Projects / AI Business Communication
Relay Hub: An AI-Powered Business Communication Platform
An AI-first business communication platform unifying chat, documents, and knowledge, scoped by what it's allowed to see.
- Company
- Relay Hub
- Timeline
- 2024 - 3 weeks
- Team
- Developers, Stakeholders, Tester, Product Designer
WHERE THIS IS TODAY
An AI layer built into every conversation, grounded in the company's own documents, storage, and past discussions, not a chatbot bolted onto a workspace.
AI-native
every module reads from and writes to the same AI-grounded knowledge layer
Full
system spanning communication, knowledge, and governance
3
AI reference modes, internal, external, hybrid
A company's knowledge, spread across every tool it uses
A typical team runs its day across separate tools for talk, for documentation, and for anything that needs an AI to think it through. Each tool does its one job well. None of them know the others exist, and none of them know the company. The cost of that isn't visible in any single tool. It shows up in the gaps between them. An employee asks a question in chat that was already answered in a doc they never saw. An AI conversation produces a genuinely useful answer, then evaporates the moment the tab closes, because nothing carries it into the next person's context. A new hire spends their first weeks discovering where things live rather than doing the job they were hired for. And every time leadership considered letting AI actually touch company documents, the same question stalled the conversation: who is it allowed to see, and on whose behalf?
“Every AI conversation I have is genuinely useful for about ten minutes, and then it's gone. Nobody else on my team ever sees it, and I'll probably ask the same question again next month.”
Disconnected by design, one root cause
Competitive research into chat tools, cloud storage, and standalone AI assistants confirmed the shape of the gap: every category solved one slice of this well, and none of them treated knowledge, conversation, security, and AI as parts of the same system.
Chat tools hold conversations, not knowledge
Chat tools are excellent at the moment of talking. None of them treat a conversation, or a meeting, as something the organization should be able to draw on later.
Documentation is scattered across drives no one tool can see
Cloud storage and docs tools each store real, current company documents, but nothing connects a document to the conversation that needed it, or surfaces it before someone has to go looking.
Knowledge stays trapped inside one-off conversations
A genuinely good answer, worked out in a thread or a meeting, rarely turns into something the next person can find. It just becomes one more thing to ask about again.
AI assistants answer generically, not organizationally
A standalone AI tool is fluent about the internet and knows nothing about this company, its policies, its customers, its own past answers.
Security was never designed for AI to have an opinion
Existing permission models govern who can open a file. None of them were built to answer the harder question of what an AI should be allowed to say about that file, and to whom.
THE GUIDING QUESTION
What if a company's AI wasn't a separate tool answering from nothing, but a layer that already knew what the company knew, and what it was allowed to say?
Following a question, not an org chart
The gap between tools is easy to describe and hard to feel. So rather than audit the stack, I followed single questions through the company and watched what each role had to do to answer them. The same question cost a different amount of work depending on who was asking, and that difference is where the product lives.
WHO THE WORK RUNS THROUGH
3 personas
Elena Marsh
Office HR · policy → onboarding
- Age
- 41
- Location
- Austin, TX
- Tools
- Chat, docs, drives
- Tech
- High, non-technical
Owns the policy documents and answers questions about them one thread at a time. The answers are all written down and have been for years; people ask her anyway, because asking her is faster than finding the current version across three drives.
MOTIVATIONS
- Wants her time back from questions she has already answered
- Being the person new hires trust in their first week
- Policy applied consistently rather than remembered differently
GOALS
- Answer a policy question once and have it stay answered
- Point people to the current version, not a stale copy
- Get new hires productive without a week of hand-holding
FRUSTRATIONS
- The same five questions arrive every week in five threads
- Nobody can tell which copy of a policy is current
- A generic AI answers confidently and gets the policy wrong
“The answer is written down. It has been written down for two years. People still ask me, because finding it is harder than asking me.”
STAKEHOLDER SESSION
3 roles · 9 questions
I sat with each role and walked a real question through their actual day rather than a documented process, anchoring every session on how the answer was found, kept, and shared.
Office HR
Policy, onboarding
- Walk me through the last policy question someone asked you.
- Where does the current version of that answer actually live?
- What happens to your answer after you send it?
Research Team
Documents, findings
- Show me how you'd answer a question spanning several documents.
- What do you do with a good AI answer once you have it?
- How would you know if an answer came from our documents or not?
Managers
Oversight, access
- How do you catch up on what your team worked out this week?
- What would have to be true before AI could read company files?
- Who decides what the AI is allowed to say, and to whom?
The answer existed; finding it cost more than asking
Almost every question traced back to a document that already held the answer. The failure was never authorship, it was retrieval, so the fix had to be a way in rather than more content.
Good answers died at the end of a session
The most valuable reasoning in the company happened inside private AI chats and one-off threads, and none of it survived into a form the next person could reach.
Permission was the gate on everything else
Every role wanted the same capability and every manager stalled at the same question. Until AI access could be scoped per person, no amount of usefulness would get it approved.
