Change-order first draft
A scope change rarely announces itself. The AHJ inspector wants a different conduit routing at the inverter pad. The owner's rep walks the array field and asks for a fence line to move. The...
The problem
A scope change rarely announces itself. The AHJ inspector wants a different conduit routing at the inverter pad. The owner's rep walks the array field and asks for a fence line to move. The geotech finds soft ground under a row of piers and the engineer says to go deeper.
The superintendent handles it that day, because the crew is standing there. The record of it lives in a text thread, a photo or two, and maybe a line in the daily log. Two weeks later the project manager sits down to write the change order and has to rebuild the story from fragments. Dates are fuzzy. Nobody remembers who verbally approved what. The extra labor hours were never tied to a cost code.
By then the other side has their own version. The change order becomes an argument instead of a document.
The gap is not that people don't care. The gap is that writing a clean narrative takes thirty focused minutes, and the field never has them on the day it happens.
How the agent workflow works
The agent reads the places where changes already show up. That means the daily log entries, the project text or messaging thread, the RFI log, inspection reports, and the photo folder for the job.
It watches for change signals: phrases like "not in scope," "owner wants," "inspector requires," "redo," "additional," or an RFI response that differs from the drawings. It also flags any log entry where crew hours land on a task that has no matching line in the schedule of values.
When it finds a likely change, it drafts a change-order first draft the same day. The draft contains:
- A plain narrative: what was found, who raised it, when, and what direction was given.
- References to the source material: the log date, the RFI number, the photos by filename.
- The affected cost codes, listed with the labor, equipment, and material items that appear to be involved.
- Open questions it could not answer, such as who gave the direction or whether the work is complete.
- A note on schedule impact if the log shows a crew or delivery was held.
The agent does not price the change and does not send anything to the owner. It sends the draft to the superintendent first, who corrects facts in a few minutes, then to the project manager. The PM owns pricing, markup, contract language, and submission. Nothing leaves the building without the PM's approval.
What you need to set it up
You need an AI model that can read documents and run on a schedule or trigger. You need read access to your daily log tool, your messaging thread export, your RFI log, and your photo storage. Most of this can be done with an automation platform and an AI subscription you may already have.
You also need your cost code list and schedule of values as a reference file, so the agent maps work to your codes and not generic ones.
Setup time is roughly a day of work: a few hours connecting inputs, a few more writing the instructions and a change-order template, and an hour or two testing against a past change you already know the story of. That last step matters most.
Where it breaks
It misses changes that were never written down. If the direction was given verbally and nobody typed a word, the agent has nothing to read. It helps only as much as the crew logs.
It over-flags. Ordinary rework and normal field adjustments will look like scope changes. Expect a pile of false alarms early, and expect the superintendent to get tired of them unless the trigger phrases are tuned.
It guesses at cost codes. When a task straddles two codes, the agent picks one, and it can be confidently wrong. The PM has to check the mapping.
It can state a cause too neatly. A text thread is messy, and the agent may turn a vague exchange into a clear instruction that nobody actually gave. Any line about who approved what should be checked against the source, not trusted.
It does not know your contract. Notice requirements, time limits for claiming a change, and what counts as a compensable event all sit outside what it reads unless you load them, and even then a person should decide.
The bottom line
This agent will not settle a dispute, but it moves the record from memory to paper while the details are still fresh. The superintendent spends minutes correcting a draft instead of the PM spending an evening reconstructing one. Start with one project, test it against a change you already lived through, and keep the PM as the only person who can send it.
Related use cases
- Two-week lookahead, built not typed Management
- Certification expiry tracker Management
- Schedule scenario planning, run before you commit Management