Architecture over assurance
A guarantee you can verify beats a clause you have to trust. If the software has nowhere to call, no policy is required to stop it calling.
About Barie · Our journey
We started with a question nobody would answer straight.
Barie was not started to build a better chatbot. It was started because a room full of competent people could not get a straight answer to one question about where their data went, and because the honest answer turned out to require different architecture rather than a different contract.
The thesis
Sovereignty is not a setting. It is where the software runs.
Most of the market sells the same architecture with different words in front of it: your data leaves, and a contract promises what happens to it afterwards. We took the view that a promise is not a control, and that the only version of this worth building is one where the question does not arise.
Built with the awkward customers
The first deployments were the ones with the hardest constraints, on purpose. Everything that came out of those rooms is in the product rather than in a services engagement.
Say the number or say nothing
Where we quote a figure, it comes from a run somebody can point at. Where we do not have one, we leave the sentence without it rather than reaching for a range.
Own the compute, own the curve
Rented intelligence gets more expensive as you rely on it more. Owned compute does the opposite, and that difference compounds over the life of a deployment.
The journey
Four years, in the order it actually happened.
2022
A question nobody could answer straight
Three of us were doing diligence on AI vendors for a bank. The same question kept coming back with the same shape of answer: where does the data go at inference? Every deck said ‘private’. Every architecture said otherwise, and usually said it on page forty of the security appendix. That gap is the reason Barie exists.
2023
Built the boundary first
The first eighteen months went on the part nobody demos: a perimeter that inference runs inside, connectors that read in place under the permissions already on the source, and an audit trail written before the feature that needed it. Nothing shipped that made a call out of the network, because once one thing does, the guarantee is gone and cannot be recovered by policy.
2024
The Company Brain
Retrieval was never the hard part. Resolution was: deciding that the customer in the ERP and the account in the CRM are the same entity, at the moment the question is asked rather than at 02:00 the night before. Getting that to run at query time is what turned a search tool into something you could reason across a whole business with.
2025
From answers to work
Customers stopped asking for better answers and started asking for the work to be done. That meant long-horizon runs that survive a restart, writes that are previewed and reversible, and a receipt for every step. It also meant admitting that most agent systems fail on state rather than on reasoning, and building for that instead of around it.
Early 2026
First regulated deployments at scale
Government, finance and healthcare deployments running on customer-owned compute, inside customer perimeters, several of them on networks with no outbound route at all. Auditor mode came out of those rooms, not out of a roadmap workshop. So did most of what shipped in the two releases after it.
Now
Making it ordinary
The work in front of us is not a new capability. It is making a sovereign deployment as unremarkable to run as the systems it sits next to: the desktop app, the release cadence, the skills your own people write. Sovereignty stops being a project when nobody has to think about it.
How we work
Three rules we have not broken yet.
Each of these has cost us a feature at least once. They are written down here so that the next time one of them is inconvenient, it is inconvenient in public.
No outbound path, ever
Not for telemetry, not for crash reports, not for a model call. A feature that needs one does not ship, and several have not.
The customer can see everything
Every retrieval, every decision and every write is in a trail you own and can export. We do not hold a copy and we do not need one.
Your data is never training data
Nothing your deployment sees improves a model anyone else uses. The commercial model does not depend on it, which is the only version of that promise worth having.
How the boundary is actually enforced is set out on the Platform page.
Where we are now
Still a small company with a large perimeter to defend.
- Deployments running in government, finance, deep tech, legal and healthcare, several of them on networks with no route to the internet at all.
- A calendar release cadence with the full change list published every time. See product updates.
- Engineering and support in the same room as the customers who run it, because the alternative produces the kind of product this one was built in reaction to.
- Writing about the parts that are hard rather than the parts that demo well, at the blog.
Own your intelligence.
If the question you cannot get answered straight is the one we started with, bring it to us.