BYO AI
You choose the AI. It can draft. It can never decide.
Clyreon does not sell you a model or mark one up. You point it at whichever model your company already uses, on your own account, and it can only ever suggest things for a person to approve.
What bringing your own AI means
It comes down to three things.
You pick the model
Whatever your company has already approved and already pays for. Clyreon ships no model of its own and has no preference.
You pay for it directly
To your model provider, at your own rate. Clyreon adds no margin and takes no cut, so nobody here has a reason to want you using more of it.
It runs where you say
Against your own account, at whatever address you configure, including a model running inside your own network. Your security team decides which model sees your material, not us.
Why you might want this: most companies have already spent months deciding which AI vendor they trust with company information. Nobody wants to have that argument again because a supplier bundled a different one.
Clyreon works with no AI at all
Everything else on this page depends on it.
Opening a decision, writing down your reasoning, attaching evidence, saying what would prove you wrong, and recording what happened all work with no AI involved whatsoever. If you never connect a model, you lose nothing that the rest of this site describes.
A memory of your decisions that came from a model would be a memory of the model, not of your company.
What a model can never do here
Not a policy or a setting. There is no way to do these things, for anyone.
A model connects to Clyreon exactly the way Salesforce and Slack do, and that connection can do precisely one thing: add a fact. There is no way for it to open a decision, take a position, or close one, no matter what it produces.
Every claim a model returns has to point at the material it was given. Anything that points at nothing, or points at something it was never shown, is thrown away before it reaches your record. Not flagged, not saved with a warning. Thrown away.
What does survive says what it is. It is recorded as something a model worked out, listing what it worked from, with a model named as its author. Nothing a model produces can look like something a person saw.
A model can draft a position and the reasoning for it. It cannot set the confidence. How sure somebody is, is a fact about that person, and a number nobody actually formed is worse than none.
It cannot record that somebody disagreed, because that would be hearsay about another person’s view. It cannot record an outcome, because what happened is something you observe rather than predict. It cannot claim who said or did anything.
What you actually see when a model drafts
Two parts, and you can tell them apart at a glance.
The prose is the model’s and is not checked. The sources beside it are checked, every one, against the material actually given to the model before you see the draft. The text shown next to a source is always Clyreon’s own wording, never the model’s summary of it.
If the model cites something that does not exist, that citation is shown as unsupported rather than quietly removed. If nothing it cited holds up, you still get the draft, clearly marked as having nothing behind it, because at that point it is a model’s opinion and you should be told so.
A draft has nowhere to go except back to a person, who accepts, edits or bins it exactly as if they had written it themselves.
For the engineers reviewing this
The short technical version.
Clyreon bundles no vendor SDK. You supply a credential and an endpoint. The only thing the rest of the system can do with a model is send a prompt and receive text back.
One adapter ships, matching the OpenAI-compatible /chat/completions shape, which is what OpenAI, Azure OpenAI and most self-hosted or gateway-fronted models already speak. Moving from a hosted model to one inside your own network is a configuration change, not a code change.
The restrictions above are enforced by the code that reads the model’s output, not by stripping fields afterwards. A field that is never read cannot be set, whatever the model emits.
What Clyreon will not build
A permanent boundary, not a gap waiting to be filled.
Clyreon will not build automatic deciding. Not as an enterprise option, not behind a setting, not for a large enough customer. The reason is not doubt about how good models are. It is that responsibility for a decision has to sit with a person, or the record stops being a record of anyone’s judgment.