Scott Falconer
EssaysBookAbout
All essays
AI Governance··2 min·Series: Grounding AI

5 Things You Can Do Today to Ground AI (and Why It Matters)

Grounding a large language model, i.e. linking it to your own documents, is the fastest way to cut hallucinations, deliver accurate answers, and build trust.

5 Things You Can Do Today to Ground AI (and Why It Matters)

Grounding a large language model, i.e. linking it to your own documents, is the fastest way to cut hallucinations, deliver accurate answers, and build trust. It extends team knowledge, saves search time, and supports better decisions.

Start small, iterate often, follow 80/20 rule for quick wins. And remember, LLMs and remember, LLMs are trained on the most imperfect of docs (the web at large), so don't overthink it. Something beats nothing.

1. Start Small with What You Have

Curate the 20% of docs that answer 80% of questions. Pull your FAQs, checklists, and "how to...?" emails.

  • Do: upload 5-10 high-impact items to NotebookLM etc. and let the AI index them.

  • Don't: dump every archive folder on day one.

  • Today: list recurring questions and upload the matching docs.

2. Add Examples and Clarity

LLMs thrive on concrete scenarios.

  • Do: work an example into each doc, e.g., "Error 405 after a password change? Follow these steps..." Explain acronyms the first time you use them.

  • Don't: assume the reader (or the AI) shares your context.

  • Today: edit one doc; add a real-world example and spell out any shorthand.

3. Keep it Simple.

Headings, bullets, one topic per file, work better than a tome.

  • Do: caption visuals ("Figure 2: three-step approval flow").

  • Don't: hide answers in a 100-page "everything" PDF, split big files by topic.

  • Today: re-head a clunky doc and break it into smaller pieces if needed.

4. Group and Label Intuitively

Make it obvious where things live, and who they're for.

  • Do: create themed folders or notebooks ("Onboarding," "Discount Steps") and title files descriptively: "Internal - Discount Process - Q3 2025."

  • Don't: mix confidential notes with customer-facing articles.

  • Today: spin up one folder/notebook and move three to five docs into it with clear names.

5. Test and Tweak, then Keep It Fresh

A quick test run exposes gaps faster than any audit.

  • Do: ask the AI a handful of real questions that you know the answer to. See what it cites, and fix the weak spots.

  • Do: Archive duplicates; keep obsolete info only if you label when and why it applied ("Policy for v 8.13 - spring 2020 customers"). Plan a quarterly ten-minute sweep, ~30 % of data goes stale each year.

  • Don't: skip the test drive or wait for an annual doc day.

  • Today: upload your starter set, fire off three queries, and fix one issue you spot.

TLDR: Perfect can wait. Ground your AI with the highest-value 20%, keep the documents clear and labeled, and refine in small bites. You'll see faster answers, and more accurate responses.

Keep reading