The Big AI Secret · ongoing research

AI can find customers. Their data stays outside.

Call extracts reveal different ways businesses draw the line around client data, from excluding it altogether to restricting where it can go. Those approaches answer different questions, and a general promise to use AI carefully does not explain which boundary applies.

5 November 2026 · 4 minute read

The agency describes an agent that helps identify potential customers and populate a spreadsheet. Yet material arriving from existing customers is kept outside AI altogether. That is a specific boundary: AI can support finding business without being allowed to read the information that customers subsequently provide.

“So Nothing my customers will give me or send me goes into any AI tool. So there's a clear firewall between anything that just sent to me by my customers and then anything that goes into AI.”Agency and marketing, under 25 staff

The accompanying call notes say this boundary appears in explicit terms of service and ways of working documents. The word firewall describes the speaker's rule. It does not establish that a technical barrier prevents an upload. What the evidence does establish is the intended scope: nothing supplied by customers should enter AI.

The source of the data can decide

This is broader than removing names or contact details. The rule turns on where the material came from, not whether someone judges each item sensitive. A customer document could contain no personal details and still fall outside the permitted boundary. That makes the stated decision relatively clear, although these extracts do not show how consistently it is followed.

Another agency describes a more selective division. Its call notes distinguish shared methods, which AI may use, from information relevant to an individual client, which it may not extract.

“What he's not allowed to do is go into individual clients' perspectives and start extracting data which is relevant to an individual client.”Agency and marketing, 100+ staff

Here, the distinction is between a business's reusable knowledge and a client's particular information. AI may work with how the agency approaches a task without receiving everything it knows about the client. This is not the same rule as excluding everything a customer sends.

The practical question is whether those categories can be separated in the material staff actually use. A working document might combine a shared method with client-specific detail. The extract does not tell us how the agency handles that case. It identifies a boundary worth testing, rather than proving that the boundary holds.

Keeping data contained is a different claim

A technology business describes its rule in terms of the environment where data is processed. Its concern is keeping information away from wider model learning.

“We have one rule with the AI structure is that all of the data is kept in a closed loop rather than exposed to the wider learning of the models.”Technology and software, under 25 staff

That statement addresses containment, not exclusion. It describes where data should remain and what it should not feed. It does not, by itself, say which client records staff are authorised to enter, who can retrieve them, or whether a client has agreed to that use.

The call extract is not a technical inspection. We cannot verify the closed-loop claim or infer contractual protections from it. Nor does the contrast establish that exclusion is always preferable. It shows why accounts of safe AI use need to name the actual restriction. Keeping information within a chosen environment and refusing to process it are materially different decisions.

The client also sets the boundary

A further agency account brings the client's terms into the discussion before use. Staff can test ideas, according to the accompanying notes, but must discuss whether the client is comfortable and what its terms allow.

“It always has to be at least entered into the discussion, how it's being used and whether the first and foremost, is the client comfortable? What are the client's terms?”Agency and marketing, 25 to 100 staff

This adds a permission question that a storage rule cannot settle. A business may consider an environment suitably controlled while a client's terms still restrict the proposed use. Equally, a client discussion does not establish that the chosen system handles data as intended. These checks complement each other rather than standing in for each other.

For a director, the useful test is to take an ordinary client document and ask what happens next. Is it excluded because the client supplied it? Can only the general method be used? Is processing permitted within a specified environment? Do the client's terms allow that task? A clear answer should identify the applicable rule, not merely repeat that staff are careful.

These are qualitative call extracts supplied to The Big AI Secret, not evidence of how common each approach is or which prevents more incidents. Speaker roles were not supplied and are marked accordingly. The finding is narrower: businesses describe distinct data boundaries, and those distinctions matter before any client material reaches AI. Reviewing the output comes too late to decide whether the input belonged there.