Last year, the organization that supports construction safety in B.C. sat down in Kelowna to talk about AI. The question was not whether the industry should be interested in it. It was what, if anything, it can do for safety work.

On September 23, 2025, the BC Construction Safety Alliance held an AI in Construction Safety breakfast at the Delta Hotels Grand Okanagan Resort, in the Columbia Room. The session, led by David Dunham, was one of four BCCSA breakfasts he ran around the province. The description is specific: an interactive session "with demonstrations of AI-supported safety documentation and communication," where participants "explore practical uses while considering privacy and responsible implementation."

That is a better question for a Kelowna GC or trade contractor than a general talk about AI transforming business. Safety paperwork, toolbox talks, inspections and follow-up already take time. If a tool can help with the writing and the tracking, it is worth understanding. If it starts to stand in for judgment, it is a problem.

Where the question came from, and the limits of the sources

A breakfast session is not a survey. It tells you what BCCSA thought local contractors and safety people needed to see, not how many Okanagan firms have tried a tool, or whether incidents went down afterward. Dunham is a safety consultant who teaches these sessions, so he has an interest in people using the methods he demonstrates. This post does not quote demonstrations that were not published, and it does not treat his later conference talks as if they were the Kelowna session. The Kelowna event, as described, was about documentation, communication, privacy and responsible use.

Other local events have asked a broader question. In May 2025, UBC Okanagan's Faculty of Management and the Kelowna Chamber of Commerce hosted a session on AI for small and medium-sized businesses, with a keynote from Dr. Wendy Cukier on what she has called the Canadian AI paradox: a country that is strong at building AI tools and slower at adopting them. That is useful context. It is not a construction safety finding. A 2024 WeBC webinar listed by the Central Okanagan Economic Development Commission warned that used poorly, AI can make you "sound like a robot," spread bias, or put customer data at risk. Again, supporting context, not a job-site study.

The region's adoption story is already covered in Why the Okanagan Lags Vancouver. This post does not repeat that argument. It answers a narrower question: if you are going to use AI at all in a Kelowna construction business, what can it do for safety paperwork, and where does it have to stop?

Why safety documentation is the right first place to look

BuildForce Canada's 2026 to 2035 outlook now treats the Thompson-Okanagan as its own sub-market, including Kelowna, Kamloops, Penticton, Salmon Arm, Vernon, Golden and Revelstoke. In that region, residential construction employment is projected to fall 11% from 2025 levels by 2035, while non-residential employment is projected to rise 6%, helped by ICI buildings and maintenance work.

That is a different picture from the Island or the Lower Mainland. The work that is expected to grow is the kind that comes with more formal safety systems, more contractors on the same site, and more records. When experienced people are also leaving the trade, as they are province-wide, the quality of those records matters more, not less. A missed follow-up or a toolbox talk that never quite gets written down is how a system looks fine on paper until it is not.

AI does not fix that on its own. It can help with the part that is writing and organizing, if a person still owns the call.

What AI can reasonably help with

The Kelowna session focused on documentation and communication. That is the right bucket. The jobs that fit are the ones you already do in a set format, where a bad first draft is easy to see and cheap to correct.

These are drafts and lists. They are not a finding that a site is safe to start, and they are not a substitute for a supervisor walking the work. If the input is thin, the output will sound finished anyway. That is the usual failure mode, and it is why a person has to read the result. The general version of that habit is in Don't Believe Everything the Robot Says.

A safety document is not finished when it sounds professional. It is finished when a person who was there can stand behind it.

If you would not let a new coordinator send the document without a review, do not let a tool do it either.

What still stays with you

WorkSafeBC's rules did not change because a new writing tool showed up. Supervisors still have duties to ensure the health and safety of workers. Due diligence still depends on what you actually did, and on whether you can show it: orientation and training, inspections and corrective actions, incident investigations, safety meetings, and the other records described in WorkSafeBC's due diligence guidance.

That is the boundary. A tool can help you write and organize those records. It cannot take the responsibility for them. It cannot decide that a hazard is acceptable. It cannot inspect a scaffold, run a lift review, or have the conversation with a subcontractor who is rushing a task. If a draft procedure is generic, or if it misses a condition that is obvious on your site, publishing it does not make you diligent. It makes you the author of a document that does not match the work.

Privacy belongs in the same conversation, and BCCSA put it on the Kelowna agenda for a reason. Do not paste incident details that identify workers, medical information, or owner documents you do not have the right to share into a public chat tool. A safety file is not marketing copy. If you are not sure whether a record can leave your system, it stays in your system.

The same caution applies to anything that looks like a decision. "The site is ready," "this is an acceptable workaround," or "we can start" are not sentences a tool should be allowed to generate and send. A person who was there has to write those, or at least sign them.

How to try this without fooling yourself

Start with one document type, not a new safety program. Last week's toolbox talk and this week's topic is enough. Give the tool your format and your notes. Time the draft and the review. If the review is longer than writing it yourself, stop. If the draft is useful, keep the prompt, and do not expand into incident reports or procedures until the simple version is reliable.

Keep a human checkpoint that cannot be skipped. Name the person. If nobody is willing to put their name on the output, the task is not ready. For a wider discussion of what software should not be allowed to do on its own, see Your Software Can Now Act Without You.

Do not measure success by how many documents the tool produced. Measure it by whether the record is accurate, whether follow-ups actually closed, and whether the people on the site still recognize the work being described. A tidy folder of generic talks is not a safety system.

One more distinction is worth making, because it is easy to blur after a polished breakfast demo. Helping a safety coordinator write better records is not the same as putting AI into the field to watch the work. Cameras, automated hazard flags and similar products exist, and vendors will sell them. They are a different decision, with a different privacy and labour conversation, and they were not what the Kelowna session, as described, was there to settle. If your first experiment leaves the page and starts watching people, you have jumped past the question this post is answering.

The same is true of training. A tool can turn a procedure into a first draft of a talk or a quiz. It cannot tell you whether the crew understood it, or whether the procedure matches the site as it stands this week. That check is still a conversation, usually held in person, before the work starts.

What to do with this

Take the safety document you write most often. Collect two good examples, plus the notes for the next one. Use a general AI tool to draft the next version in your format, then review it as if a new hire had written it. Correct the facts, cut the filler, and only then decide whether it saved enough time to do again.

Write down, in a few lines, what the tool is allowed to draft and what it is not allowed to send. Include personal information, incident details and any statement that a task is safe to start. Share that list with the people who would actually use the tool.

If you do only that, you will have answered the Kelowna question in the only way that counts: on your own paperwork, under the same responsibilities you already have.

Explore the assessment →