Our security team has blocked public AI tools. What should we evaluate?
15 min read
Public AI tools blocked by security? Compare enterprise tiers, a private workspace on approved cloud, self-hosted models and a routed mix, with a checklist.
Evaluate four options: the enterprise tier of the tools that were blocked, a private workspace that reaches models through your own cloud account, open-weight models on hardware you run, and a routed mix that sends each request to one of those by rule. The reason for the block decides which one fits, and most organisations end up with two: an enterprise tier for general work and a self-hosted or local route for the data that caused the block.
The four options compared, as of 15 November 2026:
1. Enterprise tier of a public tool | 2. Private workspace on approved cloud | 3. Self-hosted models | 4. Routed mix | |
|---|---|---|---|---|
What it is | The vendor's business product, with a contract and an admin console | A chat front end you run, calling models through your own cloud account | Open-weight models served on your own laptops or servers | A gateway that sends each request to an approved destination by rule |
Where prompts are processed | The vendor's cloud | Your cloud provider, in the geography or zone the deployment type allows | Hardware you control | Wherever the matching rule allows: local, your servers or an approved provider |
Who operates it | The vendor | You run the application. The cloud provider runs the model | You | You run the gateway and anything local |
Models | The models the vendor offers | The frontier models your cloud provider offers | Open-weight models that fit your hardware | Both |
What it fixes | Training on your content, no admin, no audit trail | Adds your own network path, region choice and logs | Data leaving the organisation | A different boundary for each class of work |
What it leaves open | Content still goes to the vendor. Personal accounts still exist | You own the application's security and upkeep | Model quality, hardware and operations | Rules have to be written and tested. More parts to run |
Effort to first user | Low | Medium | Medium to high | Medium to high |
Cost shape | Per user | Usage, plus the application you run | Hardware, power and staff time | The sum of the routes you allow |
Example products | ChatGPT Enterprise, Microsoft 365 Copilot, Claude Team and Enterprise, Gemini in Google Workspace | Azure OpenAI in Microsoft Foundry, Amazon Bedrock or Google's Gemini Enterprise Agent Platform, with Open WebUI or LibreChat in front | Ollama, LM Studio, vLLM, Open WebUI, Aventail™ Link | LiteLLM, Portkey (now Prisma AIRS AI Gateway), Kong AI Gateway, Aventail™ |
Every vendor fact below comes from that vendor's own documentation, read on 15 November 2026 and listed at the end.
Start with the reason for the block
A block is a decision about data. Ask the security team which of these it was, because each one points at different options.
Reason for the block | What it means | Options that answer it |
|---|---|---|
Consumer accounts may train on what staff paste, and nobody can see or delete it | The problem is the plan | 1 |
No single sign-on, audit trail or retention control | The problem is governance | 1, 2 |
Data has to stay in a named country or region | The problem is where inference runs | 2, 3, and 1 where the vendor offers residency for your region |
Data may not go to any third party: customer contract, classification, trade secret | The problem is the data leaving at all | 3, or 4 with a local-only rule |
Different work has different rules | One boundary does not fit all work | 4 |
If the answer is "all of them", write the data classes first. Our shadow AI and acceptable use policy guide has a red, amber and green scheme and a policy outline. It also explains why a block with no approved alternative does not hold: in Microsoft's UK research, 28% of staff who used unapproved AI tools said their employer provided no approved option.
Option 1: the enterprise tier of a public tool
This is the fastest fix. You buy the business version of the tool people already use and sign them in through your identity provider.
Named examples
ChatGPT Enterprise. OpenAI states that it does not train on business data by default, that workspace admins control how long data is retained, that sign-in is through SAML SSO, and that admins can reach an audit log of conversations through the Enterprise Compliance API. Enterprise has completed a SOC 2 Type 2 audit. The plan-by-plan detail is in Is ChatGPT safe for work?
