Announcements
Core views on AI safety
June 23, 2025
A message from the Founder: Ara Lee
Updated: May 1, 2026
After years of working in tech, I came out to build Olis with one decision: to build an AI that is safe, ethical, and real. Not another tool riding the wave of excitement and hype. Not a product dressed up for investors and stripped down for users. A tool "I" actually wanted. A tool I actually needed. A tool for my colleagues; the people I worked alongside for years deserved to have.
And it was about doing it right.
Because here's the truth. We are living in a time where the public has lost so much power and ownership over their own data. Every app we download, every tool we hand our trust to, most of them take that trust and quietly trade it for growth metrics and ad revenue. We gave them access. They gave us fine print, that writes something different tomorrow.
I wanted to prove that doesn't have to be the way. That a company can actually build something that works for people, not for the 1% at the top who never cared about the tool itself, only the value.
Olis is my proof of that. Every decision we make it all comes back to "does this actually serve the person using it? Is their data safe? Is their privacy real? Can they trust what we're telling them?"
If we ever stop asking that. If any part of how we operate drifts from that intention. Then we've failed from where we started. And I want to state that it won't happen.
You can hold me to that personally. When you see that happen, reach me directly at ara@olis-ai.com.
Why we're writing this
As Olis grows and more enterprises start relying on us as their ambient intelligence layer, we feel it's time to be direct about where we stand on AI safety. Not in a corporate, liability-driven way. But honestly, plainly, and in plain language anyone can understand.
We believe AI safety in enterprise isn't an abstract philosophical debate. It's immediate. It's operational. It's the difference between an employee getting the right answer in five seconds and making a confident decision, or getting a hallucinated answer and making a costly mistake that takes weeks to unravel.
That's what's at stake in every single company that deploys AI.
The Real Risk Nobody Talks About Enough
Employees are making decisions based on AI-generated answers that have no source, no accountability, and no way to verify.
Think about that. An employee asks an AI tool what the renewal process is for a vendor contract. The AI gives a confident, polished answer. The employee acts on it. But that answer was synthesized from generalized training data — not from the actual, current, internal process document that was updated six months ago. Nobody catches it. The contract lapses. The company loses a critical vendor relationship.
That's not a hypothetical. That's what happens when AI tools prioritize appearing helpful over being accurate. When there's no trail, no source, no accountability.
We built Olis to fix exactly that.
Our core beliefs on AI safety
Every single answer Olis surfaces comes with sources.
Not as an afterthought. Not buried at the bottom. Right there, front and center, alongside the answer. At minimum three sources per response, directly pulled from your connected enterprise applications — the actual documents, the actual systems, the actual data your company owns and maintains.
This is non-negotiable for us. If our system cannot point to where an answer came from, we do not surface that answer. Period.
Why? Because in enterprise, everything lives and dies by reference. You don't get to say something happened without proof. You put it in writing. You document it. You reference it. AI should work the same way. If you can't show me where you got that answer, I have no reason to trust it. And I shouldn't have to.
The industry has normalized hallucination. We haven't. We treat it as an unacceptable outcome, not an expected quirk.
Our RAG architecture — retrieval-augmented generation — means Olis doesn't invent answers. It retrieves real data from your real systems: your Jira tickets, your Notion pages, your SharePoint, your Slack history, your ERP. It then synthesizes across those sources to give you a direct, actionable answer grounded in what your company actually knows.
The difference between retrieving and generating is the difference between trust and risk. We retrieve. We surface. We cite.
Employees deserve AI they can actually control.
We trained our intent detection model to understand when someone is likely looking for information based on how people naturally write and work. When Olis detects intent, it highlights the text. Subtly. Non-intrusively.
You hover. You get the answer. You don't hover? Nothing happens. You keep moving.
We made a deliberate choice here: Olis does not force answers on anyone. It does not interrupt your workflow. It does not shove information in your face just because it thinks it knows what you need. It offers. You decide.
This isn't just good UX. It's an ethical commitment. AI should work for people, not the other way around. The moment employees feel like AI is surveilling them, anticipating their every move, or demanding attention — we've failed. Consent and control are not optional features. They are foundational to how we operate.
Access to knowledge should be equal, not hierarchical.
The way enterprise AI is typically built, the executives get the smart tools. The C-suite gets the dashboards. And the person in the mailroom, the new hire on day three, the admin trying to figure out who approves their expense report, they get nothing.
We built Olis for everyone.
Not for the people who already have access to everything. For the people who have never had a personal assistant, never had an executive briefing, never had someone to call who knows the answer. Because a company is not made by the leaders in the board room, but by the everyday employees on ground as well. Ollis gives every single employee in an organization — regardless of title, tenure, or department — the same quality of instant, accurate, sourced information.
That's what democratized knowledge means to us.
Data silos are not just a productivity problem. They're a safety problem.
When employees can't find the information they're entitled to, they make decisions with incomplete data. When processes aren't accessible, they get skipped. When agreements can't be located, they lapse. When nobody knows what anyone else knows, the entire organization operates in the dark.
Right now, the average enterprise uses over 100 applications. None of them talk to each other. None of them share a common knowledge layer. Employees are spending over nine hours a week just searching for information — and even then, most of what they find is outdated, incomplete, or wrong.
That's not acceptable. And it's not safe. Decisions made without the right information cost companies millions. Compliance failures happen because the right process document wasn't found in time. Projects fail because institutional knowledge lives in someone's inbox and not anywhere accessible.
Olis builds one memory layer across all of those applications. One place where all of your company's knowledge becomes findable, accessible, and actionable for everyone who's entitled to it.
How we approach safety in our own development
We train on ethical, consent-based data.
We are a young company and we are honest about that. We don't have all the answers. But here's what we're committed to:
Our intent detection model was trained on enterprise query patterns to understand how people naturally seek information at work. We do not train on proprietary company data without explicit consent. We do not sell user data. We do not use your company's information to train models that benefit anyone other than you.
We prioritize accuracy over speed. We would rather surface nothing than surface something wrong. Our source-citation requirement is a hard constraint, not a guideline. If we can't back it up, we don't show it.
We build for transparency. Every answer has a trail. Every source is clickable. Every piece of information Olis surfaces can be traced back to a real document, a real system, a real source inside your organization. You should always be able to verify that what Olis tells you is true and current.
We respect data permissions. If an employee doesn't have access to a document in Jira, Olis won't surface that document to them. We honor the permissions structures your company has already built. We don't bypass walls. We work within them.
The future we're building toward
Next category of AI: Proactive.
When all of your company's knowledge is connected, accessible, and trusted, the next step is anticipation. Olis will start to understand patterns. What information does your team typically need before a quarterly review? What processes usually get missed during onboarding? What contract renewals tend to slip through the cracks?
We will surface that information before you know you need it. With consent. With sources. With transparency.
But we will only go there once we've earned the trust to do it. We're not rushing toward proactive AI for the sake of a product roadmap. We're building toward it responsibly, with every step grounded in accuracy, source integrity, and user control.
When every piece of information becomes accessible, every choice becomes right. Every action becomes fast. Everyone becomes smart. And no one has to search for anything again.
That's the future we're building. And we're building it safely.
