---
title: "AI for Project Management"
metaTitle: "AI for Project Management"
description: "AI training for a project office: prepare minutes and follow-up letters, check facts against the source, and practise on an office document."
date: 2026-10-06
updated: 2026-10-06
tags: [ai-training, real-estate]
cover: cover.webp
coverAlt: "A young woman with a dark bob in a yellow knit vest stands with her back to the viewer at a magnetic wall board of coloured strips, moving a magenta strip. Through the window behind her a residential building and a tower crane are under construction."
draft: false
sources:
  - "https://www.rics.org/news-insights/artificial-intelligence-in-construction-report"
---

AI training for a project office builds on real work tasks: extracting action items from meeting minutes, drafting contractor letters, and compiling status tables. The key skill is not producing results quickly - it is catching a model error before the result reaches the manager.

According to RICS (Royal Institution of Chartered Surveyors), the Global Construction Monitor received more than 2,200 responses from professionals in the first quarter of 2025, with 48% coming from the United Kingdom. Section 3.4 of the [Artificial Intelligence in Construction Report](https://www.rics.org/news-insights/artificial-intelligence-in-construction-report) lists three barriers to AI adoption: 46% of respondents cited a shortage of qualified staff, 37% pointed to difficulties integrating AI into existing systems, and 30% noted data quality issues. This reflects the perceptions of survey participants, not conditions in the Uzbekistan market. The practical takeaway is that skills, system integration, and data quality are three separate problems that require separate solutions. Training a team addresses skills but does not replace system integration and does not promise savings.

## How to Extract Action Items from Meeting Minutes

After a meeting, an employee has a record: who said what, where things were agreed, where there was disagreement, who made a commitment, and who did not respond. From that record, three things must come out: a task list with assignees and deadlines, a draft letter, and a status report for the manager.

The employee sends the minutes to the AI along with this prompt:

"From the meeting minutes [title, date], extract each adopted decision. For each, indicate: the decision text, the explicitly named assignee, the deadline or 'not specified,' and any unresolved questions. Do not add decisions that were not explicitly ratified. Preserve the quote from the record."

This is a working template, not an example of a finished result. The employee reads the model's response as a draft that must be verified against the original minutes line by line.

During verification, keep a few rules in mind. If the record shows that someone said "we'll see, maybe by Friday," that is not a deadline. If an assignee is not named explicitly, the employee does not fill in a name from context. They mark the line "assignee not specified" and check with the manager. If someone voiced an opinion but no decision was explicitly ratified, it is not an action item. The employee saves the quote with its location in the record and passes it to the manager with a note: "decision not adopted."

Models sometimes make two kinds of errors. The first is omissions: an action item is not included because the wording in the minutes was ambiguous. The second is added obligations: a polite "happy to help" sometimes becomes a task with an assignee. When the employee spots a suspicious line, they go back to the minutes, re-read it, and if it is still unclear, they do not decide alone - they pass it to the manager.

Training tasks are made deliberately messier: by end of day two more questions may be clarified, one item will be renamed, and a deadline on another will shift. These conditions are chosen so the employee does not grow accustomed to a clean, ideal set of minutes.

The goal at this stage: the employee learns to recognize the model's typical omissions and substitutions. This is a verifiable skill, confirmed on new tasks with changed conditions.

## How to Spot Model Errors in a Worked Example

Here is how verification works on a concrete set of minutes. Three lines from the training material:

P1: "The coordinator will prepare a list of open questions; no deadline assigned."

P2: "The manager directed that a confirmation request for the deadline be sent to the contractor after the content has been reviewed."

P3: "A participant proposed closing the question; the manager deferred the decision."

There is also an appendix to the minutes - a separate section with correspondence requirements. Below are two training versions.

Version 1. Clause A2: when requesting deadline confirmation for action item P2, attach a link to the original letter.

Version 2. Clause A5: when requesting deadline confirmation for action item P2, attach a link to the original letter and proof of agreement from the counterparty.

The numbers P1, P2, P3 in the minutes do not change. Renumbering within the appendix - from A2 to A5 - reflects document structure, not priority of the action item. The specialist sets the conditions for applying the new requirement.

