b2b SAAs AI
Building a human-in-the-loop layer for Atom — an AI agent that answers employee queries across IT and HR.
ROLE
Product Designer
TEAM
OVERVIEW
What is Atomicwork ?
Atomicwork is an B2B SaaS startup building an AI service agent to automate internal employee support. When I joined, it was -
・Early-stage, funded startup
・Pilot customers (1–2 companies)
・Product lived inside Slack/MS Teams
・Atom learned from company documents
The promise was simple: Employees ask → AI answers. The goal was to leverage AI and enhance employee experience.
Problem
Why can’t AI answers be trusted directly?
Instead of directly jumping into designing a verification pathway, I decided to dig a little deeper and research on some hard questions like:
Why can’t AI answers be trusted directly?
→Because they are probabilistic.
Where does trust actually come from?
→From control and accountability.
What was missing in the system?
→A checkpoint between generation and usage.
That shifted the problem: From “how to make AI answers better ?” → To “How to make AI governable ?”
A brief intro to ATOM - the AI companion
Atom is Atomicwork's AI agent, built to handle work that used to need a human in the loop. It lives where teams already are: Slack, Microsoft Teams, or the Atomicwork platform itself.
Instead of routing every request through a person, Atom takes over directly. Onboarding a new hire, answering policy questions, pulling up data, guiding someone through a yearly process, it handles the back-and-forth so employees get answers fast, in one place.
That's the promise. But for Atom to deliver it, its answers had to be trusted.
challenge
Governance couldn't come at the cost of friction
The hardest part was not accuracy. It was friction. If verification felt heavy, no one would use it. If it felt optional, trust would break. We had to make:
Verification quick
Editing effortless
Approval intentional
SOLUTION
Upload documents
You can give a large multi line description
Drag and drop, or choose files to upload
PDF, DOC, DOCX, TXT documents
What software can I access.pdf
Audience *
All USA employees
Female employees
Grant content access via Atom only to these user segments
Learn more about “Feature”
Cancel
Upload
1. From Documents to Structured Knowledge
Admins upload documents into Topics and then Atom
・Reads documents
・Generates possible questions
・Suggests answers
Instead of blindly learning, the system surfaces what it thinks first.
Add an answer
Document
URL
Notion Page
Confluence
2. Verification as a Core Interaction
This became the most important layer because admins could now:
・Preview answers
・Edit them
・Mark them as Verified
Verified answers became the source of truth.
This separated: AI suggestions from Enterprise-approved knowledge. And built more trust with sensitive data
Add an answer
Document
URL
Notion Page
Confluence
3. Learning from Real Conversations
We extended this beyond documents and even progressed towards Slack/MS Teams:
・Good responses could be saved
・Converted into verified answers
This meant: Real usage → better knowledge
Add an answer
Document
URL
Notion Page
Confluence
Delete verified answer?
Once deleted, Atom unlearns the contents of the verified answer.
Cancel
Delete
4. Unlearning as Control
Deleting a verified answer:
・Removed it from the system
・Prevented future usage
This gave teams confidence to iterate without risk.
5. Direct Verified Answer Entry
Lastly, an edge case we almost missed.
Not every answer came from a parsed doc. So I decided to gave teams their own folder where they can click "Add Answer”, drop in their Q&A. No upload. No parsing. Just verified knowledge, straight in.
So to summarise the decision make process
Initial State
Let AI answer everything directly
Fast, but unreliable
Alternative option
Force verification for every response
Accurate, but too slow
Hybrid Option
AI suggests & Humans validate selectively
This balanced speed and control.
impact
Even at an early stage, this changed how the product behaved.
・Introduced a clear governance layer
・Made AI responses deterministic where needed
・Reduced risk of inconsistent answers
・Positioned the product as enterprise-ready
It turned Atom from: “an AI that responds” into: “a system that can be trusted”
reflections
What I learned and observed
This project changed how I think about AI products. Accuracy is not the only problem but control is.
It’s about defining how the system behaves under uncertainty.
