We use AI in our own work every day.

These are the AI tools running in our daily practice — for drafting, code, research, and operations. Tool-specific knowledge decays in roughly 90 days, so the list below is dated; check the footer's last reviewed stamp before you assume any of it is still current.

• Claude Code (Anthropic) — the primary surface for any work that touches code, scripts, or technical documentation. Used for the verification scripts that gate this site's deploys, The Bearing's engine code, and most of the build infrastructure under scripts/ across our active projects.

• Claude Pro + Projects (Anthropic) — the primary surface for client-facing drafting: case studies, training-pack copy, internal memos. Voice tuning happens in Project-level memory; the engagement-specific Project carries the customer's voice descriptors as instructions.

• MCP integrations (multiple) — Model Context Protocol servers connecting Claude to working tools: Obsidian for the project knowledge base, Firecrawl for source-checking research claims, Sequential Thinking for multi-constraint planning, scheduled-tasks for recurring operations. The integrations are the infrastructure that lets Claude act on real artifacts rather than just describe them.

• Microsoft Copilot for M365 (Excel, Outlook, Word) — for the work that lives inside the suite where the artifact already is. Variance recaps in Excel, draft-then-edit in Outlook, structured memos in Word. We use it because clients use it; the tool selection follows the workflow, not the other way around.

Specific artifacts shipped through the tools above. The point isn't that AI helped — that's a claim. The point is naming the artifact so the claim is checkable.

• This site's verification pipeline. Three pre-build scripts gate every production deploy: check-verifications.mjs for inline-claim flags, verify-hreflang.mjs for EN↔ZH locale-page reciprocity, verify-en-zh-parity.mjs for translation completeness, and check-freshness.mjs for the 90-day decay flag. Drafted with Claude Code, owned by Paul, run on every commit.

• The Bearing — a 60-question diagnostic engine. The recommendation engine that drafts the Phase 2 strategy pack from a discovery-call intake. 4D AI mastery questions (verification habit, manager visibility) drive section-level branching in the output: when manager visibility is absent, the engine recommends a smaller manager-first engagement instead of standard team training. The structurally honest version may scope smaller than a typical contract — that's by design.

• Case-study drafts on this site. Each Tier 1 case study was drafted through Claude Pro from source-material notes, then edited line-by-line by Paul before publication. Where a study cites a quantitative outcome, the citation links to disclosed methodology — same vendor-claim discount rule we apply to vendor pitches.

• The research-grounded scaffolds. Schema.org JSON-LD on every page, llms.txt at the site root, freshness-stamp discipline on every page that names a tool. Research from the 2026-05-05 AI-mastery memo argued for verification scaffolds over skepticism training; we shipped the scaffolds in our own infrastructure first, before recommending them to clients.

The interesting question isn't where AI shows up. It's where it doesn't. Three categories stay human in our practice; they'd be the same three categories in yours.

• Judgment calls. Whether a client is the right fit for an engagement. Whether a recommendation is structurally honest or just lucrative. Whether a draft reads true or just clean. These are the decisions that determine whether the work is good — and they sit with the practitioner, not the tool.

• Regulated decisions. Anything touching client data with regulatory weight (PIPEDA, GDPR, Quebec Law 25, sector-specific rules) goes through our governance frame, not through a generative model. AI assists in drafting a memo about the rule; the rule itself is human-read, human-checked, human-cited.

• Voice. The voice in the case studies, the voice in the proposals, the voice on this page — those are Paul's. AI assists in drafting and structure; the voice itself is checked sentence by sentence on every public artifact. If a sentence feels generative, it gets rewritten or cut.