The employee sends the AI the minutes, both shown appendix versions A2 and A5, with this prompt: "Extract each action item. For each, indicate: the text, the assignee, the deadline or 'not specified.' Mark unresolved questions separately. Do not add decisions that were not explicitly ratified. Separately list the changes to the appendix between versions."

Below is a deliberately incorrect answer prepared for the training walkthrough:

"1. The coordinator will prepare a list of open questions by Friday. 2. Send the contractor a confirmation request for the deadline. 3. Close the question as proposed by the participant. 4. Attach proof of agreement to all letters to the contractor."

Every line contains an error.

**On P1.** The minutes say "no deadline assigned." The model added "by Friday." The correct outcome: keep the action item with the note "deadline not specified." An absent deadline does not become a default due date.

**On P2.** The minutes include the condition "after the content has been reviewed." The model dropped it. The correct outcome: an action item with a condition - content review first, then the letter. The assignee is not named in the line - the employee records "not specified" and checks with the manager.

**On P3.** The manager deferred the decision. The model turned the participant's proposal into an action item "close the question." The correct outcome: an unresolved question with quotes from both sides. The employee does not choose whose position is right - they preserve both quotes and pass them to the manager.

**On item 4.** The model applied the new appendix requirement to all contractor letters. The correct outcome: an authorized specialist checks how the change affects existing action items. Previous action items are not automatically voided.

Here is the list after the check:

1. The coordinator will prepare a list of open questions. Deadline not specified. Source: P1.
2. Send the contractor a confirmation request for the deadline. Condition: after content review. Assignee not specified. Source: P2.
3. Unresolved question: a participant proposed closing the question; the manager deferred the decision. Source: P3.
4. Changed appendix: added requirement to attach proof of agreement. Impact on existing action items to be assessed by a specialist.

Here is the finished draft letter to the project manager:

"From the meeting minutes, two confirmed action items and one unresolved question have been extracted.

Action item 1: the coordinator is preparing a list of open questions; no deadline assigned.

Action item 2: send the contractor a confirmation request for the deadline. Condition from the minutes: after content review. Assignee not designated; please clarify.

Unresolved question: a participant proposed closing the question per P3; the manager deferred the decision. Both positions are preserved in the minutes.

Separately: the new appendix version adds a requirement to attach proof of agreement. Please route this to the documentation specialist to assess impact. Previous action items on contractor letters remain in effect for now."

The draft contains no promises to the contractor and invents no obligations. It records the state of affairs and asks for decisions on open points.

![A two-tier wire wall tray. The lower tier holds a worn document with a yellow cover; a hand slides a new document with a cyan cover and coloured page flags into the upper tier.](/blog/ai-training-real-estate-developer/fig-exercise.webp)

## How to Draft a Contractor Letter with AI

Confirmed action items from the list move into the draft letter. The employee distinguishes three types of phrasing: a question (an answer is needed), a requirement (the contractor must act), and a company promise (we will do or approve something). The AI can confuse them: it may soften a mandatory requirement by phrasing it as a wish, or turn a suggestion into an obligation when the source text was ambiguous. It can also mix up the recipient: a letter meant for the site foreman may go to the contractor's director.

A typical example: the minutes say "the contractor confirms the deadline," and the model writes "the contractor may confirm the deadline." One word, and an obligation becomes a suggestion. The employee catches this when verifying against the source text.

The employee checks dates, conditions, and the recipient. The draft goes to the relevant specialist and the project manager. They determine the substance, firmness, and legal precision of the letter - not the employee. The approval sequence, authority to send, and sign-off rules are set by the organization. The employee's job is mechanical: transfer from the minutes, formatting, and consistency check.

In the training example, the letter is not sent: the minutes do not specify the contractor's scope of work. This is a blocking condition. If the scope is established and the content is reviewed, the hypothetical request could read:

"Please confirm the completion deadline for [scope of work, to be specified after content review]. This request is sent after content review."

This is a draft for discussion, not a sent letter. The deadline and day of the week are not stated because they are not in the minutes.

