Designing an AI-native content operating system
An internal CMS that turns a corporate's industry and org structure into a living skills taxonomy — and automatically generates the courses and assessments needed to close every skill gap.
Duration
Dec-April, 2026
Client
Wizr
Primary Skills
#
Product Strategy
#
Information Architecture
#
Cross-functional Collaboration
Product background
What it is. An internal CMS that lets a company (a "Corporate") onboard onto the platform, define its industry-specific skills taxonomy from the ground up, and then automatically generate the learning content — courses, videos, and assessments — needed to close skill gaps across every role in the organization. Why it exists. Before this system, translating "what skills does Retail Banking at Kotak actually need" into structured learning content was a manual, inconsistent process. The business had already mapped out how this should conceptually work — industries → sectors → functions → departments → concepts → skills → roles — but there was no interface for anyone to actually execute that model.
01
Corporate
Onboard a client, configure identity & branding.
02
Taxonomy
Build the skills taxonomy, manual or AI-assisted
03
Roles
Generate roles from skill bundles
04
Courses & Assessments
Generate or sync content against skill gaps
05
Learner content
QC'd content reaches the end learner
Who uses it. Internal Catalogue/Content Ops admins who configure corporates and taxonomies, and who monitor and QC the AI-generated learning content before it reaches end learners.
Understanding the problem
Business
The taxonomy model (Industry → Sector → Function → Department → Concept → Skill → Role) was fully specified on paper but had no interface.
The same taxonomy needed to power three downstream systems — Roles, Courses, Assessments — without duplicating data entry.
AI generation needed to be fast but reviewable; the business could not risk publishing ungoverned AI output directly to learners.
User
Admins needed to create and link six-plus levels of hierarchy without losing track of what depended on what.
No way to see, at a glance, whether a given Function or Department was ready, pending, or blocked.
Reviewing AI-generated content needed to be fast, bulk-capable, and reversible.
Operational
Course Provider Platforms, Products, Courses, and Corporates form a many-to-many web — one Course can serve multiple Corporates with different commercial terms.
Content generation for courses and assessments needed parallel but independently trackable progress.
Product
Nearly every entity moves through its own lifecycle, and those lifecycles gate each other across the entire system.
Complexity in managing content. The core difficulty was that this is a nested, stateful system: nearly every entity — Corporate, Segment, ISFD, Concept, Skill, Role, Course — moves through its own lifecycle (Draft → Pending → In Review → Locked/Active), and those lifecycles gate each other. You cannot generate a Role until ISFD is locked; you cannot generate a Course until Skills are locked. Designing legible status communication across all of this was the central UX problem.
Research and discovery
The inherited business logic was thorough on rules but silent on interface. Nearly every entity-creation flow followed the same shape: name → optional AI-generated description → optional image → save & lock → auto-generate downstream ID. That repetition was the first clue that a single reusable entity-creation template could replace what the flowcharts drew as six separate flows. "Exist? Yes/No" branches appeared at every level — a signal that search-or-create components were needed everywhere, not just once.
01
One template per entity type, reused everywhere — consistency over novelty at the atomic level.
02
Every AI action is reversible before it's locked — approve/reject queues and undo, never silent auto-commits.
03
Status is always visible, never inferred — every entity carries a badge, and drilling in never loses that context.
04
Progressive disclosure over long forms — collapsible, labeled sections instead of one giant scroll.
Who uses it. Internal Catalogue/Content Ops admins who configure corporates and taxonomies, and who monitor and QC the AI-generated learning content before it reaches end learners.
Information architecture
The core of the system is a single connected entity graph — everything downstream (Roles, Courses, Assessments) reads from and writes back to it.
Screens
Challenges
Status-state sprawl — with nearly every entity carrying its own lifecycle, keeping status legible without visual clutter on dashboards showing many entities at once was a continual balancing act.
Designing without a designer's usual research runway — the business logic was fixed before I joined, so intent had to be reverse-engineered from flowcharts rather than user interviews.
AI-in-the-loop trust — every AI-generation touchpoint needed a design answer for "what happens when the AI gets it wrong," across very different content types.
Commercial complexity — the many-to-many relationship between Products, Courses, and Corporates was one of the hardest pieces to represent without turning Course creation into a spreadsheet in disguise.
Solo design on a fast timeline — three months, one designer, a dense multi-module system; prioritizing which flows got full rigor was constant.
Outcome
60%
reported improvement in process efficiency across onboarding and content operations, per the Catalogue team.
Corporate onboarding — previously undefined and inconsistent — became a single guided, repeatable flow.
Course and assessment generation, previously disconnected manual processes, now run in sync off the same taxonomy data.
Reported ~60% improvement in process efficiency for onboarding and content operations.
Key learnings
