It can draft one from what you recorded. It cannot see the site, and a report that blurs what was observed with what was inferred is a professional problem long before it is a writing problem.
Key takeaways
Yes, for the writing. A site visit produces a scatter of inputs — a field notebook, forty photographs with timestamps, a dictated voice memo on the drive back, a checklist half-completed on a tablet. Turning that scatter into a structured report with a consistent heading order, item numbering, locations and photograph references is a formatting and summarisation task, and it is one of the better uses of AI in technical practice because the inputs are all yours.
What AI cannot do is any part of the visit. It did not stand on the deck, it did not notice that the second-floor guard was loose, and it has no basis on which to say anything about the site that is not already in what you gave it. The failure mode is not that it refuses; it is that it fills gaps smoothly, and a smoothly filled gap in a field report is a fabricated observation.
Engineers Canada's national guideline on the code of ethics sets out a principle that reads like it was written for this problem. Principle 4 asks registrants to provide professional statements that distinguish between facts, assumptions and opinions, and explains why: registrants are consulted for their expertise and judgment, each of the three carries a different level of certainty, and if they are not differentiated clearly, serious consequences and misunderstandings can result. Principle 2 of the same guideline asks registrants to practise only in areas of their competence, carefully, diligently and with honest conviction.
Generative text does the opposite of Principle 4 by default. Given “cracking noted at column base” it will happily produce “cracking at the column base consistent with differential settlement”, which is an opinion you did not form, presented in the same register as the observation you did. If you use a tool for field reports, the single most valuable instruction in your template is a hard separation: an Observations section that may contain only what was recorded, and a separate Assessment section that a registrant writes.
One caveat about that guideline, stated on the page itself: Engineers Canada's national guidelines are not regulations or rules, they do not establish a legal standard of care, and the ultimate authority on any specific practice lies with the regulator in the province or territory where the engineer works. In Alberta, for example, APEGA publishes practice standards, bulletins and guidelines under the mandate of the Engineering and Geoscience Professions Act and General Regulation, and its own page is explicit that a practice standard sets the minimum standard against which a permit holder's or licensed professional's practice and conduct will be assessed by APEGA's statutory boards. Check your own regulator before you build a template around a national document.
The safety-inspection literature is more prescriptive than most engineering guidance and it is a good skeleton for a drafting template. CCOHS's guidance on effective workplace inspections sets out what the final report should contain: an introduction naming the department or area inspected, the date and time frame, and the inspection team's names and titles including any technical experts; then observations listed and consecutively numbered, with a unique item identification tag that includes the year and month; then a classification of each item by hazard type, with a note of whether it is a recurring issue.
Two of its instructions are the ones a draft most often fails. The first is specificity: instead of “machine unguarded”, state “guard missing on upper pulley #6 lathe in North Building”, because a person who was not present should be able to find the same item from the information provided. The second is linkage: hand-written notes, checklists and diagrams should be included or scanned into the final report, and photographs must be linked to the correct items. Both are checkable mechanically, which makes them ideal things to have a tool verify before a report goes to review.
A mid-sized consulting firm in Ontario runs site attendances across a dozen active projects. Reports go out three to nine days after the visit because they are written in the evenings, and the delay is where the risk sits: a nonconforming item observed on a Tuesday and reported the following week has had a week to be covered up by subsequent work.
The firm changes the sequence rather than the writing. Field staff dictate on site against a fixed prompt structure; the draft is generated the same afternoon into the firm's template with the photographs attached and the specificity check run; the reviewing engineer edits and issues within twenty-four hours. The Assessment section stays empty until the engineer writes it.
The measurable is the interval between attendance and issue, and the count of items that had to be re-observed because the first report could not be located on site. Both come straight out of the document register. Neither is an estimate.
Field records are frequently the only contemporaneous account of what a structure looked like at a point in time, and they surface years later in disputes. Treadstone Law's article on suing an engineer or architect for construction defects in Ontario walks through how those records are read in litigation, and its shorter answer on negligence claims against engineers and architects covers what a claimant has to establish. The practical consequence for your template is that an AI-smoothed sentence which converts an observation into a conclusion is not a stylistic choice. It is a statement you will be asked to stand behind under cross-examination.
The same logic applies to the site itself: liability for what happens to people on a construction site has its own rules, set out in Treadstone Law's note on liability for injury to a construction site visitor in Ontario.
Can AI generate observations from photographs?
It can describe what is visible in an image, and that description belongs in the draft as a prompt to the reviewer, not in the report as an observation. The registrant who attended is the one who observed.
Should the report say AI was used?
There is no national requirement that it should. Decide it at firm level, be consistent, and be able to describe your process if a regulator or a court asks how the document was produced.
What is the highest-value automation here?
Repeat-item detection. Recurring unresolved items are the pattern that matters and the one nobody has time to reconstruct by hand across eighteen months of reports.
A 30-minute call is enough to tell you whether AI pays for itself here.
