SaaS, private or hybrid deployment: how should an enterprise choose?

What each form actually solves

SaaS

The supplier runs the platform and the company uses it through accounts. It starts quickly, upgrades centrally and costs less up front, which suits a pilot and well-bounded general work. Check the data processing scope, the supplier's model policy and account permissions.

Private deployment

The platform and its execution resources run in an environment the company nominates. You gain control over network, storage and operations, and take on deployment, monitoring, backup, upgrades and capacity planning. Private deployment is not safer by itself: misconfiguration and over-broad permissions still apply.

Hybrid

Identity, projects, execution records or sensitive data stay inside, while some models and elastic resources are used externally under rules. It balances boundary against model choice, but demands more of connection policy, data classification and fault isolation.

Settle these before choosing

Which data may not leave the nominated environment?

Inputs, context, logs and outputs can all carry sensitive data. Checking the prompt alone misses execution files and logs. Where personal information has to cross the border, article 38 of China's Personal Information Protection Law sets out the routes that make it lawful — a security assessment, certification, a standard contract, or another condition in law — which decides whether an overseas model is usable at all, not merely where the platform runs.

Which systems does the agent connect to?

If an agent reads and writes repositories, tickets, ERP, CRM or production APIs, the architect handles network and authentication together. Draw the connections to see which components must sit close to the data. How those connections are then authorised is covered in designing agent access and audit.

Who runs it afterwards?

A private environment needs named people for upgrades, vulnerability fixes, monitoring and restore drills. Without them the system quietly stays on an old version.

How fast do models and the work itself change?

Early on, models and tasks change often. Where the data boundary allows it, run on SaaS first and move the parts that must stay inside once the work settles.

Three things that are easy to miss

Industry does not decide the form

Inside the same bank or manufacturer, tidying public material and processing production data have completely different boundaries. Task and data classification is more useful than an industry label.

A model on the internal network does not mean the data stays there

External tools, update services, monitoring and human operations can all form a data path. An architecture review has to count those connections.

Hybrid means more connections to manage

A hybrid setup involves internal and external identity, routing and logs at once. When that configuration is scattered, faults and over-permission are harder to trace.

Order of work

• pick one frequent, reviewable task;
• list its inputs, context, execution environment, external connections and outputs;
• mark where each class of data is allowed to appear;
• compare the three forms on time to launch, operational responsibility and cost of growth;
• use a pilot to test performance, permissions, failure recovery and handover of results;
• widen the scope or adjust the architecture based on what the pilot showed.

A deployment decision deserves a written basis. When data classification, model supply or business connections change, the teams involved review it again.

Common questions

When is it fine to start on SaaS?

For low-sensitivity work such as tidying public material, for well-bounded general tasks and for a pilot. Check the data processing scope, the supplier's model policy and account permissions first.

Is private deployment automatically safer?

No. You gain control over network, storage and operations, and you take on deployment, monitoring, backup, upgrades and capacity planning. Misconfiguration and over-broad permissions remain possible.

Does the industry decide the deployment form?

No. Within one company, tidying public material and processing production data have completely different boundaries, so task and data classification is the more useful test.

If the model runs on our network, does the data stay there?

Not necessarily. External tools, update services, monitoring and human operations can all form a data path, and an architecture review has to count them.