Find the useful job: define success and stopping points
We start with a repeatable task and representative examples supplied by your team. A knowledge assistant might answer questions from approved policy documents; a workflow might extract named fields for human review. We identify the user, the input, the desired output and the cost of an incorrect response. The initial scope stays narrow enough to evaluate meaningfully before adding further tools or sources.
An evaluation plan records expected answers, relevant source passages and cases where the system should decline. We include incomplete documents, conflicting information and requests outside the agreed task. The written scope defines the acceptance evidence and who reviews it. This makes usefulness assessable without promising that a model will always be correct or replacing professional judgement with generated text.
Understand the data: draw the network boundary
We map source systems, document permissions, update rules and retention requirements. For a fully private deployment, model inference runs on your own server or client-controlled infrastructure. Document parsing, embeddings, retrieval storage and response generation must also remain inside that agreed environment for the document path to stay private. We record these components separately so an external embedding service cannot hide behind a locally running model.
The architecture lists outbound connections, including hosted models, integrations, telemetry and software update requirements. In a fully private configuration, company prompts, documents, retrieved text and outputs are kept within the agreed network boundary; any exception requires an explicit design decision. A hosted or hybrid option identifies exactly which data is sent to which service. Hardware suitability and model licence terms are checked before selecting the deployment. The tools below are implementation options, with the final components selected against those requirements.
Prototype and evaluate: make limits visible
We build the agreed assistant or workflow and review it through controlled demos on a live URL. Preview access is restricted where the material requires it. Answers should expose the supporting sources where retrieval is used, and indicate when evidence is missing. We compare responses with the evaluation set, review failure patterns and distinguish retrieval problems from model behaviour before changing the system.
User identity and source permissions shape retrieval before context reaches the model. We define allowed actions separately from chat access. An agent that can draft a record need not be allowed to publish or delete it. Tools use scoped credentials and explicit approval points for consequential actions. We test permission boundaries and instruction-like content in documents as well as ordinary questions, with unresolved risks recorded before rollout.
Integrate responsibly: operate the system you own
The release scope defines model serving, access control, monitoring and recovery responsibilities. Logs are chosen for diagnosis without collecting unnecessary document text or secrets. We decide who can inspect them and how long they remain. Monitoring covers the agreed signals, such as failed retrieval, service errors and changes in evaluation behaviour. Updating a model or source pipeline triggers the relevant evaluation again rather than being treated as a cosmetic change.
Your team receives code, configuration, operating instructions and account ownership under the agreement. The handover records model and component licences, hardware requirements and external service dependencies. A prototype may lead to a wider deployment, but that expansion is a new scope decision. Connecting Automation & systems or Web & development can provide the surrounding workflow and interface without removing the AI system’s permission boundaries.