The Enterprise AI Data Perimeter

Most enterprise AI agreements make one headline promise: the provider will not train its models on your prompts and uploads. That promise matters, but it covers a small part of the real exposure. In a recent essay, The Enterprise AI Data Perimeter, Jentic CEO Sean Blanchfield sets out where company data actually leaks once AI is in daily use, and how to decide which workloads can safely go where. This post summarises the argument.

Personal accounts sit outside every contract

No-training terms apply to enterprise accounts. They do not apply to the personal accounts that employees use at work, and most organisations have a lot of that. Microsoft and LinkedIn found that 78% of surveyed AI users bring their own AI tools to work.

When customer records, contracts or internal strategy go into a personal account, the employer has no visibility, no deletion rights and no say over whether that material ends up in a future training run. Sean's point is that the transcripts themselves are the asset. They are a step-by-step record of how your people do their jobs, and that know-how can end up improving a model that is then sold back to you and your competitors.

Blocking does not work

Banning personal AI tools on work devices only moves the activity to personal phones. Sean's view is that the only reliable fix is a sanctioned platform that is better to use than the consumer alternative, because it is connected to the data, tools and APIs people need to get their work done. A generic internal chat interface will not win on its own. One that is the only route to the systems people rely on will.

Every vendor is now an AI vendor

AI is being built into CRM, HR, email, analytics, security and development tools. Each of those vendors, along with plug-ins, subcontractors and MCP servers, now receives data that has new commercial value. No-training protections need to cover all of them, not only the large model providers.

Metadata and indemnities

A no-training clause still allows providers to keep logs, telemetry and usage data for security and operations. That derived data shows which tools, models and workflows you use and which tasks succeed. In other words, it maps the business processes you are automating, which are fast becoming a company's most valuable intellectual property.

Indemnities are not a full answer either. Standard terms are rarely negotiable, liability is capped, and no cap reflects the loss if your operating know-how ends up in a public model's weights.

Trust, but verify

Customers have no way to check what happens to their data after it leaves the building. Running proprietary models through a hyperscaler adds some separation, although the hyperscalers are closely tied to the AI labs commercially. Open-weight models on independent hosting providers remove the training incentive altogether. Separately, no contract prevents breaches, insider access or misconfiguration, and for organisations outside the US, the CLOUD Act means EU data residency does not stop a US provider being compelled to hand data over.

Route by sensitivity first

Most model routing today optimises for cost, speed and quality. Sean argues that the first question should be who you trust with which data, and only then which model offers the best price and performance inside that boundary. He sets out five options:

• Personal AI: outside the perimeter. Public information only, no company data.
• Enterprise AI lab: contractual no-training terms and admin controls. Suitable for the least sensitive company work.
• Proprietary models through a hyperscaler: for sensitive work that still needs frontier intelligence.
• Open-weight models on independent providers: for highly sensitive work, with no training motive and stronger sovereignty guarantees.
• Open-weight models on your own hardware: for the most sensitive data and mature, repeatable workflows. Data never leaves your building.

What to do now

The practical steps follow from the framework:

• Stay model-neutral. Put an AI gateway in front of all model traffic, give it a mix of proprietary and open-weight models, and block direct access to AI providers.
• Outcompete shadow AI. Give employees a sanctioned platform that is easy to use and connected to the right systems, so the company context it builds stays inside your perimeter.
• Inventory every processor. Review the training rights, retention, metadata use and liability terms of every SaaS, AI and MCP vendor that touches your data.
• Decide on jurisdiction. Assess whether foreign government access is a risk you need to plan for.

The pattern Sean recommends is to use frontier models for hard and novel problems, capture the workflows that result, then move repeatable work to open-weight models that run cheaply and securely under your control.

Read the full essay on seanblanchfield.com. If you want to talk about running agents inside your own data perimeter, get in touch with the Jentic team.

---

Based on Sean Blanchfield's essay, originally published at seanblanchfield.com on 14 September 2026.