Every question started a treasure hunt
Company knowledge was spread across drives, wikis, inboxes and people's heads. Employees searched three or more places, then gave up and asked the same few experts.
01Overview
Project overview: before, turning point, after
This is the story of how employees stopped hunting through drives, wikis and inboxes, and how the colleagues they kept interrupting got their days back. I designed the end-to-end experience: user flows, information architecture, wireframes, AI interaction patterns and the design system.
Trust comes from knowing the source
People would only rely on an AI answer if they knew whose knowledge it came from, and if owners could control what went in.
Ask once, get a trusted answer
Mind AI Ninja answers questions from each department's own documents, as text, tables or charts, while owners keep their knowledge current in one library.
My design process
02Context
Company knowledge is everywhere, and nowhere
Policies, contracts, project notes and reports live across shared drives, wikis, inboxes and people's heads. As a company grows, it knows more and more, yet finding the right answer gets slower.
Mind AI Ninja works like a good reception desk. You ask your question once, and it knows which department to ask, which document is current, and how to put the answer in plain words.
Where it breaks down
Scattered sources
Files, drives, wikis and web pages each hold part of the answer, with no single place to search them all.
Department silos
HR, finance, legal and engineering each own knowledge that other teams need all the time.
The same questions, again
Experts spend hours answering repeat questions instead of doing their own work.
AI without guardrails
Public chatbots can't see internal documents, and shouldn't. Companies need AI that only uses their own content.
Key terms
- Knowledge base
- All the company documents the assistant is allowed to draw its answers from.
- AI assistant
- A chat you can type or talk to, which answers questions in plain language.
- LLM
- Large language model: the AI “brain” that reads and writes text.
- Embedding model
- The part that indexes documents by meaning, so a search finds the right idea, not just matching words.
- Department scope
- Limiting answers to one team's documents, so an HR question is answered only from HR's files.
- Base prompt
- Standing instructions that tell the AI how to behave for a particular department.
03User Persona
Designing for the person everyone asks
Rahul leads HR operations at a 2,000-person company. He owns more than a hundred policy documents, and answers over thirty questions a week that are already written down somewhere. Here's a typical Monday.
One policy question, before Mind AI Ninja
≈ 3 hours lost
- 9:00 amA new joiner needs to know the leave policy.
“How many days of leave can I carry over to next year?”
- 9:20 amShe searches the drive, the wiki and her inbox.
She finds three documents, each with a different answer.
- 10:00 amShe messages Rahul.
He stops his own work to find the current version and reply.
- 11:30 amTwo more people ask the same thing.
One by email, one in the corridor.
- 2:00 pmA manager approves leave using last year's policy.
An old PDF is still doing the rounds, and it gets used.
- 5:00 pmRahul's own report is still untouched.
His day went on answering questions that were already written down.
An illustrative scenario, pieced together from what employees and knowledge owners described in research.
User personas
Mind AI Ninja serves three kinds of people: those who own knowledge, those who need it, and those who run the platform.
Rahul Verma
HR Operations Lead
IT services company · 2,000 employees
Half my day is answering questions that are already in a document somewhere.
- Questions answered
- 30+ / week
- Policy documents
- 120+
- Knowledge sources
- Drive, wiki, email
- Role in platform
- Dept admin
Goals
- Let employees self-serve answers
- Keep one current version of every policy
- Know which content is used
- Generate FAQs and summaries quickly
Frustrations
- The same questions arrive by chat, email and in person
- Old versions of policies keep circulating
- No visibility into what people search for
- Public AI tools can't be used with internal documents
Responsibilities
- Maintain HR policies and documents
- Answer employee questions
- Onboard new joiners
- Report on HR operations
Traits
Technical skillMedium
Time pressureHigh
Trust in AIMedium
Meera Iyer
Business Analyst · new joiner
Same company · 3 months in
I don't even know which team to ask — let alone which folder to look in.
- Tenure
- 3 months
- Searches
- 10+ / day
- Teams worked with
- 5
- Preferred format
- Short answers
Goals
- Find answers without interrupting colleagues
- Get numbers as tables or charts
- Know which source an answer came from
- Ramp up faster
Frustrations
- Doesn't know where information lives
- Search returns long documents, not answers
- Feels awkward asking the same people repeatedly
- Unsure whether a document is current
Responsibilities
- Prepare analysis for project teams
- Gather requirements from departments
- Summarise reports for managers
- Learn internal processes
Traits
Technical skillHigh
Time pressureMedium
Trust in AIHigh
David Chen
IT Platform Administrator
Same company · IT team
I need to know exactly which data each model can see — and prove it.