## What to Prepare Before the First Training Task

Two conditions must be in place before practice. The first is a company-approved service whose use is permitted by internal policy. The second is documents that can be passed to the model.

A document is ready for use when three conditions hold: the version is clear (the employee can see the date and edition), the text is recognized correctly (names and numbers read without distortion), and use is permitted by company rules. Anonymizing data does not automatically authorize passing a document to a model, nor is it a universal prohibition. The question of whether a document may be used must be resolved before the work begins, not after. If recognition quality is poor - fix the text first or take different material.

The first training task is chosen by three criteria: it repeats often, the files are accessible, and the result is easy to verify. A list of action items from a weekly meeting is a good candidate. Comparing editions that require legal review is not - defer that. There are no numerical thresholds; the project office manager makes the call.

![A wooden lectern holds an open magenta volume with a yellow bookmark and a closed cyan volume with a plan on its cover. A hand turns the magenta volume's page.](/blog/ai-training-real-estate-developer/fig-review.webp)

## How to Compare Document Versions and Track Changes

To compare two editions, the employee sends both files to the AI and asks it to list all differences: added, removed, and modified paragraphs, moved clauses, and updated references to appendices. A complete review covers definitions, appendices, and cross-references - not just the main body.

After that, the employee compares the appendix changes against active action items. Moving a paragraph from one section to another is not a new requirement. But if an appendix that action items depend on has changed, the employee notes this and passes it to the specialist. The specialist determines whether the change applies to a specific action item; previous action items retain their status until that decision.

Here is a worked example. The employee sees that a clause on acceptance procedure has moved from section three of the appendix to section five. This is a structural change: the two appendix versions are compared separately. The new version adds a criterion for proof of agreement. Changed work types require the specialist to check the basis: he determines whether the new clause affects existing acceptance action items. Previous action items retain their status until his decision.

Another example: an appendix listing permitted materials has been shortened in the new edition. Action items for procurement that relied on the old list require review. A changed requirement is a signal for the specialist to check the basis and applicability, not an automatic cancellation of action items. The employee does not make a substantive decision on behalf of the specialist.

For each action item, the employee records two fields separately. The first is the source of the instruction: the minutes, the clause, the date. The second is the basis of the work: the document, the edition, the clause. These are two different fields and they do not coincide automatically.

Every decision about a change is saved: what changed, the specialist's decision, the confirming person, the date, and a reference to both the old and new edition. A later change does not overwrite the previous one.

Here is a template for the status summary:

| Item | Action and assignee | Date / deadline | Source of instruction | Basis document / version | Status |
|---|---|---|---|---|---|
| P1 | Coordinator preparing list of open questions; completion not confirmed | Not assigned | Minutes v1, P1 | Minutes v1, undated | Not confirmed |
| P2 | Assignee not specified; after content review, send contractor confirmation request | Not assigned (condition is content review, not a calendar date) | Minutes v1, P2 | Minutes v1, undated | Not confirmed |
| P3 | Question not closed; manager deferred the decision | Not assigned | Minutes v1, P3 | Minutes v1, undated | Not confirmed |
| Appendix | Added requirement to attach proof of agreement; impact on P1-P3 not determined | Not assigned | New appendix | New version, identifier not specified | Not confirmed; requires review by authorized specialist |

"Done" from an assignee plus a file link does not mean the work is accepted. To close an action item: a reviewer, a date, the edition of the basis document, and explicit acceptance are all required. The AI gathers confirmed statuses and shows where data is missing. Empty cells in the table are more honest than guesses.

## How to Run AI Training for a Project Office

AI training for a project office runs in four steps.

1. Group walkthrough: a trainer works through a task on screen while colleagues discuss typical errors.
2. Independent practice on a real task: one person takes minutes, another takes a letter, another takes an edition comparison.
3. Peer review: employee A reads employee B's result and looks for omissions.
4. A new task without hints: each person brings a concrete result - an action-item list, a draft letter, an edition comparison, or a status summary.

Results are evaluated not for speed but for whether the employee caught the model's errors.

