Illustrative cover photograph: imgix / Unsplash. Not a client engagement.
Define what private needs to mean
Private AI is not a single deployment model. For one business it means prompts never leave a controlled network. For another it means a contracted provider may process them, but cannot use them for specified secondary purposes. Those requirements lead to different designs.
Write the boundary you need before choosing hardware. Include prompts, uploaded files, retrieved text, embeddings, generated answers, logs, backups and support access. A model running locally tells you where inference happens. It does not, by itself, tell you where the rest of the application sends data.
Follow the data, not the product label. This is a deployment decision. The build-versus-buy guide addresses acquisition choices, while RAG versus fine-tuning addresses how a model uses business knowledge.
Compare the operating responsibilities
For this guide, a local model runs on hardware your organisation controls, such as a workstation or on-premises server. Private cloud means a model and application deployed within a cloud environment configured for your organisation; the precise isolation and provider responsibilities still need verification. A hosted API means you send requests to a provider-operated model service under its applicable configuration and terms.
| Option | What you need to establish | Operating work to budget |
|---|---|---|
| Local model | Device access, outbound connections, backups and physical protection | Hardware capacity, patching, model updates and recovery |
| Private cloud | Tenant isolation, regions, identity, network paths and administrator access | Cloud configuration, monitoring, capacity and shared-responsibility gaps |
| Hosted API | Processing locations, retention, permitted data use, subprocessors and support access | Integration, contract checks, access control and provider-change monitoring |
None of those labels establishes a compliance outcome. The NCSC’s secure-design guidance recommends considering threats, deployment environment and security alongside functionality and performance. It also calls for provider due diligence and controls on information sent to external APIs.
The table is a decision aid, not a claim that every provider supports the same controls.
Trace a document assistant end to end
Take a fictional internal assistant answering questions about supplier agreements. The intended users are procurement staff. The relevant restriction is that only authorised staff can retrieve each agreement, and contract text must stay within approved processing arrangements.
Start at upload. Where is the file stored, who can open it and what happens to deleted files? Then inspect parsing and embedding. Are these local services or separate external calls? Where does the search index live? Which text passages are sent to inference, and is access checked before they are selected?
Continue beyond the answer. Does the application log the complete prompt? Does an error-reporting service collect document excerpts? Are transcripts copied to analytics, support tickets or backups? Who can inspect those copies, and for how long?
A local inference server does not make an external logging service local. Conversely, a hosted service may be appropriate for an approved low-sensitivity workflow if its actual terms and configuration meet the requirement. Make that judgement from evidence, not from a cloud-versus-office preference.

Ask for evidence at each boundary
Ask vendors and your own engineers the same concrete questions. Identify the exact service tier and configuration being evaluated. Avoid treating a sales statement about one offering as a commitment covering every product the supplier sells.
Record where content is processed and stored, whether and how it is retained, whether it is used for training or other purposes, who can access it, what deletion covers and which contractual terms govern the answers. “Not used for training” does not answer retention, access or location questions.
For local and private-cloud options, request a recovery demonstration, update procedure and named operational owner. Inspect who can install models and dependencies. The NCSC advises treating imported third-party models or serialised weights as untrusted components requiring suitable checks and isolation. Running them on your own server does not remove the supply-chain question.
For UK and UAE use, identify the applicable sector, contracts, people and processing locations with qualified privacy or legal advisers where needed. Do not assume that all UAE business data must remain in-country, or that choosing a UK region settles every obligation. This article does not determine the legal requirements for your organisation.
Test the workload before buying the infrastructure
Use representative, authorised test documents and questions. Include difficult cases: conflicting agreement versions, no answer in the source, access denied, long documents and simultaneous requests. Measure answer quality and source accuracy as well as latency. A fast wrong answer is not an infrastructure success.
Set practical acceptance criteria with the users. Can the proposed model answer the actual questions reliably enough with review? What happens when it cannot? Test the complete application, including retrieval and permissions, rather than benchmarking the model in isolation.
Estimate capacity from measured requests and concurrency. A quiet demonstration with one user says little about a deadline morning when twenty people ask questions at once. Record warm-up behaviour, queues, timeouts and fallback. A fallback to an external API must obey the same approved data boundary; it should not quietly bypass the reason you selected a private deployment.
Include the person keeping it running
Compare costs over the same period and workload. Include hardware or cloud capacity, service usage, storage, monitoring, backup, implementation and operating labour. Keep cash expenditure separate from allocated staff time, and show uncertainty rather than burying it in a single precise number.
For an illustrative GBP monthly comparison, a hosted option has £300 of service usage and 6 operating hours at an assumed £60 per hour: £660. A private-cloud option has £800 of infrastructure and 16 operating hours at £60: £1,760. A local option allocates £250 for hardware, power and backup plus 20 operating hours at £60: £1,450.
These invented figures show the arithmetic only. They are not supplier prices, total implementation budgets or evidence of equivalent capacity. They exclude tax and initial setup and assume each option has already passed the same quality and boundary tests. Replace them with quotes and measured effort. For an AED comparison, collect AED inputs rather than applying an unrecorded exchange rate.
The cheapest option that fails a mandatory requirement is not a saving. But buying local hardware without anyone available to patch, monitor and restore it is not a privacy strategy either.
Record the choice and the exit
Use hard requirements to eliminate unsuitable options before comparing preferences. Then record the selected deployment, evidence still missing, accountable owner and review triggers. A change in document sensitivity, supplier terms, workload or model capability should prompt another look.
Test portability while the deployment is small. Can you export documents, access rules, prompts and evaluation cases? Which parts must be rebuilt for a different embedding model or inference service? Keep a usable copy of the evidence and configuration you are entitled to retain.
The editable data-flow worksheet includes the contract-assistant example, boundary questions, blank rows and the three illustrative cost calculations. It contains plain values, no formulas or external connections. Complete the flow first. The deployment choice should be the consequence of that work.
Download the editable worksheet (CSV) ↓
Frequently asked questions
Can a local model still send data outside the business?
The surrounding application may use external parsing, embeddings, telemetry or support services. Inspect those connections and the device’s outbound behaviour rather than assuming local inference covers the whole workflow.
Does no training mean no retention?
These are separate questions. Verify the applicable service terms and configuration for retention, access, location and deletion as well as training use.
Can different workflows use different deployment models?
Yes, if the data classification, access rules and processing arrangements support that separation. Define routing explicitly and test that a restricted workflow cannot silently fall back to a less-restricted service.
Sources checked 8 November 2026. Worksheets and worked examples are suggested working methods, not official standards or reported client results.