Microsoft 365 Copilot. Microsoft states that prompts, responses and data reached through Microsoft Graph are not used to train foundation models, and that Copilot only surfaces content the individual user already has permission to view. For EU users, traffic stays within the EU Data Boundary. Claude models that Microsoft offers through Anthropic as a subprocessor are excluded from that boundary.
Claude Team and Enterprise. Anthropic states that by default it does not use inputs or outputs from its commercial products to train models. A conversation can be used if a member sends feedback with the thumbs up or down button, and an owner can switch that button off for the organisation.
Gemini in Google Workspace. Google states that content is not used for generative AI model training outside your domain without permission, and that existing Workspace data protection controls apply.
Choose it when the block was about consumer plans, and the data in question may go to a vendor under contract.
Check that the terms you read apply to the plan you are buying, where inference runs as well as where data is stored, which sub-processors and third-party models are switched on, and what happens to personal accounts on the same domain. An enterprise licence does not remove a member of staff's private account.
Option 2: a private workspace on approved cloud
You call models through a cloud account you already govern, and you run the chat application yourself. Prompts follow your network path, your identity provider and your logging.
Named examples
Azure OpenAI in Microsoft Foundry. Microsoft states that prompts and completions are not available to OpenAI and are not used to train foundation models without your permission. Prompts are processed within the geography you specify unless you choose a Global or DataZone deployment type, which can process them more widely. Abuse monitoring can store a sample of prompts for review unless you are approved for modified abuse monitoring.
Amazon Bedrock. AWS states that model providers have no access to Bedrock logs or to customer prompts and completions, and that you can connect privately from your own virtual private cloud with AWS PrivateLink.
Google's Gemini Enterprise Agent Platform (the documentation sits under the Vertex AI path). Google states it will not use your data to train or fine-tune models without your permission. Zero data retention is not the default: Gemini models cache customer data by default, in memory for up to 24 hours, and some features store prompts. Google lists the actions a customer must take to reach zero retention.
The front end. Open WebUI describes itself as a self-hosted AI platform that works with Ollama and OpenAI-compatible APIs. LibreChat is a chat application you install yourself and connect to a model API key. Either gives staff a chat window in front of the cloud endpoint.
Choose it when you need to control region, network and logs, the cloud provider is already an approved processor, and you have a team to run an internal application.
Check the deployment type before you approve an endpoint, because the service name does not tell you where a prompt is processed. Review the front end as you would any internal application: sign-on, conversation storage, plugins and patching. Deployment patterns are in deploying a private AI workspace across laptops and servers.
Option 3: self-hosted models
Open-weight models run on hardware you control, so the prompt does not go to a third party. This is the only option that answers "the data may not leave".
Named examples
Ollama. Runs models on your computer. Ollama also offers cloud models, so a self-hosted deployment has to make sure the model in use is a local one.
LM Studio. A desktop application for macOS, Windows and Linux that downloads and runs models locally, with a local API.
vLLM. A library for model inference and serving, built for throughput, with an OpenAI-compatible API server.
Open WebUI. A self-hosted chat interface that its documentation says is built to run entirely offline.
Aventail Link. An inference engine and gateway that runs on each machine, on macOS (Apple silicon), Windows x64 and Linux x64. Aventail™ Control manages devices, models and configuration centrally, which is the part a team needs when there are more than a few machines. Link is not open source today.
Choose it when the data that caused the block cannot go to any outside processor, or when the work has to continue without a connection.
Check three things. Test the model on your own tasks, because answer quality differs by model and task. Size the hardware per machine. Give every serving endpoint its own access control before other machines can reach it. The fuller comparison is in our guide to self-hosted AI, and disconnected deployments are covered in air-gapped AI.
Option 4: a routed mix
Most organisations have more than one kind of work, so they end up with more than one of the options above. A routing layer makes that deliberate: each request goes to a destination that a written rule allows.
Named examples
LiteLLM AI Gateway. A self-hosted server that gives applications one OpenAI-compatible endpoint for more than 100 model providers.
