to the confirmation. First, does the named owner know about it? A task that someone did not hear in the meeting is still theirs according to the list. Second, is the entry part of the minutes or only a thought from the conversation? You cannot tell from the text.

Comparing with the next minutes

The benefit shows at the next meeting. Once the new minutes are available, the AI compares them with the list and reports four kinds of deviation:

• An item on the list appears in the new minutes as done. It proposes closing it, with a source reference.
• An item appears with a new date or a new owner. It proposes the change.
• An item no longer appears in the new minutes although it is still open. This is the most important report, because this is where something gets lost.
• The new minutes contain a task that is missing from the list.

Here too the AI closes nothing itself: an item that no longer appears in the minutes can be done, forgotten or deliberately dropped. Only someone who was there knows. The report shows you where to ask.

What not to expect

The list knows only what is in the minutes. Agreements made by phone, on site or by email are missing until someone enters them. I have not put a figure on time saved, because I have no measurement. The benefit lies more in less being lost between sets of minutes, and that can only be checked on your own projects.

How the minutes themselves come about is described in the article Meeting minutes with AI: from recording to sign-off. It also explains why checking against the transcript is part of the sign-off. The list presupposes approved minutes. Recording has an article of its own: Recording meetings: what to clarify first.

Minutes and lists contain names and commitments. They are processed only on a platform that can be operated in line with the GDPR, not in a consumer chat service. That requires a data processing agreement (Auftragsverarbeitungsvertrag). Your data protection officer clarifies the individual case.

What you can do now

1. Decide the fields of your list and write them down. Missing details should be allowed to stay empty.
2. Go through the latest minutes by hand and enter all open items. You will see how many there are and where owners or dates are missing.
3. Decide who confirms the list and when, ideally the day after the minutes.
4. At the next minutes, compare list and minutes item by item before you use an AI for it.