THE GUIDING QUESTION
What would it take for a question to be answered once, kept where the next person will find it, and shown only to whoever is allowed to see it?
Six decisions, one idea: an agent for every task
The product could have been a chat app with a search bar bolted on, or a wiki with a chatbot bolted on. Each decision below was really the same one, made again in a new part of the product: treat knowledge, conversation, and AI as one system, never three.
Chat, research, and every workspace read from and write to one shared knowledge layer.
Every module risked becoming its own silo, the same failure as the tools it replaced.
Every response can retrieve and cite the company's own knowledge, not just what a model already knew.
A generic AI answer is only marginally better than no answer at all.
Every recurring task type gets its own configured agent, scoped to that job and its documents.
A single fixed assistant either tried to do every job at once, or fell short on tasks it wasn't built for.
Connect to those sources directly and retrieve live, instead of copying files into a new system of record.
Cloud storage already held the real, current documents.
Any conversation can run two models against the same prompt, side by side, before committing to an answer.
Settling for whichever model answered first risked settling for the weaker generation.
One continuous path: a question to a grounded, permissioned answer
An employee opens the agent built for the job, HR, onboarding, research, rather than a generic chat. The question carries their role and access scope before any retrieval happens, so the agent pulls only from what that person is allowed to see: the Knowledge Base, past answers, and connected drives like Google Drive or Dropbox. For higher-stakes questions, they can run the same prompt against two models side by side and pick the stronger answer.
- 1
Employee opens a task agent
EMPLOYEEInstead of a generic chat, they pick the agent built for the job, HR, onboarding, research, or continue an existing thread.
AI Workspace - 2
Question is scoped to a permission set
SYSTEMThe request carries the employee's role and access scope before any retrieval happens, not as a filter applied after.
Role-Based Permissions - 3
Agent retrieves from the Knowledge Base and connected drives
TASK AGENTThe agent pulls only from documents that person is allowed to see, company policy, past answers, files in OneDrive, Google Drive, or Dropbox.
Knowledge BaseCloud Integrations - 4
Two models can be compared before answering
EMPLOYEEFor higher-stakes questions, the employee can run the same prompt against two models side by side and pick the stronger generation.
Model Comparison - 5
Answer returns with its source
TASK AGENTThe agent responds in plain language and cites what it drew on, so the answer reads as grounded rather than guessed.
AI Workspace - 6
Conversation is saved to the right Hub
SYSTEMThe exchange is kept with the project or department it belongs to, so the next person asking the same question finds it instead of starting over.
Hubs
One shared knowledge layer, an agent built for every task
The information architecture splits into where people work day-to-day, where knowledge actually lives, and how collaboration stays governed, every layer feeding the same AI context rather than three separate products sharing a login screen.
AI Workspace
A chat experience deliberately familiar from modern LLM interfaces, extended into a business environment. A teammate can talk to AI or to a colleague in the same thread, continue a conversation across projects, share it with the team, reference company knowledge, or hand off to a task agent set up for that kind of question, built to feel like a business teammate rather than a standalone chatbot.
Knowledge Base
The structured home for SOPs, HR policy, onboarding guides, technical documentation, research, customer information, and internal process, organized so AI can reason over it directly rather than just index it as a pile of files.
AI-Powered Knowledge Retrieval
Chat is the primary way employees reach company knowledge. Instead of navigating folders, they ask a plain-language question and AI retrieves the relevant information, summarizes it, and answers in context, turning document management into conversational discovery.
Cloud Integrations
Direct connections to Google Drive and Dropbox. Rather than duplicating documents into a new system of record, AI retrieves straight from wherever the company already keeps them, respecting each source's own access permissions.
Task Agents
Instead of one generic assistant, a company configures agents around its own recurring work, an HR agent for policy questions, an onboarding agent, a research agent, each one instructed for its task and scoped to the same permissions and documents as the person asking, so business communication runs through the agent built for that job rather than a single catch-all bot.
Hubs
Dedicated spaces set up per project, a research hub, a client hub, any recurring body of work, each holding its own chats, documents, and agents so context stays contained to what that project actually needs.
Model Comparison
Two models can run against the same prompt side by side, so a team can weigh outputs directly and choose the stronger generation before committing to an answer.
What agents built for the job are worth, measured where it counts.
This is a working product, not a finished launch. It's in the product stage. The core flow is built and in use, and we're actively refining it, so the honest version of "impact" here is a status report: what's shipped, what's being sharpened right now, and what's still ahead.
~60% fewer tool switches
2x faster policy answers
~3 weeks faster ramp
Conversations retained
Knowledge has to be the center, or the product just becomes another silo
Early explorations treated Knowledge Base as one module among several. Once it became the shared substrate every other module read from, the product stopped feeling like a bundle of features.
Trustworthy AI is a UX problem before it is a model problem
The single biggest lever on whether people relied on an AI answer was not accuracy alone. It was whether they could see what it was grounded in and cite it back.
Permission has to live inside retrieval, not just at login
Treating access control as a feature added after the AI works undersells how central it is. Permission scope became an input to every retrieval call, not a filter bolted onto the response afterward.
