Stay at the helm
You direct the work and check it. You own every number.
AI in contaminated sites. The RemTech 2026 workshop.
Notes from our workshop at RemTech 2026, “AI in Contaminated Sites Work: A Practical Introduction”: how these tools work, where they break, and what to ask before you trust them with site work.
Paul Fuellbrandt, P.Ag. · Court Sandau, PhD · Jordan Yeo Use AI to find the facts. Keep the conclusions yours.
Trace every fact
If you can't trace it to the page, you can't defend it.
Build the tool
Use AI to write code you can check, not as the calculator.
Six things to watch for
Don't take its word as fact
AI isn't a database or a search engine. It writes the most likely-sounding answer, so it can be confidently wrong.
Ask what it actually read
Site files are too big for AI to read all at once. It only sees the passages its search picks out, so ask which ones.
Check how documents were read
Scans, handwriting and old tables are hard for software to read. One misplaced decimal can change a conclusion.
Don't trust its own confidence
AI will call a wrong reading 95% certain. Compare independent reads instead, and have a person settle the differences.
Don't let it do the math
Ask the same question 15 times: same answer?
Double-check well and site IDs
To AI search, MW-3610 and MW-3601 look 99% the same. Make sure the cited page names the same well.
Client data and chat memory
Chat apps can carry details from one conversation into the next. Turn memory off or use a separate project per client and site.
Questions to ask an AI vendor.
Use these with any AI vendor, or with your own IT team, before client data goes into a new tool. A vague answer is an answer.
| Ask | Red flag | A good answer |
|---|---|---|
| Where does our data go, and for how long? | “It’s in the cloud.” | A named region and a written retention period. |
| Is our data used to train models? | An opt-out buried in the settings. | No by default, and it’s in the contract. |
| How are answers cited? | A document name, or nothing at all. | Page and passage, so you can check the exact text. |
| What evaluations do you have? | “We tested it and it looks great.” | Benchmarks on real documents, re-run on every model change, with results you can see. |
| What happens when it’s wrong? | “It doesn’t make mistakes.” | It flags uncertainty, shows its sources and logs every step. |
| Which models do you use, and can that change? | Silence. | Named models, notice before they change, and re-testing when they do. |
Habits that hold up. Name the site and ask for dates and sources. Re-test old prompts when models change. Don't cite AI; cite the source it helped you find.
Part of our work with ITRC.
We contribute to the Interstate Technology and Regulatory Council's team on artificial intelligence and machine learning in the environmental field. Much of this workshop material is work we are contributing to that effort: helping practitioners keep up with fast-moving AI tools and use them responsibly.
Take control of your contaminated sites.
Bring us one site's documents. We'll show you what's in them.