Start with the work. Then choose the AI boundary.
Use this guide to match a real workload to one of our 14 publicly available products, decide where the data should run, and define the human review required before an output is used.
Reviewed 23 August 2026
Every named product has a verified public acquisition path. Examples describe appropriate workflows, not customer outcomes, sector approval or professional advice.
Choose the route before the model
The same task can need a different architecture when its inputs, users or consequences change. Record these answers before evaluating output quality.
What data enters?
Classify documents, personal data, recordings, images, intellectual property and secrets. Use only content you are authorised to process.
Where may it run?
Choose on-device models for one user, customer-hosted AI Server for shared inference, or an optional provider only when its terms and data route are approved.
What happens next?
Separate low-impact assistance from outputs that affect rights, safety, money, employment, health, education or access to essential services.
Who verifies it?
Name the person accountable for checking sources, facts, permissions and fitness for purpose. “A human is involved” is not a review procedure.
Match the job to a public product
These are starting configurations. Confirm platform support, model quality and the exact data path during a pilot; optional cloud-provider routes have separate terms and boundaries.
| Work to do | Best starting product | Practical data boundary | Human review before use |
|---|---|---|---|
| Ask questions of PDFs, notes and reference documents | AI PDF Reader AI Notepad |
Use a local model on the reader's device. For a shared internal knowledge service, evaluate AI Server and the client that will retrieve the approved sources. | Open the cited page or note. Check that the source supports the answer and that relevant documents were not omitted. |
| Draft, rewrite, summarise and improve tone | AI Rewrite AI Assistant AI Notepad |
Keep drafts on-device with a local model when the work is confidential. Remove unnecessary personal or restricted data before any optional provider route. | The named author verifies facts, meaning, tone, confidentiality and approval status before sending or publishing. |
| Translate text, documents and images | AI Translate | Use the offline local path for private content. A deliberately selected cloud provider receives the selected content under that provider's terms. | A fluent or qualified reviewer checks names, numbers, terminology and meaning for contracts, healthcare, safety or regulated publication. |
| Explore options, reason through problems and plan | AI Reasoning AI Chatbot |
Run locally for individual analysis. Keep tool or Agent actions within approved systems and permissions. | Treat the output as a proposal. Test calculations and code, inspect sources, challenge assumptions and keep the final decision with an accountable person. |
| Transcribe speech or generate a voice | AI Dictation AI Voice Generation |
Process recordings and generation on-device. Establish permission to record and the right to use any voice sample. | Check the transcript against the audio. Obtain required consent and clearly govern or disclose synthetic voice use. |
| Create, caption, search and prepare images | AI Image Generation AI Image Tool |
Keep source images, prompts and generated files on the Windows device with local models. | Check rights, licences, likenesses, sensitive details, accuracy of captions and any disclosure required for synthetic media. |
| Give teams or internal applications private AI APIs | AI Server AI Admin Console |
Run inference on customer-controlled Windows, macOS, Linux, Docker or Kubernetes infrastructure. Add the Console when organisation identity, entitlement, policy and fleet visibility are needed. | The operator controls keys, network exposure, TLS, models, retention and monitoring. Each consuming application still owns its output checks and fallback. |
Useful patterns, not sector certification
An industry label does not make a deployment suitable. These examples identify narrow assistance tasks; the organisation must assess its actual data, law, professional duties and risk.
Legal and professional services
Search cited source documents, prepare first drafts, compare wording and translate working material inside an approved boundary.
Boundary: Not legal advice, privilege management, conflict checking or a substitute for lawyer review.
Finance and insurance
Summarise internal material, query cited documents, draft explanations and explore analyst hypotheses.
Boundary: Not an autonomous credit, claims, trading, investment or customer-eligibility decision.
Healthcare and life sciences
Search approved literature, transcribe or draft working notes, translate non-final material and organise research documents.
Boundary: Not diagnosis, treatment, triage, clinical validation or a medical-device claim.
Public sector and critical environments
Evaluate internal knowledge, drafting and translation on customer-controlled or tested offline infrastructure.
Boundary: Not an autonomous eligibility, enforcement, immigration, policing or essential-service decision.
R&D and manufacturing
Work with protected specifications, manuals, images, code, plans and multilingual documentation without defaulting to a public model endpoint.
Boundary: Engineering, quality and safety owners verify every consequential instruction or change.
Education and non-profits
Create notes, translate material, improve accessibility with speech tools and give staff a private writing assistant.
Boundary: Not autonomous grading, admissions, safeguarding or decisions about access to support.
Do not automate the final consequential decision
Local deployment reduces one data-transfer risk. It does not make a model accurate, unbiased, lawful or authorised for a particular decision.
Before the pilot
- Define allowed and prohibited inputs, users, actions and decisions.
- Name the accountable owner and qualified reviewers.
- Prepare representative tests, including failure and misuse cases.
Before production use
- Set confidence limits, escalation and a non-AI fallback.
- Record model, prompt, source and configuration changes.
- Monitor quality and harm, and give the owner authority to stop use.
For a technology-neutral risk process, use the NIST AI Risk Management Framework. It is a voluntary framework, not a Software Tailor certification or legal opinion.
Prove one workload before you scale
A useful pilot produces evidence for a go, change or stop decision. Do not score only the best demonstration prompts.
Measure task success, source support and severe-error rate on representative and adversarial examples.
Verify the actual network route, content handling, storage, optional providers and organisation metadata.
Measure latency, throughput, hardware headroom, update effort, monitoring and recovery.
Time the real review process and test escalation, accessibility, training and non-AI fallback.
Document accepted limits, open gaps, owner, evidence, rollback trigger and the next review date.
Direct answers before you choose
Does private AI mean every use is fully offline?
No. The local-model path runs on the device, and AI Server runs on customer-controlled infrastructure. Optional cloud models and organisation services have separate, documented data paths.
Which app should I start with?
Choose the narrowest specialist app that fits one repeatable task. Choose AI Suite for several Windows workflows, or AI Server when multiple people or internal applications need shared inference.
Does running locally make answers more accurate?
No. Location changes the deployment and data boundary, not the truth of an answer. Model choice, sources, prompting, testing and human review still determine whether an output is fit for purpose.
Can these products be used in a regulated organisation?
They can be evaluated for a defined workload and control environment. Suitability depends on the use, data, configuration, jurisdiction and organisation's duties; no product makes the deployment compliant by itself.
Do individual users need AI Admin Console?
No. It is for organisation administration: identity, members, entitlements, policy, enrolled servers, usage and audit. Individual apps and a simple AI Server evaluation do not require it.
What should procurement review next?
Review the Trust Centre, exact product page, public catalogue, Privacy Policy and relevant AI Server operations guidance. Then test the real workload and boundary.
Turn the evaluation into procurement questions →Bring one real workload and its constraints
For a team deployment, record users, data, model capabilities, concurrency, latency, hardware and network boundary. The deployment planner turns those inputs into a reviewable starting architecture.