When a research team adopts an AI knowledge base, one decision shapes everything else: where does the AI actually run? There are three common models - fully self-hosted, vendor-managed, and bring-your-own-key - and they trade off data control, cost, and setup effort in different ways. There is no universally right answer; the right choice depends on how sensitive your corpus is and how much operational work your team can take on. This post lays out each model honestly and offers a way to decide.
The three models
- Self-hosted (local or on-prem): the AI models and the whole pipeline run inside your own environment. Documents, queries, and embeddings never leave infrastructure you control.
- Managed: the platform vendor runs the AI for you, typically by calling a hosted provider on your behalf. You get the least setup and the fastest start, with data handled under the vendor's and provider's terms.
- Bring-your-own-key (BYO-key): the platform orchestrates everything, but AI calls run against your own provider account using your own API key, so usage and terms sit with you rather than the vendor.
Data control
Data control is the axis researchers care about most, and the three models sit on a clear spectrum. Self-hosted offers the strongest guarantee: if the work is unpublished, regulated, or export-controlled, keeping it entirely inside your environment removes whole categories of risk because nothing transits a third party. BYO-key is a middle position - data does reach a provider, but under your own contractual relationship and account controls rather than a shared one. Managed is the most convenient and the most dependent on the vendor's and provider's data-handling promises.
Crucially, the run-location choice is separate from tenant isolation. A good platform keeps each tenant's and workspace's data segregated regardless of where the AI runs, so one team's documents are never visible to another. Run location decides whether data leaves your walls; isolation decides who inside the system can see it. You want both.
Cost
Cost behaves differently in each model. Managed and BYO-key are usage-based: you pay per token or per request, which is cheap to start and scales with how much you use. The difference is who holds the bill - with BYO-key it lands on your own provider account, which some teams prefer for budgeting and procurement reasons. Self-hosted inverts the curve: you take on fixed costs for hardware or reserved compute and operational time, but marginal query cost is low, which can be economical at high volume or when data rules make it the only option.
Whatever the model, runaway spend is a real risk with usage-based AI. Hard spend caps - enforced per workspace so a single heavy week cannot produce a surprise invoice - matter more than they first appear, especially for managed and BYO-key setups where a loop or a large batch can rack up calls quickly.
Setup and operational effort
Setup effort runs opposite to control. Managed is the fastest path: no keys to provision, no infrastructure to stand up, useful within minutes - ideal for evaluating quality before committing. BYO-key adds one step, configuring your provider key, in exchange for keeping usage on your account. Self-hosting asks the most: provisioning compute, deploying and updating models, and owning availability. That is real work, and it is worth being honest that it requires capacity many smaller teams do not have - which is exactly why having all three options matters.
Which fits sensitive vs non-sensitive corpora
A practical way to choose is to sort your documents by sensitivity and match each tier to a model:
- Non-sensitive (public papers, general background): managed AI is usually fine and gets you moving fastest.
- Moderately sensitive (internal notes, pre-print drafts you control): BYO-key keeps usage under your own account and terms while staying low-effort.
- Highly sensitive (unpublished IP, regulated, NDA, or export-controlled data): self-hosted or on-prem, so nothing leaves your environment.
Many teams use more than one model at once - managed for an open literature workspace, self-hosted for the confidential corpus - which is why flexibility beats lock-in. If you are weighing whether a general assistant is safe for confidential work at all, our post on ChatGPT and confidential research goes deeper, and our piece on why AI fabricates citations explains why grounding answers in your own sources matters regardless of where the model runs.
How Researcher Center fits in
Researcher Center supports all three models - platform-managed, bring-your-own-key, and fully local or self-hosted - so you can match the run location to each corpus instead of compromising on one setting for everything. Per-tenant and per-workspace isolation keeps data segregated whichever model you pick, and hard spend caps keep managed and BYO-key usage predictable. Answers stay grounded in your own documents and cited to the passage they came from, with the system declining when the evidence is missing.
The best way to find your fit is to start small: pick a non-confidential workspace, try managed AI to judge answer quality, then decide which corpora justify BYO-key or self-hosting. Early Access is an invite-only free trial - request access on our pilot page and configure the model that matches your data.