I have started to think about what AI knows about me as a kind of spider diagram.

I sit in the middle. Around me are different spokes: the businesses I am building, the projects I am working on, the way I communicate, the subjects I keep returning to, the problems I am trying to solve, and the context that shapes how I see them.

The useful bit is the connection between them. If an AI understands which business I mean, how I like to work and what we have already decided, I can get on with the next question rather than brief it all over again.

On 14 January, Google launched Personal Intelligence in Gemini, initially as a beta for eligible US subscribers. With permission, Gemini can connect information from Gmail, Google Photos, YouTube and Search and use that context to produce more personalised answers.

That announcement caught my attention because I think it points towards a much bigger change. For the last few years, the AI industry has been obsessed with making models more intelligent. I suspect the next leap in usefulness may come from something slightly different.

AI is starting to know who it is working for.

Intelligence without context has a ceiling

Most of the discussion around frontier AI still revolves around capability. Can the model reason better? Can it write better code? Can it understand images? Can it solve harder problems?

Those things obviously matter. But imagine hiring an exceptionally intelligent person who knows almost nothing about your business. They do not know your customers. They do not know the projects already underway. They do not understand how you make decisions. They do not know which ideas have already been rejected, how you like information presented, or what happened in the meeting yesterday. They may be brilliant. You will still spend an enormous amount of time briefing them.

That is roughly how most of us have used AI. We compress enough context into a prompt to make the model useful, get an answer, and then repeat the exercise when the context changes or the conversation starts somewhere else.

Memory begins to change that relationship. OpenAI had already been moving in this direction. In April 2025, it expanded ChatGPT’s memory so that, where enabled and available, it could reference previous conversations rather than relying only on explicitly saved memories.

Google’s Personal Intelligence pushes the same underlying idea further by connecting an AI assistant to information distributed across products people already use. The interesting question is no longer simply: how intelligent is the model? It is: how much relevant context can the model bring to the problem?

Personalisation is more than remembering facts

There is an important distinction between memory and adaptation. Knowing my name is memory. Knowing that I prefer a direct answer, dislike unnecessary padding and would usually rather have three useful words than ten polite ones is more interesting.

The difference becomes noticeable in the interaction itself. Over time, the AI can begin to respond in a way that fits how I work. I do not have to keep explaining that I want the answer tightened, that I am comfortable challenging an assumption, or that I want the commercial implication rather than another paragraph describing the technology.

Having spent years selling software, I am fairly sceptical of the phrase human-centred. We have used it for decades while largely requiring humans to adapt to software. We learn the interface. We learn the workflow. We learn where the button is. We learn what information the system expects and in what order. AI creates the possibility of reversing part of that relationship. The software can begin adapting to the person.

The most useful AI may not be the one that knows the most. It may be the one that needs the least explaining.

I still want some personality in the answer. Direct does not have to mean sterile. The useful adaptation is knowing when I want the detail and when three words will do.

Context compounds

The spider-diagram analogy becomes more useful when I think about work. I am usually moving between several projects. Historically, every change of context carried a cost.

You brief somebody on one business. Then another. You explain the background to a product decision. You reopen an old document. You reconstruct why a previous decision was made. You remind people which constraints matter.

AI can reduce some of that overhead if context persists. A conversation about one project does not necessarily have to begin from zero. A preference established somewhere else can inform how information is presented. Previous decisions can become useful context for the next decision.

That does not mean an AI remembers everything perfectly. It doesn’t. It means the economics of context begin to change. The more useful context a system can retrieve at the right moment, the less time the human spends rebuilding the world before useful work can start.

For an individual, that can feel like convenience. Inside a business, it could be much more important.

The real opportunity may be organisational memory

Most organisations already possess enormous amounts of context. It is just badly distributed. Some sits in email. Some in CRM records. Some in documents. Some in Slack or Teams. Some in project systems. Some in meeting notes. A great deal sits in people’s heads.

This is one reason apparently simple business questions can be surprisingly expensive to answer.

