For a UK CISO, the shortlist for shadow AI discovery starts with the security platform that already carries your web traffic, plus the Microsoft controls you may already own. Zscaler and Netskope are the two inline platforms worth comparing first, and Microsoft Defender for Cloud Apps with Purview is the sensible baseline for Microsoft-heavy estates. No vendor page proves a winner on detection, so the practical test is a pilot against your own applications, accounts and data flows.
What counts as shadow AI discovery, and what does it need to see?
Discovery means finding which AI tools staff actually use, which accounts they use them with and what data goes in. That is different from blocking a site, and different again from completing a risk assessment. Microsoft's own deployment guidance puts the order plainly: use Defender for Cloud Apps to discover and monitor AI app usage, then decide which apps to sanction or unsanction.
Ask every vendor to show four things on your own traffic. Can it attribute use to a named user or department? Can it tell a corporate AI account from a personal one? Does it see AI features embedded inside ordinary SaaS apps? Does it still work for staff off the office network? A large catalogue of known AI tools is a useful signal, but vendors count them differently, so treat catalogue size as a starting question rather than a score.
Which shadow AI tools give the clearest visibility for regulated firms?
Zscaler describes decisions made per user, per app and per risk across 3,500+ AI tools, with approve, isolate or caution as alternatives to blocking outright. That middle path suits regulated teams that want to keep staff productive while exposure is controlled.
Netskope One offers Skylight AI Security for governance and deep discovery, with an orchestrator that gives a single point for policy management across web, cloud and AI security. That suits firms consolidating several security controls at once. Microsoft covers a different layer, assessing each discovered app through Defender for Cloud Apps so you can sanction or unsanction it. Palo Alto Networks and Cisco also sell AI access controls, but this guide has not compared them, so test them only where they already run your network security.
For governance rather than traffic inspection, ServiceNow AI Control Tower periodically scans connected platforms and adds the AI assets it finds to an inventory as unmanaged until someone reviews them. OneTrust automates privacy impact, vendor and AI risk assessments, including DPIAs. Neither is a substitute for a network or browser sensor, so pair them with one.
How do Netskope and Zscaler compare for detecting and blocking AI apps?
Both belong on the shortlist, and the published evidence does not crown either. Zscaler is the clearer in public on graduated access. Netskope is the clearer on where traffic is processed.
| What you need | Zscaler | Netskope |
|---|---|---|
| Graduated responses | Approve, isolate or caution as well as block | Single orchestrator for policy across web, cloud and AI security |
| Account separation | Confirm corporate and personal policies in a demonstration | Confirm in a demonstration |
| Processing location | Ask for processing and logging locations in writing | Configurable processing zones, all global data centres by default |
Run the same scripted test on both: an approved AI tool, an unapproved one, a personal login and an AI feature built into a SaaS app. Score what each platform actually sees and blocks.
What do UK rules say about monitoring AI use?
Monitoring staff is itself processing of personal data. The ICO says a data protection impact assessment is an ideal way to demonstrate your compliance when AI processes personal data, and that you should consider the nature, scope, context and purposes of the processing. Run that assessment on the monitoring tool as well as on the AI services it watches.
The law has moved recently. The ICO confirmed that most remaining data protection provisions of the Data (Use and Access) Act came into force on 5 February 2026. Ask your privacy lead to confirm which of your templates and notices need updating, and decide whether you need prompt content at all or whether application metadata is enough.
For firms under FCA or PRA oversight, add operational resilience, supplier access and exit terms to the evaluation. Those belong in the contract, not in a slide.
Does UK processing mean your data stays in the UK?
Not automatically. Netskope's documentation says that by default all customers are configured to use all global data centres, and that customers can choose other zone configurations to meet geographic and legal needs. The sensible reading is that UK-only inspection is something you specify and verify, not something you assume.
The same discipline applies to the AI services themselves. Microsoft's announcement says that local inferencing for Microsoft 365 Copilot interactions is expected in the United Kingdom by the end of 2026. A security gateway decides who may reach an AI service, while the location of the model is a separate question for the AI provider. Ask for separate written answers on inspection, logs, management data and model processing.
What does it cost in pounds?
Of the vendors covered here, only Microsoft publishes UK list prices for this kind of control. For firms on Microsoft 365 Business Premium, Microsoft's UK pricing page lists add-ons priced per user per month, paid yearly, and the page states a maximum of 300 seats. The prices were checked on 7 November 2026, so confirm them on Microsoft's own site before you buy.
Microsoft Business Premium security add-ons, £ per user per month
Published UK list price per user per month, paid yearly, excluding VAT, checked 7 November 2026.
View the data
| Category | £ per user per month |
|---|---|
| Defender Suite | £7.70 |
| Purview Suite | £7.70 |
| Defender and Purview together | £11.50 |
Netskope, Zscaler, Palo Alto Networks, Cisco, ServiceNow and OneTrust do not publish comparable GBP list prices for these modules, so request a written UK quote covering licences, connectors, log retention, deployment and renewal. Compare like with like, because a per-user add-on is not the same scope as a fully deployed enterprise service.
How should a CISO choose and run a pilot?
Start with what you already own. Check your Microsoft entitlements before funding overlapping controls, then pilot one inline platform for four to six weeks with a defined user group. Agree the success measures first, such as apps discovered that were previously unknown, personal-account use caught and false positives handled.
Then connect the result to a decision process. Each discovered app needs an owner, a purpose, a risk decision and an assessment. Use governance software for that workflow, and keep human sign-off on every data protection assessment, including any drafted with AI assistance.
What changes next?
Expect more in-country processing options from the large AI providers during 2026, more embedded AI inside everyday SaaS tools and tighter questions from regulators about monitoring and transfers. Build your contract and pilot around specific data flows now, and review them each quarter.
Sources
- Microsoft, Microsoft 365 pricing plans for business, 2026
- Microsoft, Step 1 Discover AI apps, Microsoft Learn, 2026
- Microsoft, in-country data processing for Microsoft 365 Copilot, 2025, updated April 2026
- Zscaler, AI Guard for Users, 2026
- Netskope, Netskope One, 2026
- ServiceNow, Configuring integrations, AI Control Tower, 2026
- OneTrust, Privacy Operations, 2026
- Netskope, Netskope One, 2026
- ServiceNow, Configuring integrations, AI Control Tower, 2026
- OneTrust, Privacy Operations, 2026
- Netskope, Data Transfer at Netskope, 2026
- ICO, accountability and governance implications of AI, 2023
- ICO, commencement of the Data (Use and Access) Act, 2026
