Redesigning Delta for Business as one connected platform for travelers and managers.

The business-model redesign was joint with my teammate; everything below was mine.

Before the redesign, Delta's corporate offering was scattered across three disconnected web properties and two separate tools. The first research job was auditing what already existed and understanding why users got lost in it.

Moderated tests and contextual research surfaced the core problems of the existing web and app:

I started with Delta's own metrics next to Qatar, Emirates, Singapore, United and American. The differentiation gap was clearest around premium-service consistency and loyalty mechanics for high-spend corporate accounts.

Jorge Jimenez Rolo (SEO & Behavioral Manager), Najah Bradley (frequent traveler), Sarah Kling (Sr. UX Researcher, Amazon), Stephanie Krell (Service Designer & Professor), and Stehle Craig (Industrial Designer & Professor).

I mapped the end-to-end boarding & travel journey to find where status decisions and friction collide. Friction points are highlighted.

Instead of personas, I defined the audience as adoption archetypes by company size and placed them on Rogers' Diffusion of Innovations curve, asking which organizations adopt a new B2B loyalty platform first, and which hold back. Each archetype was read against Rogers' five attributes: relative advantage, compatibility, complexity, trialability and observability.

Startups scored 20/25, ready to adopt, sitting with the Early Adopters. Mid-Market travel managers scored 14/25, not yet, holding back in the Late Majority. Small businesses adopt fast for clear ROI, while large enterprises carry the highest value but move slowest, gated by procurement and security. Observability was the weakest attribute across the board (2/5). Teams can't easily see other companies' results, which is exactly why "add customer case studies" became a P0 fix.

A new business-tier system designed to make high-spend corporate accounts feel rewarded inside Delta rather than penalized by it.

I built a tokenized design system in Figma — 137 tokens for color, type, spacing, radius and shadow, extracted straight from Figma variables — and wrote the instruction system a 5-agent Claude Code team read before touching a file: 8 markdown specs (sitemap through agent instructions), a 1,172-line CLAUDE.md, and an orchestrator prompt. I didn't hand-write the site's TypeScript. I wrote the constraints that made the agents' output predictable, and made the visual judgment calls the agents couldn't.

34 commits over 15 days: 21 features, 13 fixes. The fixes are corrections to what the agent team had already shipped, and the pattern holds across all of them — the agents got the structure and logic right (the right component, the right route, the right data), and missed the calls that are pure visual judgment. Those 13 fix commits are the closest thing this project has to my fingerprint.

None of these were logic bugs — tsc passed, the token grep passed, the build passed. They were the fix commits doing the job the verification gates couldn't: judging whether it looked right.

I unified three Delta properties (Business, SkyMiles for Business and Delta–AMEX Business Cards) into a single hub, with a hard line in the information architecture: everything business and corporate, we build; general consumer features (book a flight, personal SkyMiles, flight status) link out to delta.com. Enrollment is self-serve for small business and contact-sales for mid-market and enterprise.

Rich snippets as shortcuts, and better ranking. Every page ships JSON-LD for BreadcrumbList, SiteNavigationElement and FAQPage. Google renders that markup as rich-snippet sitelinks that act as shortcuts straight into the funnel, and the structured data also improves the page's ranking. One result becomes a column of links dropping users directly onto Programs, Cards or the tools, and it surfaces above competing carriers.

The hierarchy stays shallow, at most three levels below the homepage. Sections (Level 1) open into sub-pages (Level 2) and, only where needed, detail pages (Level 3). A flat tree keeps every page a few clicks away and gives crawlers a clean, predictable path. Each node below is tagged with its level. Drag, scroll with two fingers, or open a node to read what it does.

The live Delta for Business landing page came out of the same system described above: the SEO / GEO markdown specs, the tokenized React components, the agent-ownership boundaries, and the verification gates between phases. Below are captures of the live pages — open the full site in a new tab to click through it yourself.

The Benefits Calculator was built independently in Figma Make by Brian Glennon. It matters to this case study as the trigger for the broader progressive-disclosure / guided-selector work that now sits on the public site.

The numbers below are real and verifiable in the transcripts, except where marked projected.

What surprised me about the agentic build was where the split actually fell. I expected to spend my time on structure and let the agents handle the small judgment calls. It was the other way around. Five agents with clean file-ownership boundaries produced 63 correct pages and 55 working components without me touching the file tree — routing, data shape, component boundaries, all of that held up. What they couldn't do was see their own output: a red stroke reading as an error state, a grid with the wrong proportions, a badge on the wrong background. Thirteen of 34 commits were me fixing exactly that, not bugs. Agentic build earns its place anywhere a rule can be written down and mechanically checked — tokens, types, routes, metadata. It doesn't earn its place anywhere the only real test is "does this look right," which is most of what a design system is actually for. What I took from the project as a whole: owning structure across three connected products at once; running a tokenized design system and an agentic build pipeline end-to-end while knowing exactly which decisions I couldn't delegate; separating what users said from what we interpreted; and knowing when to lead, when to delegate, and when to absorb work to protect the deliverable.