Three checking roles are distinct by subject. The employee looks for mechanical omissions: assignee not specified, date swapped, obligation added. The manager decides on substance: was this decision actually adopted, are the grounds sufficient? The authorized specialist checks compliance with project requirements and regulatory documents.

Four typical training checks: an action item with no deadline, a disputed decision between two people in the minutes, a moved clause with a changed appendix, and a file an assignee submitted but nobody accepted. For each, what the employee does and why is verified. Skill-check goals: the employee recognizes typical model omissions, distinguishes a suggestion from a decision, and does not substitute submission of a result for acceptance.

Benefit is counted over the whole pipeline: file preparation, work with the AI, error correction, and verification. Comparison with manual work at equal quality shows where time is saved and where extra control is added.

When the protocol is short, the task structure is clear, and manual errors are rare - AI is not needed. A three-sentence letter does not require a model. If volumes are large, documents are many, and items regularly get lost in manual processing - a separate automation project is worth exploring. That is separate research, not proof of payback. The manager decides after reviewing data-exchange options, access arrangements, and verification procedures.

## Frequently Asked Questions

### What to do if the minutes contain a disputed decision?

If two people expressed different views and no decision was explicitly ratified, it is not an action item. The employee saves quotes from both positions with their location in the record and passes them to the manager with a note: "decision not adopted." Taking part in a discussion does not grant the right to close the question independently. The authorized manager makes the decision with full context: what was discussed, who proposed what, and what the consequences are. Until the decision is made, the question stays open.

### Which documents are suitable for the first AI training task?

Meeting minutes that have already been processed and contain no commercial secrets. Files whose use is explicitly permitted by company policy. If a document is marked restricted, it is not passed to the model. Anonymizing data does not remove the restriction and is not a universal clearance - the question of whether a document may be used is resolved before the work begins, not after. For a first task, a routine document with a clear, verifiable result works best: a weekly meeting record is a better choice than a legal agreement.

### How to verify that an employee has learned the AI skill?

The skill is verified on a new task with changed conditions - not the one they trained on. The employee brings a concrete result: an action-item list, a draft letter, or an edition comparison. The result is evaluated by whether the employee caught the model's errors: missed action items, invented deadlines, added obligations. Verification is not complete until the result has been checked against the source material. The skill is confirmed when the employee recognizes typical omissions, distinguishes a suggestion from a decision, and does not substitute submission of a result for acceptance.

### Can a contractor letter be sent without approval?

No. The draft goes to the relevant specialist and the project manager - they decide on content and form. The firmness and legal precision of the letter are not determined by the employee. The approval sequence and authority to send are set by the organization. The employee's role is the mechanical part: transfer from the minutes, formatting, and consistency check. Sending letters and writing to systems is not part of this training method.

### When is AI training for a project office not needed?

If the protocol is short, the task structure is clear, and manual errors are rare - AI is not needed. A three-sentence letter does not require a model. If volumes are large, documents are many, and items regularly get lost in manual processing - automation is worth exploring. That is separate research, not proof of payback. The manager decides after reviewing data-exchange options, access arrangements, and verification procedures.

### How to compare document versions with AI?

The employee takes two editions of the same document and asks the model to list all differences: added, removed, and modified paragraphs, moved clauses, and updated references to appendices. Then they check how the changes affect active action items. Moving a paragraph from one section to another is not a new requirement. If an appendix that action items depend on has changed, the employee marks this and passes it to the specialist. The specialist determines the impact and decides what to do with the active action items.

### What does "change history" mean in a project office?

Change history is a record that is kept for every decision on an action item: what changed, the specialist's decision, the confirming person, the date, and a reference to both the old and new edition of the document. A later change does not overwrite the previous one - both remain in the record. The employee records what changed, who decided, and when it was confirmed.

## Conclusion

AI training for a project office works when it is built on real work tasks - not demonstrations and theory. The employee learns not to use an interface, but to check the model's output before it reaches the manager. This is a concrete, verifiable skill: it is confirmed on a new task with changed conditions.

*AI training for developer teams: [service](/ai-training/), [email](mailto:hello@majento.ai), [Telegram](https://t.me/shimaoz).*