- Organisations
- 3
- Departments
- 18
- Users managed
- 2,000
- Integrations
- Drive, Confluence
Goals
- Set up departments and models quickly
- Control access by role and organisation
- Monitor uploads and failures
- Review AI conversations when needed
Frustrations
- Failed uploads go unnoticed
- Access rules differ across tools
- No audit trail of AI answers
- Every model change needs a deployment
Responsibilities
- Configure departments and LLMs
- Manage users and organisations
- Connect content sources
- Monitor usage and logs
Traits
Technical skillHigh
Need for controlHigh
Trust in AIMedium
Rahul is the primary persona; Meera and David are secondary.
What this meant for design
The product needed two faces: a simple assistant for employees, and a clear control panel for the people who own and run the knowledge.
04The Problem
Searching took longer than the work itself
Employees had to dig through several tools, out-of-date documents and busy colleagues just to answer simple questions, and the whole company slowed down because of it.
The answer usually existed. It was just buried in the wrong folder, in an old version, or in the head of someone who was already too busy.
Challenge → business impact
The problemWhat it cost the business
Documents scattered everywhereHours lost every week searching across tools.
Systems that don't talkTeams worked from different versions of the truth.
Out-of-date documentsDecisions made on old or wrong information.
Everyone asks the same expertsSpecialists are constantly interrupted, and become bottlenecks.
Summaries written by handRoutine reports and FAQs take far too long.
Problem statement
How might we give every employee one trusted place to ask anything about the company, answered from their own organisation's knowledge?
05Research & Discovery
How employees actually find answers
I wanted to know where people look first, which sources they trust, and how the owners of that knowledge keep it current. The goal was an assistant people would trust enough to ask first.
Understanding the domain
Where knowledge lives (files, drives, wikis, web pages)
Which department owns what
How the AI models are set up
How people chat and search
When people prefer to speak
Who is allowed to see what
Stakeholder interviews
EmployeesNeed answers
Department HeadsOwn knowledge
IT AdminsRun the platform
HR & ComplianceSet policy
LeadershipMake decisions
Key questions I asked
Finding information3 questions
- Where do you look first when you need an internal answer?
- How long does it usually take to find it?
- When do you give up and ask a colleague instead?
Owning content2 questions
- How do you keep your department's documents up to date?
- How do you know whether people are finding them?
Trusting AI answers2 questions
- What would make you trust an AI answer about company policy?
- What format do you want answers in: text, tables or charts?
Access & control2 questions
- Who should be able to see which department's knowledge?
- How would you check what the assistant has told people?
What we found
The same patterns came up on both sides, among those asking and those answering. Paraphrased from the interviews:
“I check three or four places before I give up and ask someone.”
“I can't tell an old document from the current one.”
“Honestly, the real search engine is Rahul.”
“If it's about numbers, show me a table or a chart, not a paragraph.”
“When I upload documents, I never know if they actually went in.”
“Nobody can check what's been asked or what was answered.”
Key insight
Trust in an AI assistant comes from its sources. People need to know which department's knowledge an answer came from, and owners need to control what goes in.
06Competitor Analysis
Search finds files. People need answers.
I compared how companies already handle internal knowledge. Each approach covered storing it or searching it, but few gave a trusted answer drawn from the right team.
| Approach | Good at | What's missing | Verdict |
|---|---|---|---|
| Shared drives & wikisStorage | Storing knowledge. | Answers. Search returns documents, and nothing flags the out-of-date ones. | Partial |
| Company search toolsSearch | Searching across many sources at once. | The answer itself. You still have to read and interpret the results. | Partial |
| Public AI chatbotsPublic AI | Understanding and writing language. | Access to your own documents, and any control over who sees what. | Not scoped |
| Asking colleaguesPeople | Accurate answers with context. | Scale. Every answer costs an expert's time. | Bottleneck |
| Mind AI NinjaThis project | Department-scoped AI chat, voice and a content library over the company's own documents. | No guesswork. Every answer comes from known sources that admins control. | Opportunity |
Opportunity
Combine AI's way with words with each department's own documents, and give the owners control over both.
