Interviews

Brad Wardell, Founder and CEO of Clairvoyance AI – Interview Series

Antoine Tardif
Add Unite.AI to your preferred sources on Google
Brad Wardell, Founder and CEO of Clairvoyance AI
Brad Wardell, Founder and CEO of Clairvoyance AI.
Key PointsWardell argues routing work between local and cloud hardware matters more than raw…
View more
  • Wardell argues routing work between local and cloud hardware matters more than raw model size.
  • Clairvoyance stores agent memories in a graph so updates propagate and stay current.
  • Wardell predicts 90 to 95% of AI work will run locally within 18 months.

Brad Wardell is Founder and CEO of Clairvoyance AI and the founder and CEO of Stardock. He has spent more than three decades building desktop software and strategy games, including products such as Fences and Galactic Civilizations. His work on game development pipelines, creative assets, and everyday business reporting shaped his interest in persistent AI agents that can work directly with local files and applications.

Clairvoyance AI develops a workspace for named AI staff with ongoing assignments and memory, connecting local files and development tools with cloud services and models running on a user’s own hardware. Incubated at Stardock, it supports workflows that produce reusable notes, reports, and other artifacts. In this interview, Wardell discusses how game development informed the product, where local models fit into daily work, and the routing, memory, and permission controls needed to make agents useful.

You’ve spent more than three decades building desktop software and strategy games at Stardock, including products such as Fences and Galactic Civilizations. What problem inside your own business prompted you to build Clairvoyance AI, and when did you realize it could become a company of its own?

The real pressure point was Ara: History Untold, which Stardock helped Microsoft and Oxide Games finish. A huge amount of time was going into just moving creative assets into the game. Writing came in as .DOCX files that had to be painstakingly hand-converted into XML or CSV. 3D assets had to be carefully brought into our engine. A single balance change meant touching a dozen different files. Localization meant matching translated strings back to the right keys in the right files by hand. Clairvoyance AI made all of that painless and let us ship on time.

By late 2025, we had an internal debate about whether to make it public at all. It was a real competitive advantage, like having a superset of something like OpenClaw a year early. What pushed us toward going public was watching friends and colleagues lose jobs to AI automation that they could have kept if they’d had better tools. At that point, we knew we wanted to open Clairvoyance AI to the public.

You’ve argued that many AI agent tools reveal a lack of experience shipping consumer software. What mistakes do you see most often, and what lessons from your own career should AI developers be applying?

The biggest tell is terminology. If you don’t already know the words, like “glyph” for the little graphic on a button, you’re at a real disadvantage talking to an AI, because it’s trained on people who already know the lingo. Most of the teams building these tools are very smart but very young, and it shows in the usability choices. A good example is when Anthropic added the ability to drag text from one terminal session to another, that got millions of views. We’ve had agents that you can just @ by name to hand off work automatically, no dragging anything. Theirs requires the user to do the work the AI should be doing.

The lesson from 30 years of shipping desktop software is that you have to actually use the thing you’re building, every day, with real deadlines, before anyone outside the company sees it. A lot of AI products feel like they were built by people who haven’t had to live inside their own software.

You describe the next phase of AI as a routing problem: deciding what runs on a device, across a local network, or in the cloud. Could you walk us through a real workflow and explain how those decisions should be made?

Routing isn’t really an individual decision anymore, it’s becoming a company-level one. We’re about to release a feature called Hivemind that makes this concrete. Most enterprises already own plenty of capable machines that sit idle every night. Hivemind lets a company pool those idle GPUs across its own machines so anyone on the team can run AI work on them instead of sending it to the cloud. If you’re tracking tens of thousands of shipments across a supply chain, or auditing where a company’s money went, you don’t need a frontier cloud model for that. You need enough compute, and most companies already have it sitting unused after hours.

At the individual level, the same logic applies, just on a smaller scale. I run dashboards, crash reports, sentiment reports, and sales data every night on a local model instead of a cloud one. I used to run that same workload through Claude and it cost about $100 a month in tokens. Now it runs for free overnight on hardware I already own, and I read the results in the morning.

So the actual decision isn’t “local versus cloud,” it’s “does this need to happen while someone is staring at the screen, and does the company already have the hardware to do it without paying per-token.”

