When harmless data becomes almost mind-reading
Most of us started using AI in much the same way: Ask a question → Get an answer.
Then many of us went further: better instructions, more context, perhaps a few documents: Information + Instructions → Deliverable.
A report. A presentation. An analysis. A plan.
But there is another way of using AI that I find much more interesting: Scattered Information → AI → Inference.
Instead of asking AI to produce something from the information we give it, we can ask: What do these pieces of information tell us when we connect them together?
And that can produce some surprisingly powerful results.
AI can help us see what nobody actually said
Imagine two colleagues strongly disagreeing about a decision. Give AI both sides of the conversation and ask: What might each person be seeing that the other is missing? What concerns might be driving their positions?
Or imagine preparing for an important customer conversation: Give AI the previous discussions, what you are hoping to achieve, and relevant public information about the customer’s organisation. Ask it to identify possible priorities, concerns, conversation entry points and likely objections. The same applies to organisational change: Finance, Engineering, Sales and Operations might hear the same announcement very differently. AI can help anticipate those perspectives before we communicate.
None of this is mind-reading. It’s inference – something humans have always done by listening, observing and connecting information. What is changing is how easily AI can do this across large amounts of information, including unstructured information that previously took considerable human effort to connect.
Increasingly, anyone with access to a general-purpose AI tool can simply ask: “What do you think is really happening here?” And this is where something incredibly useful starts to become something we need to think about carefully.
When harmless data stops feeling harmless
Think about the information we leave behind every day. A public social media post. A few searches. The articles we read. The questions we ask. The words we choose. Even when and how we respond. Individually, many of these signals may reveal very little. But AI does not have to look at them individually. It can connect them.
From apparently ordinary information, it may be possible to infer interests, concerns, preferences, circumstances or likely behaviour that someone never explicitly disclosed. Add other digital signals – cookies, tracking pixels or browser fingerprinting, for example – and the picture can become richer still. AI cannot literally read our minds.
But perhaps it doesn’t need to. If AI can connect enough seemingly harmless fragments to infer, with reasonable probability, what someone wants, worries about, or may do next, an organisation can potentially act on something it believes it knows about a person, even though that person never explicitly disclosed it.
And that is where the real impact begins. The important question is no longer simply what AI can infer. It’s what happens when someone acts on that inference.
The same inference. Completely different outcomes.
Imagine an organisation uses AI to analyse a range of ordinary signals and concludes: This customer may be under increasing financial pressure. The customer never said that. The AI inferred it.
What should happen next?
Perhaps the organisation uses that insight to recognise that the customer may need additional support. Perhaps it avoids offering them an unsuitable product. Those could be genuinely beneficial uses of AI. But the same inference could be used differently.
Perhaps the organisation decides that this person is now a less attractive customer. Or perhaps it discovers that financial pressure makes a particular message, offer or sense of urgency more likely to influence their decision.
And there is another possibility: AI inference is probabilistic, and it can be wrong. A collection of signals that appears to tell one story may have a completely different explanation. That doesn’t mean we should stop using inference. The governance question is how we use it, what problem we are trying to solve, and what happens if the inference is wrong.
The value and the risk come not only from what AI can infer, but from what we choose to do with it.
This isn’t a regulatory vacuum
Inference itself is not a new concept in privacy. Profiling and inferred information are already established concerns under data protection law. Existing principles around fairness, transparency, purpose limitation, data minimisation and automated decision-making remain highly relevant.
The EU AI Act adds another layer for particular AI practices and higher-risk uses. So the interesting question is not whether somebody suddenly discovered a completely unregulated activity called “AI inference”. What AI changes is how easily sophisticated inference can now be performed, by far more people.
From everyday AI use to enterprise scale
An employee using an ordinary general-purpose AI tool can already combine a conversation, internal context and public information and ask what it might reveal about a customer. AI policies therefore need to consider not only what information goes into AI, but what new information may be inferred from it, and how that inference is used.
At enterprise scale, the same capability can be built into products, customer journeys and business processes. There is enormous potential value here. Better understanding can improve customer service, identify genuine needs and support better decisions. So the answer cannot simply be: don’t use it.
The challenge is to be deliberate about why we are making the inference, how confident we are that it is right, and what we allow to happen as a result. For larger investments, organisations should also ask whether customers would reasonably expect their information to be used this way and whether the capability will remain sustainable as regulation, customer expectations and the operating environment change.
A capability can work exactly as designed and still become a poor investment if its use later becomes difficult to justify, control or adapt. That is where digital trust becomes practical. AI is becoming remarkably good at learning things people never explicitly told us. What can AI learn about someone without them ever telling us and what should we do with that knowledge?
Browse our website for more information on Resillion’s AI services.