07Design Decisions
Why it works the way it does
The core idea: knowledge belongs to departments, and every answer should be traceable to where it came from. Here's what we noticed, what we could have done, and what we chose.
Departments as the unit of knowledge
We noticed: people trusted answers when they knew which team they came from. So each department gets its own admin, its own AI settings and its own document sources. Answers stay in scope, and owners stay accountable.
Instead ofOne big pile of every document
We choseA knowledge space for each department
Chat with memory and scope
We noticed: real questions come in follow-ups. So the chat keeps a history of conversations and lets you pick which department you're asking, and follow-ups stay in context.
Instead ofA single search box
We choseDepartment-aware chat with history
Answers can be data, not just text
We noticed: for questions about numbers, paragraphs got in the way. So the assistant can answer with tables and charts, which can be copied or regenerated.
Instead ofText-only answers
We choseCharts and tables right in the chat
A content library with clear status
We noticed: admins uploaded documents and simply hoped they worked. So the library brings in files, web pages, Confluence and Google Drive, and shows each item's status (uploaded, in progress or failed) at a glance.
Instead ofUpload and hope
We choseOne library with a status for every file
Voice, with a record of every conversation
We noticed: nobody could check what had been asked or answered. So a voice assistant opens access beyond the keyboard, and every call is logged with a recording, a transcript and an analysis, letting teams review it later. Every screen comes in light and dark themes for long working sessions.
Instead ofVoice questions that leave no trace
We choseVoice logs with transcripts & analysis
The hardest part
Keeping the employee side as simple as a chat, while giving admins real control over models, sources and access, without one side's complexity leaking into the other.
08The Solution
Ask, get answers, keep knowledge current
Mind AI Ninja brings AI chat, voice, a content library and department settings together in one place: simple for the people asking, and in control for the people answering.
The same question, with Mind AI Ninja
≈ 2 minutes
- 9:00 amThe new joiner asks the assistant, with HR selected.
“How many days of leave can I carry over?”
- 9:00 amThe answer comes from HR's current policy.
It's drawn from HR's own documents, not an old PDF.
- 9:01 amShe asks a follow-up: “Show it as a table by year.”
The chat remembers the context and answers with a table.
- 10:00 amRahul uploads the new policy to the content library.
Its status shows “uploaded”, so he knows it's live.
- 10:05 amHe generates an FAQ from it in one click.
Then he gets back to his own work.
Key features
01 · AI chat
Ask anything, in context
Employees ask questions in plain language, choose the department they're asking, and get answers as text, tables or charts. Their conversation history is always one click away.
Fixes“I check three or four places before I give up and ask someone.”
02 · Content library
Bring knowledge in from anywhere
Admins add content from files, web pages, Confluence or Google Drive, and see each item's status, owner and size in one library.
Fixes“When I upload documents, I never know if they actually went in.”
03 · Departments
Knowledge spaces with owners
Each department has its own admin, AI settings, connected sources and standing instructions, so answers stay accurate and ownership stays clear.
Fixes“I can't tell an old document from the current one.”
04 · Voice & insights
Every conversation, reviewable
Voice calls are logged with recordings and transcripts. Documents get AI summaries, a read of their tone (sentiment), and auto-generated question sets.
Fixes“Nobody can check what's been asked or what was answered.”
Product screens
09Impact
Less searching, more doing
Mind AI Ninja was built to change how much time people spend looking for answers, and how often experts get interrupted. These are the outcomes it was designed around.
Knowledge anyone can reach
The company's knowledge is now reachable through one place you can simply ask.
Less time spent searching
Employees get answers, summaries and FAQs directly, instead of hunting through documents.
Room to grow
A solid foundation, ready for future AI features as the company grows.
Remember Rahul? “Half my day is answering questions that are already in a document somewhere.” Now the document answers for itself.
10Reflection
What this project taught me
Mind AI Ninja turned complicated company-knowledge work into something that feels like asking a helpful colleague. Pairing AI conversations with a clear dashboard for owners made information easier to reach, and left room for the product to grow.
Scope builds trust
Answers felt reliable once people could see which department's knowledge they came from.
Admins are users too
The assistant is only as good as its content, so the owners' side deserved as much care as the chat.
Match the answer to the question
Numbers deserve charts. Letting the format follow the question made AI answers far more useful.