You believe a substantial share of everyday agentic work can already run on local hardware. Which tasks have you demonstrated this with, on what hardware, and how do you measure success beyond whether a model produces a plausible answer?

We have had a lot of success with using local AI for adding metadata to our files. Like many companies, we have terabytes of data on our network. Stardock develops both software and games and on the game side, there’s decades of 3D and 2D art assets, videos, work products, etc.

Around the company we have a lot of pretty decent video cards from AMD and NVIDIA. We have found that 8B to 12B parameter models are more than enough to do everything we need and those will run on even relatively older hardware. Right now, we have a Qualcomm Snapdragon X2 laptop doing some of this work in the background during the evening.

Local AI is really good at just slowly chewing through and reviewing these files and adding metadata to it so that when we need to find something, you can search for it and find it quickly. Hence, we discovered that we have 16 different 3D models of wolves. Sigh.

What should trigger escalation from a smaller local model to a more capable cloud model? How do you detect when the local model is struggling before it makes a consequential mistake?

Right now, anything that requires open-ended judgement should probably go up to the cloud. The frontier models have largely solved hallucinations. When you go to smaller (local) models, hallucinations return. So on local models we tend to need to keep a lot of guardrails on them.

Clairvoyance organizes agents as AI staff with persistent memory and ongoing assignments. What have you learned about preserving useful context over time without allowing outdated instructions, incorrect assumptions, or conflicting memories to accumulate?

Under the covers, Clairvoyance stores its memories in a graph structure so when one thing updates, it tends to recursively trigger updates to other nodes. This helps ensure that the memories stay organized and up to date.

The challenge has been to prevent infinite recursion and that has required a lot of trial and error over the last couple of years.

Giving agents access to files, applications, and development tools makes them more useful, but also increases the consequences of errors. How do you structure permissions, human approvals, and recovery when an agent takes the wrong action?

The one concrete point we have is File Protection: Clairvoyance can block agents from wiping out directories, enforced at the app level rather than through hooks that an agent could work around.

When a workflow combines local execution with cloud inference, “local-first” can mean different things. What information leaves the user’s machine, and how can users understand and control that boundary?

Cloud security is about keeping other people out. It doesn’t answer whether the model provider should have your company’s secrets in the first place. Clairvoyance is local-first by design: your notes, projects, and staff memory live on your own machine. What goes to a cloud model is your call, made per staff member, not a blanket setting. Run a local model and nothing has to leave your hardware at all. We’re hearing this directly from enterprise customers who are frustrated about uploading entire confidential workflows to Anthropic or OpenAI just to get work done.

Clairvoyance has over 40,000 users. What have you learned from how people actually use Clairvoyance, and can you share a measurable example of a recurring workflow where it has delivered sustained value?

The retention rate is highly dependent on the amount of data the user is generating through their sessions.

Most users of AI haven’t really crossed the threshold where they are taking the data the AI generates to build back onto it. Most users still think in terms of prompts. The users who stay with Clairvoyance are the ones generating a lot of Exhibits and Reports which in turn act as the foundation for further work.

Looking ahead a year, what developments would most expand the work agents can reliably handle locally? What would need to happen for you to reconsider your view that routing matters more than model size?

I think within 18 months, 90 to 95% of AI work will be local, with maybe 5% needing the cloud for frontier work. That’s mostly because almost nothing is being delegated to AI yet, even trivial tasks like screening resumes for location or checking shipment timing are still done by hand, and a 4B model can already do that on a laptop. What expands this is Hivemind-style pooled compute (letting a company’s underused GPUs serve everyone on the domain) and routing getting good enough that nobody has to know what a 4B model or an Opus even is.

Thank you for the great interview, readers who wish to try a desktop app for managing persistent AI agents across their computers should visit Clairvoyance AI.

Antoine is a visionary leader and founding partner of Unite.AI, driven by an unwavering passion for shaping and promoting the future of AI and robotics. A serial entrepreneur, he believes that AI will be as disruptive to society as electricity, and is often caught raving about the potential of disruptive technologies and AGI.

As a futurist, he is dedicated to exploring how these innovations will shape our world. In addition, he is the founder of Securities.io, a platform focused on investing in cutting-edge technologies that are redefining the future and reshaping entire sectors.