An afternoon spent inside project records
That renovation project ran for three months, from the first consultation to the construction handoff.
Afterward, I wanted to reconstruct what had happened: what the client cared about at each stage, what changed across four revisions of the quote, which questions had been confirmed, and which remained verbal promises. The material included hundreds of messages, more than ten voice notes, notes from two meetings, and four versions of the electronic quote.
It took an entire afternoon.
There was no shortage of information. Each source simply had its own order. Messages followed send time, quotes followed revision history, voice notes contained temporary decisions, and new site conditions could change an earlier agreement. Search, transfer, and version checks consumed the time that should have gone into risk judgment.
That was why I began studying AI seriously. The problem in front of me was specific: could scattered information be organized into one structure, leaving my attention for the decisions that required experience?
I had the order wrong at first
Like many people, I initially collected "universal prompts." They could produce polished text quickly, but they knew nothing about my projects. They did not know why the word "charged separately" in a quote deserved another question, or who would be accountable for a verbal promise later.
A prompt detached from a real task produces work that cannot be accepted or checked.
I changed the process. When I handle similar records now, I first remove direct identifiers such as names, phone numbers, and addresses. Then I put messages, voice transcripts, meeting notes, and quote revisions into one task and ask AI to organize them around five fields: time, participant, point of dispute, missing information, and version difference.
The result is an index, not a conclusion.
I still return to the original records and verify each item. If a material price changed, was it due to a new specification, a quantity change, or an omission in the previous quote? If a process is listed, is it specific enough to inspect? Could "we will discuss it later" become an unaccountable extra charge during construction? AI can surface inconsistencies. I must decide the risk level and the next question.
How I decide whether the result is usable
I use four acceptance conditions for this kind of work:
- Important information is not lost through compression.
- Every summary can be traced to an original record.
- Uncertain points are marked instead of completed with invented answers.
- I review the final material and remain accountable before it reaches a client.
If one condition fails, the organized material cannot move directly into delivery.
With that division of labor, AI handles transfer, classification, comparison, and omission checks. I handle judgment, trade-offs, follow-up questions, and final expression. The tool did not complete the project for me. It showed me that experience held in my head could be turned into work with inputs, a process, and acceptance criteria.
Start with the task, then choose the tool
My learning order changed after that project.
I now begin with a recurring piece of real work. I define its inputs, required result, and final reviewer before deciding which step belongs to a tool. If a tool cannot improve actual delivery, I stop chasing it. If a process works again on the next project, I keep it and revise it.
I later documented the detailed quote-review process in a real AI-assisted workflow and the wider learning curve in three stages of AI adoption for a traditional-industry practitioner. Both articles are currently in Chinese.
The conclusion I took from this project is precise: AI's first value is making scattered work inspectable. Once experience is recorded, judgment can be used again in the next piece of work.