Redesigning Qi Studio so domain experts can author what engineers used to build, without taking the canvas away from the engineers who already lived there.
Qi Studio is GEP's AI agent and workflow platform. It already worked for engineers. The procurement experts who actually understood the workflows couldn't build them. This is the story of how we redesigned the front door so the people who own the work could finally do the work.
I co-led the UX for Qi Studio with another senior designer. Between us we held the IA, the interaction patterns, the conversational design surface, the visual system, and the design system tokens that hold the platform together.
My counterpart took the lead on the visual workflow canvas (Advanced mode). I took the lead on the conversational surface (Specifications mode) and the cross-platform interaction patterns. Most of the structural design — the dual-mode model, the primitive taxonomy, the checkpoint system — was authored together.
We worked alongside ML and platform engineering (who owned the AI behaviour), product management, and — most importantly — the procurement domain experts who became our north star.
We weren't designing into a vacuum. Qi Studio already shipped — with two dozen orchestrations, 151 connectors, a visual canvas, and a small purple AI helper called QiPilot in the toolbar. The platform worked. It worked for the wrong audience.
The tools catalogue was a wall of raw function names:
getchunksfromQdrant, pvaireviewContractDatagrid0013,
bulkUpdateContractMetadata. The IA was a developer's map. A sourcing
lead landing on this screen had no idea where to start.
A translation layer. Tools, skills, knowledge, and connectors got procurement-vernacular names (“Contract Clause Lookup”, “Compare Contract Versions”) and short, plain-language descriptions. The IA collapsed from seven workspaces to six clear primitives. The developer surface didn't disappear — it moved one tab deeper.
Twenty-four orchestrations already existed, hand-built one at a time. They sat in a card library, scrolled past, and got rebuilt the next time someone asked for something similar. No one was searching the library before building. There was no shared memory.
Capability discovery as the first move. When a user describes intent, Qi searches the existing library and reports back with confidence scores. The default recommended path: “Combine existing and create only what's missing.” We called it Compose Before Create.
The catalogue as it shipped — two dozen agents in a card grid, scrolled past and rebuilt the next time someone asked for something similar.
The same intent, answered before anything is generated: three of four capabilities matched, each scored.
Enterprise procurement is a compliance-heavy domain. “The AI did it” is not an acceptable answer when an auditor asks why a supplier was approved. Pure-prompt AI tools are black boxes by default. Our users needed inspection, not magic.
Every AI action shows its work. Confidence scores on reuse. Sentence-level diffs on spec changes. Checkpoints at every step. The principle: a domain expert should inspect what Qi did, why, and what changed — without learning to read code.
We couldn't remove the existing power surface. Engineering had built sophisticated agents on the visual canvas — with state passing, ReAct strategies, custom tools, model selection. They needed all of it. But every new feature shipped for engineering pushed procurement further away.
A two-speed builder. Same workflow, two cognitive modes. Specifications mode (conversational, fast, domain-vernacular). Advanced mode (visual canvas, model picker, system prompt editor). Test mode for QA. Not two products. One product, two reading levels.
These four challenges aren't independent — they're symptoms of one structural cost. Every time a procurement expert wanted to build a workflow, they paid a hidden tax: explain it to engineering, wait, review, ask for changes, wait again. The whole redesign is an attempt to take that tax to zero.
We didn't begin with sketches. We began with two parallel research tracks running for six weeks — watching the experts struggle, and interviewing the engineers who had been translating for them.
The procurement experts couldn't build. The engineers were spending most of their time translating. The translation work itself was the product opportunity.
Five connected design moves — and one cross-platform pattern that ties them together. Everything from here is the design response to The Translation Tax.
Same artifact. Two cognitive modes. One QA gate. Specifications is the new procurement-friendly entry. Advanced is the existing developer canvas — preserved, not replaced. Test is the gate where both audiences meet.
Chat on the left. Artifact on the right. Both live.
The Specifications surface is a two-pane layout that does something subtle: it makes generative AI legible. The user watches the agent assemble in real time — Qi paraphrases intent on the left, while sections of the spec stream into existence on the right. Pure-chat AI tools hide the artifact in the conversation. Pure-builder tools strip away the narration. The two-pane model gives users both.
Search the library first. Generate second.
