HH Global / Noosh · 2024 – 2026
Designing AI into the procurement workflow
An AI companion, smart search, spec creation, AI email and an agentic mode for a B2B platform that manages $1.6B+ in marketing spend — designed around what users needed, not around the technology.
Overview
Noosh is HH Global's marketing-execution platform: enterprise teams use it to spec, source, order and pay for print and marketing materials at scale. The workflows are dense, form-heavy, and repetitive by design.
Starting in 2024, I designed a new layer of AI across the product: a natural-language AI Companion, Smart Search across every record type, AI spec creation, AI-assisted email, and an agentic mode that carries out multi-step tasks on the user's behalf. The question that drove all of it wasn't "where can we put AI?" but "what do our users actually need, and how do we make it usable?" — for a procurement lead who lives in the product and a marketer who opens it twice a quarter alike.
My role
Designer, from discovery to delivery
I was the designer on the AI work end to end — part of the earliest opportunity and discovery conversations, with a voice in where AI could make our users' lives easier and where it wouldn't. From there I did the research, designed every feature from scratch, built the design system they follow, tested with real users, and implemented UI changes directly in the React codebase alongside engineering.
- Opportunity & discovery: mapped end-to-end user flows with the product team to find the steps where AI would remove real effort — and spoke up about the ones where it wouldn't
- Research: user interviews, usability testing, and competitive analysis of AI assistants in adjacent enterprise tools
- Design: every AI feature, from concept to hi-fi, plus the states AI introduces (thinking, validating, confirming, failing)
- Design system: components, patterns and rules that keep AI surfaces consistent with the rest of Noosh
- Delivery: partnered with the HH Global AI R&D team and front-end engineers; shipped UI changes in code
Discovery
Start with what users need
The tempting move was a chat window. I pushed against that. Our users aren't looking for a conversation — they're trying to get a project spec'd, sourced and ordered with as few steps as possible. So the first design task wasn't a screen, and it wasn't "where does the AI go"; it was understanding what people were struggling with.
Three problems kept surfacing in interviews. People couldn't get a succinct answer to "what do I need to do today?" without hunting across projects. Small issues — a date slip, a missing file, a quick question to a supplier — took a full email cycle to resolve. And the core flows, creating a project or writing a spec, were long, click-heavy processes that punished anyone who didn't do them every day.
I walked the core flows step by step and asked one question at each point: is the effort here about judgment, or about mechanics? Judgment stays with the user. Mechanics — finding the right record, re-entering data that already exists, turning a brief into a structured form, writing the same update for the fifth time — are where AI earns its place.
A simplified view of the workflow. Highlighted steps are where mechanical effort was highest — and where the first AI features went.
The other constraint came from who uses Noosh. A procurement lead who lives in the product all day and a marketer who opens it twice a quarter have opposite needs — one wants speed, the other wants guidance. Every AI feature had to work for both without a "beginner mode": the same surface, with the AI carrying more of the load for whoever needs it.
Principles
Four rules that shaped every feature
The user's task comes first, the AI second
Every feature started from a real thing someone was trying to get done. If AI didn't make that easier, it didn't ship — however impressive the demo.
Inside the workflow, not beside it
AI actions land the user in the real Noosh screen with values pre-filled — never in a separate AI silo they have to copy out of.
Show, draft, confirm
Every AI surface opens with what it can do and concrete examples. AI drafts and pre-fills; the user confirms before anything is created. Validation is visible, not silent.
One surface, novice to expert
No separate "easy mode". A first-time user gets guided through; a power user types one line and moves on. Same components, different amount of help.
Design system
Making AI feel native
Noosh had grown a lot of UI over the years. Adding AI on top without a system would have made the fragmentation worse, so I built one first — and then designed every AI feature within it.
Prompt bar, action chips, result cards
A single prompt pattern reused everywhere, with chips that scope intent ("Create Project", "Copy Spec") so users don't have to guess the right phrasing.
Thinking, validating, confirming, failing
AI introduces states classic forms don't have. Each got a defined treatment so the product never feels like it's stalled or guessing.
What AI touched
Anything drafted or pre-filled by AI is visibly marked until the user accepts it, so trust is earned field by field.
When AI may act
Read-only actions run immediately; anything that creates or changes a record requires confirmation. The rule is baked into the components.
The system itself is internal, so it's described rather than shown here.
The work
Five features, shipped
One place to ask, across the whole product
The Companion turns natural language into real actions: Smart Search across eight record types (projects, orders, RFEs, RFQs, invoices, quotes, estimates, specs), creating and copying projects, copying specs, drafting messages, pulling a task list, running reports.
