Providers is where the intelligence comes from. It is one page that behaves in two completely different ways, and which one you get is decided by how your NearSync is deployed rather than by a setting you choose here.
Managed
Most organisations are here, and it is the shorter half of this article on purpose.
The intelligence is provided as part of the platform. Each provider shows as active, with the models available and how much of your monthly allowance you have used.
There are no key fields, because there are no keys for you to hold. There is no vendor account to open, no card to put on file with a model company, and no second bill to reconcile.
There are no dollar figures, because you are not paying per request. Usage is measured in credits, which is the unit your plan is actually written in.
That is genuinely the whole page. If it looks like there is nothing to do, there is nothing to do.
Your own keys
The other path, and the one that needs work.
Each provider card opens into a key field. You paste the key from your account with that vendor, and from then on requests go through your account and your bill.
Test the connection before relying on it. The button next to the key checks it, and a bad or expired key produces a clear error there rather than a mysterious failure the first time somebody asks a question in a meeting.
That is the single most useful control on the page. The failure it prevents is the worst kind: everything looks configured, and the AI simply does not answer.
You will see costs in dollars, because you are paying them.
Who has to do it this way
Self-hosted and on-premise deployments. There is no shared infrastructure to draw on, so the keys have to be yours.
Who might choose it
Organisations with an existing vendor relationship, a negotiated rate, or a procurement rule about where inference happens. Those are real reasons.
"It will be cheaper" usually is not one. Between the vendor accounts, the keys to rotate, the bills to reconcile and the failures nobody notices until somebody complains, the administrative cost is larger than it looks on paper.
Adding a key safely
If you are on the second path.
Use a key made for this, not one already in use elsewhere. When it has to be rotated, and it will, you want to rotate one thing.
Give it the access it needs and no more, if your vendor supports scoping.
Write down where it came from. The most common cause of an AI outage months later is a key somebody revoked during a tidy-up because nobody could say what it was for.
Set a spending limit with the vendor. Not because anything here runs away, but because a limit is the difference between a surprise and an incident.
What happens to what you send
The question every organisation asks second, after cost.
Material sent to a model to answer a question is sent to be answered, not to be kept. It is not added to anybody's training data, and it does not become part of a model other customers use.
On managed, that arrangement is ours to hold and is covered by the same terms as the rest of the platform. On your own keys, it is between you and the vendor you chose, under whatever terms you agreed with them, which is one of the genuine reasons an organisation might want that path.
Either way, the data that reaches a model is the data needed to answer, drawn under the asking person's permissions. Somebody who cannot see a salary does not cause a salary to be sent anywhere by asking about it.
Which providers to turn on
More than one, if you can.
A second provider is a fallback, not a luxury. Model vendors have outages, and an assistant that stops working entirely because one company is having a bad afternoon is an assistant people stop trusting.
Different models are genuinely better at different things, which is what Choosing your model is about. Having two connected is what makes that choice available.
What this page does not do
It does not choose which model answers. That is Model settings.
It does not control what the AI knows about you. That is Teaching the assistant.
It does not set tone or limits. That is Persona and guardrails.
It does not show you what is being used. That is Watching AI usage.
Providers answers one question only: where does the capability come from.
What you do not have to do
Worth stating, because people arrive expecting these.
You do not choose a model per person. It is set for the organisation, which keeps answers consistent between colleagues.
You do not manage capacity. There is nothing to scale, provision or reserve.
You do not have to keep up with the model market. New models become available and are added; nothing you configured stops working because a vendor released something.
One more habit
Write down which path you are on. Six months from now, somebody will ask whether the keys are ours or yours, and the answer determines who to call when it breaks.
If the AI stops working
Start here, and in this order.
Check the provider is still active. On managed, check the allowance has not run out. On your own keys, run the connection test.
Check the model still exists. Vendors retire models, and a retired model produces failures that look like an outage.
Check a second provider is connected. If one is down and there is no fallback, everything stops.
Then look at usage, where a rising error rate will tell you whether this started recently or has been happening quietly for a week.
Did this answer your question?
No, ask a person