Before Qi generates anything, it searches existing agents, skills, tools, and knowledge — and returns reuse candidates with confidence scores. The Recommended path is always to combine existing pieces with only what's net-new. Compose Before Create is the anti-sprawl pattern expressed in the interaction itself.
Git, but for natural-language specs.
Versioning code is well-understood. Versioning natural-language agent specifications — where a single sentence can change behaviour — is not. Every meaningful step in the build flow generates a Checkpoint with a View Changes affordance. Opening it shows sentence-level diffs: Modified, New, Removed. Read like an email amendment, not a code commit.
Surface every gap during construction.
The old Qi Studio would let you save an unconfigured orchestration and find out at runtime that a connector was missing. The redesign surfaces every structural gap during build — missing connectors, ambiguous inputs, untested edges — with inline warnings and Configure CTAs. The agent cannot ship until the system has called its bluffs.
One taxonomy. Six pieces. Many workflows.
Underneath every authoring surface is a single architectural taxonomy. Agents are made of Skills. Skills use Tools and Knowledge. Tools and Connectors talk to data. Widgets surface results. Each primitive gets its own catalogue page, its own procurement-vernacular naming, and — critically — its own two-speed builder.
Same pattern. Different primitive.
The Specifications surface authors agents, skills, tools, knowledge, widgets, and connectors — every primitive — with the same chat-left, artifact-right composition. The grammar is the platform. Add a new primitive type in the future and the platform absorbs it without a new builder paradigm.
Engineers didn't lose their tool.
The Advanced canvas is the original Qi Studio orchestration editor — refined but intact. Node-and-wire workflow on a dotted grid. Per-agent versioning (v2.1, v1.8, v2.3). Model picker exposed. System prompt editor with markdown, state references, and ReAct strategy selection. Anything authored in Specifications opens cleanly here. No conversion. No data loss.
Legibility before personality.
We anchored the visual system on legibility before personality. The result is intentionally calm. The interesting things happen in the artifact pane and the chat — the chrome stays out of the way. Typography, status colors, iconography, tags, pills, and multilingual scaffolding all serve one goal: help procurement experts inspect what Qi is doing without making the product feel technical or intimidating.
The redesign is rolling out in waves across GEP's customer environments. We don't have a full quarter of measured outcome data yet — but the early signal is qualitative, observable, and consistent.
I used to file a ticket and wait three weeks. Now I describe what I want and watch the agent appear on the right. The first time it surfaced an existing agent I didn't know we had, I felt like the platform had finally caught up to me.
Category Manager · Internal Pilot
Without engineering hand-off, in early pilots. The translation step is collapsing exactly where the research predicted.
Adoption of Advanced mode is unchanged among the engineering audience. The two-speed model didn't force a migration — it added a lane.
The 24 legacy orchestrations were rebuilt from scratch in 80% of new requests. Early pilots suggest Compose Before Create is meaningfully bending that curve.
Framed against industry benchmarks — targets, not measured results.
Industry benchmark for procurement IT-led workflow buildouts. Target with conversational authoring: hours to days on the happy path.
Enterprise no-code AI platforms typically need 40+ hours of training. Targeting under 2 hours through native procurement vocabulary.
Legacy baseline was approx 20% reuse on new requests. Compose Before Create targets 60%+ within the first year of rollout.
The single design move that organised everything else. Reframing QiPilot from a toolbar button to the primary surface unlocked the rest of the system.
The hardest stakeholder wasn't procurement — it was the engineering team who built the original canvas. Designing with them kept their work load-bearing in the new platform.
The conversational + canvas pattern scaled across every primitive type with no architectural rework. The grammar was right.
Watching a domain expert review a checkpoint diff and catch a regression was the moment the system felt trustable.
The brief was almost never to add capability — it was to widen access without breaking the existing audience.
The translation step between procurement and engineering was invisible until we interviewed both sides in parallel.
Trust comes from inspection — confidence scores, diffs, checkpoints, build-time warnings. The conversation is the entry; the artifact is the contract.
Renaming getchunksfromQdrant to “Contract Clause Lookup” is the difference
between a tool a procurement person adopts and one they file a ticket about.
Kirana Store Admin System & B2B Order App
read the case studythinking of something that could make the world a better place? let's talk.
free 1:1