Released early 2025; a second version followed mid-2025 based on how people actually used the first.
AI Companion V2 — public release demo.
Open with capabilities and example requests instead of an empty chat box.
In testing, an empty prompt froze people — they didn't know what was "allowed". Showing what the Companion can do, with copy-able examples, got first-time users to a successful request far faster.
Search results come back as structured tables and charts, not as a paragraph of text.
Our users compare, filter and export. A chat-style answer would have been a dead end; structured results plug straight into the way they already work.
Action chips scope the request ("Copy Project ×") before the user types.
It removes the phrasing guesswork, sets expectations for what will happen, and lets the AI validate against the right fields.
"My Tasks" became a first-class Companion action, not a search you had to phrase.
The most common thing people wanted was a succinct answer to "what do I need to do?" Making it a one-tap action — defaulting to everything due in the next week — solved the top request in interviews without anyone learning a query.
From a brief to a structured spec
Specs are the backbone of print procurement and the most form-heavy thing in the product — a long, many-step process that experts tolerate and newcomers get lost in. AI now drafts the spec from a plain-language request, and the user reviews it in the familiar spec form. The expert saves the clicks; the novice gets a working starting point.
AI Spec Creation — public release demo.
The AI drafts into the existing spec form rather than a new "AI spec" view.
Users trust the form they know. Keeping the draft editable in place meant zero retraining and made it obvious that they, not the AI, own the final spec.
Pre-filled fields stay visibly marked until reviewed.
A spec with a wrong quantity is expensive. Marking what the AI filled makes review deliberate instead of a skim.
Communication that writes itself from context
A large share of a project manager's day is small updates, requests and follow-ups. Email with AI lets users address a quick issue from inside the project — the draft is written from the project's own data, so it's accurate before the user even starts editing.
Email with AI — public release demo, 2026.
Drafts are grounded in the project record, not written from a blank prompt.
Generic AI email is fluent and wrong. Pulling dates, quantities and status from the record makes the draft useful on the first try.
Built for the quick issue, not the long thread.
In interviews, the pain wasn't writing big updates — it was the dozens of small ones: a date slipped, a file's missing, a supplier needs a nudge. Designing for a thirty-second resolve, inside the project, is what made it worth using.
From answering to doing
The agentic mode takes a request like "create a project named Spring Campaign, due in two weeks" and carries it out: it validates the inputs, walks the user through anything missing, and lands them on the project overview with everything pre-filled and ready to confirm.
Agentic Noosh — public release demo, 2026.
The agent ends on the normal overview screen, not in the chat.
This is Principle 1 at its most literal. The user finishes exactly where they would have by hand, with the same controls — the AI just did the walking.
Missing or invalid inputs become a guided step, not an error.
Enterprise data is messy. Turning validation into a short guided flow keeps momentum instead of bouncing the user back to the prompt.
Where it started
The first release focused on a narrow set of high-frequency tasks — create and copy projects, find records — to prove the pattern before widening it.
AI Companion V1 — public release demo, early 2025.
Research & testing
What users taught me
Interviews and usability sessions ran alongside every release. Internal artifacts stay internal, but the shape of what I learned is worth sharing.
"I want to know what I need to do — in a succinct way."
The request that came up most often, paraphrased. It shaped the Companion more than any feature idea did.
People don't trust invisible work
Users wanted to see what the AI changed. Marking AI-filled fields turned skepticism into confidence.
Examples beat instructions
Nobody read guidance text. Concrete, copy-able example requests were what got people started.
Scope beats freedom
Open-ended prompts produced open-ended failures. Scoping intent with chips raised success on the first try.
Experts and novices fail differently
Power users skipped guidance and wanted fewer steps; occasional users wanted to be walked through. One surface that adapts the amount of help, rather than two modes, served both.
Outcome
Where it landed
Five AI features shipped to Noosh's global user base between early 2025 and 2026, with a shared design system underneath them and a workflow-first pattern that new features now follow by default.
- AI Companion, Smart Search, AI Spec Creation, Email with AI and Agentic Noosh live for 6,000+ users in 44 countries
- A design system that makes every AI surface consistent with the rest of the product
What I'd carry forward. Where the AI sits matters — but it's not the biggest lesson. The biggest is that the work is still about people: how they interact with it, whether they trust it, and whether it made their day easier. The features that landed best weren't the smartest ones; they were the ones that made an existing step lighter for the person doing it.