OVERVIEW
Bringing AI integrations to Acelab's Material Hub
As the industry shifted toward AI-native workflows, material selection demanded instant assistance and intelligent discovery.
This led us to design Ask Ace, Acelab’s in-platform AI assistant, alongside an automated product matching engine. I spearheaded the design of both the co-pilot companion and foundational platform AI flows, optimizing them for natural material discovery and workflow querying.
Collaborators
Purti Hardikar- Lead PD
Hardik Makrubiya- Lead PM
Nolan Tremling - CTO
Annika Harmsen - Full Stack
Amen Abe- Data Eng
Timeline
Q3 2024 – Q3 2025
Status
Live and shipped 🚀
PROBLEM & OPPORTUNITY
We had launched Acelab's Co-pilot Assistant
The first copilot started from a blank prompt. Before asking a useful question, people had to explain what they were working on.
but..it was not context aware!!
HOW DID WE SOLVE THIS?
AI was now a part of the architect's workflow
Bringing AI to Acelab meant empowering architects to increase productivity through smart, thoughtful interactions specifically by offloading redundant manual work.
To achieve this, we designed two high-impact AI features:
Ask Ace (a context-aware AI Assistant)
Ask Ace evolved into a true work assistant, gaining real-time awareness of your active projects, help center resources, saved products, and more.
Smart Product Matcher:
Acelab's AI product matcher drops project imports and products straight into an architect's daily workspace, cutting out the friction of manual specification.
PRODUCT KEY HIGHLIGHTS
Compare products & Send RFI for material Specs.
Ace is your material data best friend.
Real-time awareness of your projects.
AI builds smartly on your team and it's data so decisions are not made in silos.
IMPACT
Early signals after launch
Usage from Acelab’s PostHog dashboards, to show the impact these projects made.
Ask Ace
529 messages in 30 days, up 55% on the month before.
195 people asked Ace at least once, up 56% month over month.
Most conversations start in the side panel, right beside the work, which confirms the decision to dock Ace there.
AI matching
86% of picks are the top suggestion, and 98% fall in the top three, which backs the three-match cap.
13 of 14 people who pick a suggestion go on to apply the match.
Still open: only 32% of people who open the panel pick a suggestion. That drop-off is what I’d study next.
Reliability over the same period: 0.49% of Ace responses ended in an error and 1.67% used a fallback. Both are watched after every release.
MY ROLE
A sole designer & pm on the project.
As a lean team at Acelab, I took on dual roles as Product Manager and Product Designer for a high-priority project, delivering two core, interconnected experiences:
I designed two connected experiences: Ask Ace, an assistant beside the workspace,
& AI Matches, a way to connect imported schedule product rows to Acelab's catalog products.
Established foundational AI design tokens and component libraries for Ask Ace, ensuring consistent hierarchy across text responses, interactive cards, and data citations.
Architected modular response structures that dynamically adapt to query types such as material lookups and performance comparisons to reduce cognitive load for architects.
Created scalable interaction patterns for conversational UI, establishing guidelines for loading states, inline actions, and dismissible previews within the chat stream.
Process Alignment: Defining how the team works streamlines collaboration between design, product, and engineering.
Where should the assistant be placed in Acelab?
Originally launched as a general co-pilot for material data, Ask Ace operated in a vacuum, it was not able to help architect's make decisions based on active project parameters.
An isolated AI could never safely make autonomous decisions on their behalf.
Three Places to put an AI assistant:
DESIGN DIRECTION WHY WORKSPACE WON?
Ace belongs in the workspace
By docking Ask Ace beside the primary canvas rather than using a floating overlay, we eliminated the cognitive split between asking and looking.
2
4
The call
Give Ace a permanent place next to the page, so asking never means leaving what you’re looking at.
The trade-off
The page gives up width. Every page now resizes to make room, which Front Door covers in more detail.
While initial iterations featured an overlay chatbot and side panel, user feedback revealed that architects required a more deeply integrated workspace solution.
CONVERSATIONAL UI
Being beside the work wasn’t enough. Ace needed to understand it.
Moving Ace into the workspace solved the placement problem.
Next, we connected each task to the context it needs: the active project, selected products, schedules, help content, and representative information.
Context informs the answer. Changes still require the architect’s approval.
CONVERSATIONAL UI
More capabilities needed a consistent response language & a design system
As Ask Ace expanded, single text responses fell short. Each task required a custom structure while maintaining Acelab’s cohesive experience.
Build from Acelab's Design System. Extended for AI Experience
FINAL DESIGNS
One system, different kinds of help.
Below is a look at how Ask Ace maintains continuous context as users navigate across different pages of the platform.
A canvas view for complex data
Schedules was Acelab’s data-heavy material spec tool which demanded extra workspace to handle dense information.
To ease cognitive load, we introduced expand/minimize states alongside a dedicated canvas view to streamline complex workflows.
Closed
The page uses its full width. Ace is one click away in the header.
Docked & Canvas
Ace sits beside the page, and the page resizes instead of being covered.
Full screen
For longer conversations. History and sharing live in the panel header.
DESIGN SOLUTION
Could AI help connect the products we already had?
The imported files contained information we could use: product names, manufacturers, and other identifying details. Acelab already had the catalog to compare them against.
The opportunity was to use those details to suggest likely matches, then give architects a clear way to review them.
But how would the mapping parameters work?
Post import the users were given an optipn to syaty mapping thei rprofdusxWe mapped out the whole flow and detailed nuances on how we would wnat the user flow to look like
Make the suggestions easy to compare.
We mapped information from each imported row to catalog products and surfaced up to three candidates. I kept the original product visible beside its suggestions so architects could compare what they had imported with what AI had found.
A deliberate limit
Up to three matches per row
Fewer, stronger options are quicker to compare and easier to trust than a long ranked list.
Row 7 · Insulated Metal Wall Panel
Micro-Rib Insulated Panel
Exact
96%
Box-Rib Metal Wall Panel
Close
87%
Flat-Seam Wall Panel
Close
82%
Weaker candidates stay hidden
Compare from schedules anytime or skip for later
user could skip mapping prodycts and ma[ them at a latetr stage users could also decide which prodcucts to map and whcih they wated to skip, they could alos anuammy relace a product by serahcing through the product catalog.
Matching from the schedule
The AI icon appears only on custom products and opens matches right at the row, so people never leave the table.
Mapping also worked for existing projects
We also extended mappign to exisying prodject which had custom products this making the mapping feature a wholistic feature.
KEY TAKEAWAYS
A PERSONAL REFLECTION
“I wanted AI to take on the repetitive work while keeping architects in control of the decisions.”
Understand the data before designing the answer
Matching exposed gaps and inconsistencies in product data. I relied on our engineers to understand what we could reliably match, then designed clear paths for uncertainty and missing information.
Make uncertainty understandable
A score alone could not explain a recommendation. Plain-language labels and specific differences gave architects a clearer basis for deciding what to trust.
Automate the repetition. Keep the judgment.
AI could narrow the search and prepare matches, but the architect approved replacements. Knowing when to ask for confirmation mattered as much as making the task faster.
Scope AI around real value
I learned to weigh capability, effort, and access together. Roadmap and plan-limit decisions needed to reflect real pain points, so useful assistance stayed within reach when people needed it.