  • Why did we price the deal this way?
  • What did the customer object to six months ago?
  • Why was this feature deprioritised?
  • Which assumption changed?
  • Who has worked on something similar before?

The answer may exist, but finding and reconstructing it takes time. If AI systems become better at working across organisational information, then proprietary context could become an increasingly important business asset.

The frontier model may be shared by millions of customers. Your organisation’s context is not.

That suggests an interesting competitive dynamic. Two businesses may have access to exactly the same underlying model and still get very different value from it because one has made its knowledge accessible, structured and governable while the other has not. The advantage may not be owning better AI. It may be giving AI a better understanding of the organisation it is working inside.

Knowing more also creates new boundaries

There is an obvious problem with all of this. The more useful AI becomes through context, the more consequential the question becomes: which context should it have?

My instinct with my own information is relatively permissive. If sharing something about myself makes the system more useful to me, I can make that choice.

Information about other people is different. That distinction matters.

Personalisation can easily move from helpful to intrusive if systems infer things people did not knowingly provide, combine information in ways they did not expect, or surface context that belongs to somebody else.

Google appears conscious of this tension. Personal Intelligence is opt-in, users choose which apps to connect, and Google explicitly acknowledges that the system can make incorrect connections or “over-personalise”.

That last problem is particularly interesting. Data is not the same as understanding.

A system might see hundreds of photographs of someone at football matches and infer that football is their great passion. The reality might be that their child plays football and they are there because being a parent matters to them.

The observations can be accurate while the interpretation is wrong. That is a useful warning for businesses too. More context does not eliminate the need for judgement.

Enterprise AI will make this harder

Consumer AI can often frame personalisation as an individual choice. Enterprise environments are more complicated.

I have spent enough time around large organisations to know that technology adoption is rarely determined by capability alone. There are data policies, regulatory requirements, security teams, employee consultation, access controls, procurement processes and legitimate questions about what an employer should be able to infer about the people who work there.

Those constraints are not an argument against contextual AI. They are part of implementing it properly.

The companies that move fastest will not necessarily be those with the fewest rules. They may be the ones that become best at defining the boundaries.

  • Which information can an AI access?
  • Who authorised that access?
  • What is personal and what belongs to the organisation?
  • Which sources can be combined?
  • What should never be inferred?
  • When does a human need to verify the answer?

Those questions sound like governance. They are also product-design questions. Because if the system cannot access enough context to be useful, employees will ignore it. If it accesses context in ways people do not trust, they will resist it. The value sits somewhere between the two.

Context may become part of the moat

Frontier AI models will continue improving. Some will be better than others. The rankings will move. Capabilities that look extraordinary today will become standard features tomorrow.

That makes me wonder whether we are overestimating the long-term value of model access and underestimating the value of context. If every serious business can access powerful intelligence, what differentiates the output?

Part of the answer will be people: judgement, ambition, creativity and the ability to identify worthwhile problems. But another part may be the context surrounding the intelligence.

  • Your customers.
  • Your decisions.
  • Your processes.
  • Your history.
  • Your proprietary knowledge.
  • The relationships between all of them.

Businesses have spent years accumulating data and talking about becoming “data-driven”. Much of that conversation focused on dashboards, analytics and structured databases. AI potentially makes a much larger portion of organisational knowledge usable. Not automatically. Not safely by default. And certainly not perfectly. But usable in ways that were previously too expensive or cumbersome to attempt.

The next AI question is who it understands

Google calls its new feature Personal Intelligence. The name is revealing.

The first phase of generative AI was largely about giving ordinary people access to increasingly powerful general intelligence. The next phase may be about combining that general capability with specific context.

For a business, I would start with two questions: What would a capable AI need to understand about us to become more useful? And what should it never be allowed to know?

Those two questions belong together. Because the opportunity is not to build an AI that indiscriminately knows everything. It is to build systems that can access the right context, at the right moment, with the right permissions, and turn it into something useful.

The model still matters. But perhaps the most interesting intelligence will come from the space between what the model knows about the world and what it understands about you.