AI Risk Controls
AI Tool Use, with the Risks Controlled
The short answer
AI is a tool, not a tether. We use it deliberately, for defined jobs, and we control the risks it brings: where your data goes, who approves what it produces, and what happens when a provider changes its terms or fails. Nothing we build needs a permanent connection to an AI provider to keep working.
The Problem with Permanent Connection
Many AI products are built so that their value only exists while you stay connected. Every visitor question, every document, and every answer passes through someone else's model, on someone else's terms, often priced per message.
When the provider changes its terms, its prices, or its model, your service changes with it. When it has an outage, so do you. For an organization that handles personal or sensitive information, it also means that data leaves the building every time someone types.
That is dependency, not capability.
The Risks, and How Each One Is Controlled
Each row below can be recorded in a risk register as a risk with its control.
| Risk | What can go wrong | How we control it |
|---|---|---|
| Data leaving your control | Visitor questions, documents, and content are processed on a provider's systems, outside your jurisdiction | Retrieval-only answers and an air gap by default; data held in the jurisdiction you choose |
| Personal data exposure | Personal details typed by visitors are stored, or sent on to a third party | Common identifiers removed before anything is stored; retention period you set; nothing sent to an AI provider by default |
| Inaccurate or unapproved answers | An AI-written answer misstates fees, eligibility, or compliance status | Answers come only from approved content; regulated statements given word for word from the Verified Answer Bank; AI output stays a draft until a person approves it |
| Provider dependency | Your service changes or stops when a provider changes its terms, prices, or models | Benefits kept in structures you own; answer systems work with no AI connected |
| Cost escalation | Per-message pricing rises as use grows | AI used for defined jobs, on your own provider account, with a key you can remove |
| Service outage | A provider's downtime takes your service down with it | Answering runs on your own infrastructure and does not depend on a provider |
| No audit trail | Nobody can show where an answer came from, or who approved it | Every answer traces to a source page or an approved answer, with a named approver |
Two Ways to Use AI
| Compared on | AI as a permanent connection | AI as a tool |
|---|---|---|
| Visitor questions | Pass through the provider's model, every time | Answered from your own content; nothing sent by default |
| Cost | Often metered per message, rising with use | Predictable: AI is used for defined jobs |
| Provider changes terms or prices | Your service changes with them | Nothing you rely on stops working |
| Provider outage | Your service goes down too | Answers keep working |
| Who approves outputs | Often no one, before visitors see them | A person, before anything is applied |
| If you disconnect | Little of the value remains | Content, approved answers, answer system, structured data, and measurement all remain |
Four Principles for AI Tool Use
Draft, Not Decide
AI suggests answers, wording, and vocabulary. Every suggestion arrives as a draft for a person to approve, and nothing is applied automatically.
Use Once, Keep the Benefit
We use AI to do the heavy lifting on your content, then keep the results in structures you own: approved answers, vocabulary, and structured data. The benefit stays after the AI is gone.
Air Gap by Default
Nothing is connected unless you choose. When AI is useful, a person copies a prompt into the model of their choosing and pastes the result back. See Air Gap by Default.
Connect Only When It Earns Its Place
If a live connection makes sense for a task, you connect your own provider account, do the work, and remove the key afterward.
AI Agents Now Come with Ordinary Hosting
Not long ago, running an AI agent meant renting a separate server and building the plumbing yourself. Today it can arrive as a standard feature of a basic web hosting account, sitting alongside the file manager and the email settings.
Eco CPanel Hosting is one example. Its plans include an AI coding agent in the file manager that can read a site's files and server logs and write code, and its largest plan adds a persistent assistant for scheduled site health checks and recurring tasks. See what ecocpanel.com includes.
Nothing is switched on for you. The agent only works once the account holder connects their own ChatGPT or Claude account, an approach known as bring your own key. You choose the provider, you hold the account and its terms, and you can disconnect it whenever you like. That is the tool-not-tether principle built into the product.
| Tool | What it does | Plans | How it is switched on |
|---|---|---|---|
| AI coding agent | Works in the file manager: reads site files and server logs, runs commands, and writes code | All plans | You connect your own ChatGPT or Claude account |
| Persistent assistant | Runs scheduled site health checks, summaries, and recurring tasks in the background | Largest plan | You connect your own ChatGPT or Claude account |
One point to weigh before connecting: what the agent sends to your AI provider is processed under that provider's terms and wherever it operates. Your hosting location decides where your files live. It does not decide where your agent's requests are processed.
What Keeps Working When the AI Is Switched Off
- Your content, restructured so people and machines read it the same way.
- Your Verified Answer Bank, given word for word.
- Your private answer system, answering from your own content in retrieval-only mode.
- Your structured data and authorship, which answer engines keep reading.
- Your measurement, showing which visits and enquiries answer engines send you.
The Test Applied to Our Own Products
Every product we deploy passes the same test: switch the AI off, and it keeps working. See what we build.
| Product | What it does | Data held | AI provider |
|---|---|---|---|
| My Chat Assistant | Our self-hosted answer engine, answering only from content you have approved. | Your own server, in the jurisdiction you choose | Not needed. Optional help drafting suggested answers, using your own key; nothing is used until a person approves it |
| CQIP Static | Fast static pages with answer-first structure and structured data built in. | No database; pages hold only what you publish | None |
| CQIP Trust Declaration | States who wrote and reviewed each page, so answer engines can verify it. | Your own website | None |
| CQI Referrer Attribution | Shows which visits and enquiries come from AI assistants. | Your own website's database | None |
| Semantic Journey | Shows which pages are read, and which lead people to make contact. | Your own website or server | None |
| Eco CPanel Hosting | Hosting in a named jurisdiction, for sites and answer systems. | The jurisdiction you choose | Not needed. Optional AI agent tools stay off until you connect your own key |
Why This Matters at Board Level
Costs stay predictable, because value does not depend on a per-message meter. Data stays in the jurisdiction you choose. Every answer can be traced to an approved source, which is what auditors and regulators ask for. And no single provider's decision can switch your service off.
The organizations that gain most from AI will be the ones that can switch it off and keep going. That is the standard we build to.
See it in practice
Built and Run by Us
My Chat Assistant
Our self-hosted answer engine is built this way: air gap by default, retrieval-only answers, and AI only when you choose to connect it.
Questions
Frequently Asked Questions
How do these controls fit our risk register?
Each risk in the table can be recorded with its control. For each deployment we provide a written description of how every control is configured, for your risk and data protection assessments.
Are you against using AI?
No. We use AI where it does a job well, such as drafting answers and suggesting vocabulary. We object to making your service depend on a permanent connection to it.
Do your answer systems work with no AI connected at all?
Yes. They run in retrieval-only mode by default, answering directly from your approved content.
Can we use our preferred AI provider?
Yes. When a connection is useful, you use your own provider account, on your terms, and you can remove the key when the task is done.
What does bring your own key mean?
You connect an AI tool to your own account with a provider such as ChatGPT or Claude, rather than a vendor's shared account. You choose the provider, you control the account, and you can remove the connection at any time.
What happens if an AI provider changes its terms or prices?
Nothing we deliver stops working. The benefits live in structures you own, not in a provider's service.
Keep reading
Related
Start with a Confidential Chat
A no-obligation conversation with H D Fraser about how answer engines represent your organization today, and what a privacy-first approach would involve. NDA on request.