Private AI: checking client requirements

File storage answers only part of the question

Suppose a client has approved document processing in a designated environment for one engagement. A consultant uses an application that stores files there. IT discovers, however, that the selected model service processes requests globally. Pointing to the folder does not explain where work on the content takes place.

IT presents that difference to the person responsible for the client arrangements and environment approval. That person decides whether the proposed workflow meets the agreed requirements or needs changing. In this example, the requirement comes from the project arrangements. The client's country alone does not define the application's configuration.

Vendor documentation makes this distinction too. Microsoft describes deployment types for Foundry Serverless API models, separating storage geography from inference processing location: global, data zone, or geography. Availability depends on the model and deployment type. This description applies to Serverless API.

One workspace, several processing locations

The consultant sees a document and a question box. Behind that view, the application may extract text and search for passages before sending them to the model that produces an answer. The buyer needs an understandable account of that path for the specific client's workspace.

The person approving the environment should be able to establish which services receive content and where saved questions and answers remain. Backups and access for people supporting the tool also belong in that picture. They can then compare the actual workflow with the requirements already agreed.

Employees need to see which client's workspace they are using. Selecting a project should lead to the processing path intended for it. An administrator's model change may alter that path, so its effects need to be checked against the project arrangements again.

Work Syntalith can undertake

Syntalith proposes an AI application with workspaces fitted to individual clients. The starting point is a description of the current document flow, from retrieving a file to saving an answer, including the services involved. We compare it with requirements supplied by the company and establish which connections need changing. The person responsible for those requirements decides whether to approve the environment.

Further scope could cover workspace selection in the application, routing requests to an agreed service, and a clear record of how work for the client was processed. The team supplies project requirements and a description of consultants' tasks. The administrator identifies existing connections and agrees on their ongoing support arrangements.

Our article on local AI for NDA-bound client projects discusses handing documents to the next person and keeping clients' material separate. Here, the first question is which processing path a workspace uses.

A managed service or a separate environment?

If an available managed service offers processing that meets the approved requirements, selecting the appropriate option and connecting it to the application may be enough. Different client arrangements do not always require separate servers. Check the current solution's actual configuration before planning to replace it.

A separate private environment, including a company-owned server, is an option when requirements and team workload justify it. Alongside processing location, consider how many consultants need answers at once and how long they can wait. A trial using their tasks can help establish whether the selected approach fits daily work.

After launch, someone needs to track changes to services and application connections. Support scope should identify who checks a changed processing path after an update and whom they notify if it differs from project arrangements. Both managed services and company-operated environments need that allocation of responsibility.

Tell us about one client requirement that is difficult to trace through the current AI application. The first conversation can cover it alongside the document's path, without sharing confidential content. Pricing information is available on our pricing page.