Skip to content

On-premises, cloud, or hybrid? Where do we build the next project?

B2B requests and documents portal between a local server rack and connected cloud infrastructure

Series II: From infrastructure to agentic commerce projects in production. Episode 01 of 20.

A B2B portal for requests and documents may seem like a single application. In reality, the interface, database, files, backups, and any AI service all have distinct requirements. For a small or medium business, the useful decision isn’t “do we move everything to the cloud?” but rather where each component runs, who manages it, and what happens if a connection fails. The 90-day adoption plan helps select the process; here, we set where to build it.

Three terms, one component-based decision

On-premises means infrastructure on the organization’s premises and under its administration: servers, storage, network, power, and backups. Public cloud means on-demand resources provided by an operator, with elastic allocation and pay-as-you-go consumption. Renting a virtual machine alone doesn’t deliver all the benefits of managed services; equally, a group of in-house servers doesn’t automatically create a “private cloud.” According to the NIST definition, the latter means a cloud infrastructure used exclusively by one organization, which can also be hosted off-site.

In project terminology, hybrid architecture distributes components between on-premises infrastructure and cloud services. The strict NIST term “hybrid cloud” requires two or more distinct cloud infrastructures, linked to enable data and application portability. This distinction matters when comparing offers: the label alone doesn’t describe migration capabilities. None of these options automatically guarantees security, control, or low costs.

Five questions before choosing

Criterion

Question to check

What location might change

Data & access

Where are documents stored, who can process them, and what data leaves the environment?

Keeping data on-premises may require secure links to a cloud application.

Latency & connection

How long do real operations take between application and database, including under degraded connections?

Components that frequently exchange data may need to run close together.

Availability

What happens if the site, provider, or their connection fails?

Backups and recovery mechanisms must be designed for the relevant failure.

Operations

Who handles updates, monitors incidents, and can intervene outside business hours?

Managed services reduce some tasks but don’t remove the company’s responsibility.

Total cost

What do we pay for hardware, energy, team, consumption, transfer, recovery, and exit?

The winner can differ by volume, traffic variations, and project duration.

The matrix is a discussion tool, not a universal score. For each criterion, note the measurable requirement, available proof, and who approves the trade-offs. If traffic or requirements are unknown, start by measuring with a pilot, not by drawing an architecture based pure on vendor preference.

Inventory your workflows before requesting offers: who uploads files, where processing happens, which systems the portal queries, and how often data crosses environment boundaries. Map authentication and administration flows separately. A system might store files on-premises yet send names, addresses, text fragments, or telemetry to other services. Check these pathways with your data team and providers, then restrict access to only what’s needed for each function.

Hypothetical example: the partner portal

Suppose the company receives requests and documents from distributors, and an operator validates each response. An AI agent integrated into the portal could classify requests or suggest summaries, but not approve documents on its own. The AI agents used by the dev team to write and test code are a separate matter; where their tools run doesn’t automatically determine where partners’ documents go.

One option to consider, not a pre-validated recommendation, is to place the UI and API—the layer through which systems exchange requests—in a managed cloud service. The database and documents stay, initially, on in-house infrastructure, accessible via a controlled connection. The external AI service receives only permitted and minimized data fragments; if this cannot be ensured, we test local inference—running the model on the company’s hardware—or postpone the function. Logs, identity data, metadata, and backups must be tracked separately: “documents are local” doesn’t tell you where all the project’s data ends up.

This setup can work poorly if each page view crosses the link to the database. The practical test is to measure end-to-end operations under realistic traffic, including uploading and searching for documents. Test what happens if the connection fails: does the portal show a clear error, queue the request, or stop uploads? Choose based on business need and test results. One alternative is to move both the API and the data together to a suitable environment, with the required controls; another is to host the entire portal on-premises. Neither option should be chosen just because it looks simpler on a diagram.

Responsibility doesn't leave with the server

In the cloud, the provider is responsible for certain layers of the infrastructure, while the company remains responsible for configuration, identities, applications, and data, to the extent specified by the chosen service. For a virtual machine managed by the company, OS updates may still be up to them; a managed service shifts more operations to the provider. The contract and concrete architecture determine the details. On-premises, the company has to cover equipment, power supply, connectivity, component replacement, and incident recovery as well.

In every setup, decide who can read the documents, where the encryption keys are kept, what logs are retained, and how restoration is verified. A private connection does not replace access control. If using an external AI model, explicitly check the request path, data retention, and provider terms before sending actual content. Don’t assume all data is stored in the region chosen for the database.

Costs and the decision sheet

Compare two or three scenarios over the same period and with the same availability requirements. For on-premises, include equipment purchase and replacement, space, power, cooling, licenses, staffing, maintenance, and backups. For cloud, factor in resources, managed services, storage, traffic between environments and outbound, monitoring, and support. For both, add migration, recovery, testing, exit strategy, and AI consumption. Don’t turn an hourly rate into a “total cost” without traffic volumes and usage policies. Case studies on sizing can show approaches, but don’t pick a winner suitable for every company.

The result of this episode is a worksheet with a row for interface/API, data, documents, AI, logs, and backups: proposed location, rationale, responsible party, network dependence, estimated cost, and fallback for failures or migration. Mark your assumptions and what needs to be measured in the pilot. Update the sheet if volume increases, connection becomes a bottleneck, or the team can no longer support operations. In the next episode, we’ll choose the runtime form: from virtual machines to containers, and only if justified, Kubernetes.

Sources

Choose the right infrastructure for your project

Want to select infrastructure for a specific project? Talk to the i8 team about a decision worksheet and a measurable pilot.