Case study: Running an AI-native content system as a service
BrightHire is an AI-powered interview intelligence platform, now part of Zoom, that records and analyzes hiring interviews to help teams make faster, fairer decisions.
I built and ran an AI-native content system as a service for BrightHire: 25 full-length pieces across three strategic programs, produced end to end through repeatable, gated workflows.
The challenge
BrightHire created the interview intelligence category, but the market conversation was moving to a newer, faster-growing term: the AI interviewer. Ownership of that term was being decided in real time in ChatGPT and Google search results, but BrightHire had no content covering it.
This topic demands precision. AI hiring is actively regulated in the EU and several US states, so every claim has to hold up. The highest-value pages named competitors head-on. And the company’s own product facts were scattered across internal docs that didn’t always agree.
The engagement itself added a third constraint: the team ran async and time-boxed, with no clear access to subject-matter experts, plus restricted access to internal systems. My method had to produce first and clarify later, without ever publishing an unverified claim.
The work
A truth-first foundation. Before any drafting, the client’s existing material (internal messaging docs, stakeholder transcripts, live site pages) was synthesized into a foundation layer:
- a knowledge base
- a glossary of approved terms
- a voice and tone guide
- anti-AI writing patterns
- and a truth register in which every marketable claim carries a stable ID and a verification state.
Contradictions between sources went into a conflict register instead of blocking production, so pages could ship with unresolved questions hedged or omitted rather than guessed at. The register grew to 85+ tracked claims, each traceable to the pages that use it, so when a fact changed, every affected page was findable in seconds.
Three content programs, gated end to end. Twenty-five full-length pieces across three programs with distinct strategic jobs:
- a 10-page AI Interviewer cluster playing offense on the growth term
- a 9-page candidate fraud cluster capturing an emerging buyer concern
- a 5-page core category cluster defending the interview intelligence term BrightHire coined.
Every page was structured for answer engines — an extractable 40–60-word lead answer, question-shaped headings, schema — and ran through a sequential gate system covering meaning, accuracy, coherence, craft, and voice.
The hardest class, named-competitor comparison pages, carried an extra gate stack: every competitor claim sourced to the competitor’s own public materials with a last-checked date, a mandatory “where they’re strong” section, and legal review before publish.
Research fed the system under the same discipline: six independent deep-research passes ran on the fraud cluster alone, and every one of them corrected or killed a claim from the planning briefs before it reached a page.
A voice that compounds. Each human editorial pass was diffed against the canonical draft, and every recurring correction was harvested into the style guide as a durable rule. One editing cycle produced roughly a dozen permanent voice rules; later pages arrived needing less correction than earlier ones, because the system had learned the client’s voice from my manual editing.
Results
- 25 full-length pieces produced (~88,000 words), spanning category education, compliance, candidate fraud, competitor comparisons, and buyer’s guides
- Nine pages live on brighthire.com within six weeks of onboarding, including four published in a single day
- The legally hardest page class shipped: BrightHire vs. HireVue and HireVue Alternatives cleared competitive-truth and legal review, built on reusable templates so future competitor pages start from a gated skeleton rather than a blank page
- The category hub, AI Interviewers for Hiring: The Complete Guide, anchors the cluster every spoke links into
A content operating system I’d already proven in-house ran here as a service, on someone else’s brand, under legal review and regulatory constraint, at a pace the constraint was supposed to make impossible.
The bottleneck on every page was verification and review, not writing, which is the point. When drafting is no longer the constraint, the advantage goes to whoever can verify fastest.