Portkey, now Prisma AIRS AI Gateway. A gateway with fallbacks, guardrails and conditional routing, which can also route to privately hosted or local models.
Kong AI Gateway. Routing and load balancing across AI providers, with plugins that include redacting personal data before it reaches the provider. It is managed through Kong Konnect, with data plane nodes in your environment.
Aventail. Rules are written in Aventail™ Control and evaluated by Link on the device that made the request. A rule's destination is the device itself, another approved machine running Link, or an approved provider through your own account, with up to three fallback hops. If no permitted destination exists, the request fails with an error.
Where each fits. A server-side gateway fits when your models are already served and traffic already passes through a central proxy. Aventail™ fits when the mix includes models that run on staff machines and the rule has to be applied before the request leaves the device.
Check what happens when the preferred destination is down. A fallback that sends restricted work to a cloud provider without anyone noticing defeats the purpose. The rule has to state the allowed destinations and the failure condition, and you have to test both. The patterns are in LLM gateway vs inference engine and how to route between local models and approved cloud APIs.
Evaluation checklist
Score each shortlisted product against the same twelve criteria. Ask for the evidence in the last column in writing.
# | Criterion | The question | Evidence to ask for |
|---|---|---|---|
1 | Training use | Is our content used to train or improve any model, including through feedback buttons? | The contract clause and the admin setting |
2 | Retention | How long are prompts, files and logs kept, and can we set or shorten it? | Retention settings and the deletion period |
3 | Processing location | Where does inference run, as distinct from where data is stored? | Region or deployment type per model |
4 | Sub-processors | Which third parties and third-party models can receive a prompt? | Sub-processor list and the switch for each |
5 | Identity | Does it support SSO, automatic provisioning and removal of leavers? | SSO and SCIM documentation |
6 | Audit trail | Can we see which user or device sent a request, to which model, and export it? | A sample log export |
7 | Admin controls | Can we turn connectors, sharing, plugins and agents on or off by role? | Admin console walkthrough |
8 | Failure behaviour | When the approved route is unavailable, does the request fail or go elsewhere? | A test with the approved route stopped |
9 | Coverage | Which apps, browsers and devices does the control govern, and what sits outside it? | A written scope statement |
10 | Quality | Does the model do our actual tasks to an acceptable standard? | A pilot on ten real tasks with synthetic data |
11 | Total cost | What are seats, usage, hardware and staff time over a year? | A costed pilot |
12 | Assurance and exit | Which audits does the supplier hold, and can we change model or provider without rebuilding? | SOC 2 Type II or ISO 27001 report, DPA, export options |
Criteria 3, 8 and 9 are the easiest to skip and the first the security team will ask about.
Run a two-week pilot before you buy
Pick one team and one workflow that the block stopped.
Write its data class and its allowed destinations in one sentence.
Set up the shortlisted option with synthetic data only.
Send a harmless request and confirm where it was processed from the product's own logs.
Stop the approved route and send the request again. Record whether it failed or went somewhere else.
Give ten real tasks, with synthetic data, to the people who do the work. Record quality and time.
Take the evidence to the security team with the checklist filled in.
Where Aventail™ fits
Aventail™ belongs in options 3 and 4. It suits an organisation that wants staff to keep the AI apps they use while the security team decides where each request may run.
Aventail™ Link runs on each machine as an inference engine and gateway. Aventail™ Control holds the routing rules and sends them to devices ahead of time, and Link evaluates them on the device. Aventail™ Connect configures supported apps (Claude Desktop, Claude Code, OpenCode and Codex) to send their model requests through Link, with no endpoint or key to enter in the app. External providers are used through your own account. The hosted Control service receives management, device and routing events. Prompts and responses do not pass through it.
In practice that gives two lanes:
General work: the enterprise tier you chose in option 1.
Restricted work: an assistant routed by Aventail™ under a policy that permits only On device. If the local model cannot serve the request, it fails with an error and is not sent to a disallowed destination.
