Write the job before choosing the model

A company does not need a model in the abstract. It needs a useful task completed under specific conditions. Write that task as a short scenario: an authorised employee asks a question about an approved manual and receives an answer with the relevant source. Then describe a failure: the manual is outdated, the question has no answer or the employee cannot access the document. These scenarios make deployment choices easier to assess than a list of impressive features.

Specify what the system may do with its answer. Finding a document is different from changing a customer record. Preparing a draft is different from sending it. Those boundaries affect permissions, review and recovery. If the task can be solved by structured search or a simpler workflow, keep that option in the comparison. AI should earn its place by performing a defined job, rather than become a requirement before the problem is understood.

Draw the complete data path

List the source documents, retrieval index, model endpoint, prompts, outputs, logs and backups. For each, name its location, the people who can access it and the service responsible for processing it. Add supporting systems such as error reporting and external tools. A private model does not make every surrounding service private. An architecture diagram is useful only if it includes the paths information can take beyond the main inference call.

Use this map to define what must stay inside your environment and what, if anything, may leave it. Do not rely on the label enterprise or private as a substitute for documented settings and terms. With a hosted service, examine the specific product’s data handling and administration options. With a self-hosted system, examine storage, access and external connections. The useful question is whether the whole arrangement meets your requirements for this task.

Keep permissions attached to the evidence

A knowledge assistant creates a new route to existing information. Design that route so source permissions still apply when a document is retrieved and when an answer is assembled. Test a user who should not see a document, a user whose access has changed and a document removed from the collection. A permission check on the visible interface is insufficient if the underlying retrieval path can still expose restricted content.

Treat source quality as an operating responsibility. Someone must approve the collection, identify current versions and remove obsolete material. A citation helps a person inspect evidence, but a cited old policy is still an old policy. Make missing or contradictory information a valid outcome. The system should be able to say that it cannot establish an answer, and the interface should give the user a useful next step without inventing certainty.

Private infrastructure still needs an operator

Self-hosting gives the business a different set of responsibilities. Name who maintains the environment, controls access, applies updates and restores service when something fails. Establish how the system will be observed and what information its logs should contain. Assess model and hardware suitability with the intended workload rather than an isolated demonstration. A prototype running once is not evidence that an unattended production service will be manageable.

A hosted service can be sensible where the task and data boundary allow it. It still needs account ownership, configuration review and a plan for changes in the dependency. Compare the complete operating effort, including support and evaluation, instead of only the visible subscription or hardware expense. A hybrid architecture also needs a clear route for each kind of data; it should not become an unexplained compromise that nobody can describe.

Evaluate before expanding access

Create a small evaluation set from representative questions and agreed expected behaviour. Include straightforward tasks, ambiguous requests, missing information and attempts to retrieve restricted sources. Record the source collection and configuration used for each run so later changes can be assessed. Decide which failures block release and which can be managed through a clearly explained review step. The consequence of an error should shape the release decision.

Run the prototype with people who understand the task and can challenge plausible answers. Ask them whether the output is useful, inspectable and within the system’s authority. If a workflow writes to another system, test approval and recovery separately from answer generation. Expand the source collection or user group only after the narrow use case has been evaluated. More documents and more users create new conditions; they are not merely a larger version of the same demonstration.

Separate infrastructure control from privacy review

The European Commission explains that GDPR applies to personal data processing regardless of the technology used. Keeping a system inside your infrastructure does not answer every question about the processing. Your privacy lead should review the intended use and responsibilities before release. The technical team can provide the data map, access design and operating controls needed for that review. The Commission’s overview is available at https://commission.europa.eu/law/law-topic/data-protection/data-protection-explained_en.

A useful next step is a short decision document: the job, permitted data path, access rules, evaluation criteria and operating owner. Compare hosted, private and simpler non-AI approaches against that document. Bridgera can scope a focused prototype or private deployment around these decisions. The result should be a system your business can understand and operate, with explicit boundaries around its data and actions, rather than a deployment chosen because its label sounds reassuring.

Give the decision an owner and a review trigger. A new source collection, an external tool or a wider user group may change the original boundary. Record who can approve that change and which checks must be repeated. This turns the architecture into an operating agreement rather than a one-off diagram. It also gives procurement and technical teams a shared way to compare proposals: ask each supplier to explain the same task, data path and responsibilities before comparing the presentation.