From business case to a private AI application.

Establish what must improve first. Then choose the data, authority and technical shape that fit within accountable boundaries.

From decision to delivery

A private environment starts with a specific capability.

This page addresses the step after initial AI assessment: shaping a bounded application that uses organisational data for a defined operational purpose.

‘Private generative AI’ is neither a product label nor an automatic guarantee here. It means an arrangement with demonstrable boundaries around sources, permissions, storage, processing, supplier and administrator access, logging and outbound data flows.

The existing generative-AI page helps determine whether AI is appropriate. This page sets out what is required to deliver and operate an agreed business case responsibly.

Decision brief

Make assumptions testable before technology sets the direction.

A demonstration may show that something can work. A business case must establish whether it creates sufficient value in the live process, at an acceptable cost and level of failure.

Process and user: Which constraint is being removed, for whom, and what are the current quality, lead time and workload?
Outcome and capability: Must the application retrieve, summarise, draft, classify, compare or carry out a bounded action?
Benefits and whole-life cost: Include setup, licences, integration, security, oversight, adoption, operations and exit; do not rely on unproven savings percentages.
Failure and acceptance: Define critical failure modes, tolerances, test cases and authority to reject before delivery begins.
Alternative: Process improvement, search, decision rules or conventional automation may be cheaper, sufficiently reliable and easier to explain.
Ownership and go/no-go: A named accountable person decides on evidence from the intended setting and may stop introduction, scale-up or continued use.

From source to operation

Permitted data is connected for a defined purpose.

A private data silo is a bounded input, not an unrestricted pool of knowledge. Each application specifies which source is available for which purpose and role.

Business case

Capability and measure

The operational objective determines what information and output are required.

Data & governance

Sources and permissions

Classification, personal data, ownership, authorisation and retention govern access.

Private application

Storage and processing

Model use, prompt and log storage, integrations, supplier and administrator access are configured explicitly.

Assurance & operations

Sources and evidence

Citations, logs, tests, monitoring and change control make use traceable and recoverable.

Governance as a decision framework

Decide in advance what data may and may not do.

Purpose limitation must become an operational and technical rule, not remain a policy statement.

The data owner approves sources. The process owner owns the measure of success and failure tolerance. Technology and security teams implement permissions, logging and change control. A named decision owner accepts residual risk and controls the go/no-go.

Where personal data is involved, requirements such as lawful basis, purpose limitation, data minimisation, transparency and security remain relevant, with a data protection impact assessment where required. AI Act duties also depend on role, intended use and risk class. Private or European hosting does not by itself demonstrate compliance.

Source and purpose

Permit only data that is necessary and suitable for the defined capability. Separate production, test and training data.

Access and model use

Define roles, least-privilege access, technical administrator access and whether data may be used to improve a model.

Retention and traceability

Set retention for prompts, outputs and logs. Preserve citations and decision records where assurance requires them.

Supplier, jurisdiction and exit

Assess storage and processing, remote support, subcontractors, outbound connections, portability and deletion. See also Data sovereignty.

Purpose-built applications and processes

Independent execution is not unlimited decision authority.

An application may retrieve information, draft content, route a case or carry out a predefined action when its purpose, permissions and impact limits are explicit.

Low-risk work may proceed automatically within approved rules. Material consequences, uncertain evidence, special-category personal data or exceptional transactions require human review or a hard stop.

Every executing application needs monitoring, a named owner, a stop control, a recovery procedure and an escalation path. Changes to models, prompts, sources or integrations trigger targeted retesting.

Illustrative example — not a client case

An internal assistant for service reports.

Suppose staff spend too long searching approved manuals and previous service reports. The required capability is to retrieve relevant passages and draft a response with citations — not to make a customer decision independently.

Measure current search time, findability and correction effort, then set acceptance criteria for citation accuracy and incorrect responses.
Connect approved document collections only. Personal data and confidential customer records remain out of scope unless purpose, lawful basis, authorisation and retention are separately established.
Staff review the draft and its sources. Only low-risk routing may run automatically within a defined mandate.
Monitor errors, source changes, use and cost. Stop or recover if the measure of success, data boundary or control position is no longer met.

The valid outcome may still be that conventional search is sufficient, the trial needs redesigning or the application should not proceed.

Law and governance frameworks

Apply proportionately; do not presume evidence.

The GDPR and AI Act require an assessment suited to the data, role, use and risk. NIST AI RMF and ISO/IEC 42001 can structure governance and lifecycle controls; referring to them is neither certification nor proof of legal conformity.

Connected decisions

Explore the adjacent choices.

Make the business case, data and accountability decision-ready.

Discuss your AI question