Three limits apply. Aventail™ governs model requests routed through Link. It does not discover or block tools used outside it, so the network block stays in place for services you have rejected. Routing to other machines and hosted providers is in beta. Routing policies need Control for Enterprise. The Free plan covers three connected devices, including business use. Aventail™ holds SOC 2 Type II and Cyber Essentials.
Frequently asked questions
Why do security teams block public AI tools?
Because consumer plans give the organisation no contract, no admin view and no record of what staff entered, and some train on that content by default. A block stops the exposure while an approved route is chosen.
Does blocking ChatGPT stop staff using it?
On company devices and networks, mostly. It does not reach personal phones or home laptops. In Microsoft's UK research from October 2025, 71% of employees had used unapproved consumer AI tools at work.
Is ChatGPT Enterprise enough on its own?
It answers the training, sign-on, retention and audit concerns. It does not answer a rule that certain data may not go to any third party, because the prompt is still processed in OpenAI's cloud.
What is the difference between a private AI workspace and self-hosted AI?
A private workspace on approved cloud runs your own application but calls a model that a cloud provider hosts. Self-hosted AI also runs the model on hardware you control, so the prompt does not leave.
Does Azure OpenAI or Amazon Bedrock keep our data in the UK?
It depends on the deployment. Microsoft processes prompts within the geography you specify unless you pick a Global or DataZone deployment type. Check the region and the deployment type for each model before you approve it. Our AI data residency guide covers the checks.
Which option is cheapest?
It depends on use. An enterprise tier is a fixed cost per user. Cloud endpoints are billed by usage. Self-hosted models have no per-request fee but need hardware, power and staff time.
What should developers use if public AI coding tools are blocked?
The same four options apply. For coding agents on local models, see local AI for a development team.
Related guides
Sources
All read on 15 November 2026.
OpenAI, Enterprise privacy (updated 8 January 2026): https://openai.com/enterprise-privacy/
Microsoft Learn, Data, privacy, and security for Microsoft Copilot: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy
Anthropic Privacy Center, Is my data used for model training? (commercial products, 18 August 2026): https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training
Google Workspace Help, Generative AI in Google Workspace Privacy Hub (updated 6 October 2026): https://support.google.com/a/answer/15706919
Microsoft Learn, Data, privacy, and security for Foundry Models sold by Azure in Microsoft Foundry: https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/data-privacy
AWS, Amazon Bedrock data protection: https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html and Amazon Bedrock security and privacy: https://aws.amazon.com/bedrock/security-compliance/
Google Cloud, Gemini Enterprise Agent Platform and zero data retention (updated 12 November 2026): https://cloud.google.com/vertex-ai/generative-ai/docs/data-governance
Open WebUI documentation: https://docs.openwebui.com/
LibreChat documentation: https://www.librechat.ai/docs
Ollama documentation: https://docs.ollama.com/
LM Studio documentation: https://lmstudio.ai/docs/app
vLLM documentation: https://docs.vllm.ai/en/latest/
LiteLLM, AI Gateway: https://docs.litellm.ai/docs/simple_proxy
Portkey (Prisma AIRS AI Gateway), AI Gateway: https://portkey.ai/docs/product/ai-gateway
Kong, AI Gateway: https://developer.konghq.com/ai-gateway/
Microsoft UK Stories, Rise in 'Shadow AI' tools raising security concerns for UK organisations (13 October 2025): https://ukstories.microsoft.com/features/rise-in-shadow-ai-tools-raising-security-concerns-for-uk
Aventail™, Connect overview: https://docs.aventail.co.uk/docs/connect/overview
Aventail™, Routing: https://docs.aventail.co.uk/docs/fleet/routing
Aventail™, Compatibility and versions: https://docs.aventail.co.uk/docs/introduction/versions-and-supported-models
Aventail™, Security and access: https://docs.aventail.co.uk/docs/guides/security
Aventail™, How it works, Security and trust, and plans: https://aventail.co.uk/how-it-works, https://aventail.co.uk/trust and https://aventail.co.uk/#pricing