AI knows an extraordinary amount.
Ask a frontier model about accounting principles, manufacturing processes, sales methodologies, software development, marketing strategy or supply chains and it can often produce an impressively useful response.
It can analyse information, reason through problems, conduct research, write, compare and summarise. Increasingly, it can use tools and take actions.
But ask it why one of your largest customers receives an unusual discount and it probably doesn’t know.
It doesn’t know why your operations director insists on checking a particular order manually.
It doesn’t know which apparently small customer your CEO considers strategically important.
It doesn’t know why an exception was added to a process six years ago.
It doesn’t know that one of your internal documents is technically current but everyone experienced enough to know better ignores it.
It doesn’t know what your best salesperson notices in the first ten minutes of a conversation.
It doesn’t know why you tried something three years ago and decided never to do it again.
It doesn’t know the peculiarities, history, relationships, exceptions and accumulated experience that make your organisation your organisation.
This creates an interesting paradox.
AI is becoming extraordinarily knowledgeable about the world while remaining remarkably ignorant about the business asking it for help.
And as the underlying models become more capable, solving that second problem may become increasingly important.
General intelligence isn’t organisation-specific intelligence#
Last month I wrote about the distance between AI capability and AI value.
The basic argument was that technological capability doesn’t automatically become organisational capability.
Context is one reason why.
Imagine asking a highly intelligent external consultant to walk into your business on Monday morning.
They might know your industry exceptionally well. They might understand finance, operations, strategy and technology. They might have worked with hundreds of similar companies.
But they still wouldn’t know everything required to make good decisions about yours.
They would ask questions.
How does this process actually work?
Why do you do it that way?
Which customers matter most?
Where does the margin come from?
What happened last time you changed this?
Who needs to be involved?
Which data should I trust?
What does this field actually mean?
Is this policy still followed?
What’s the exception?
They would need context.
AI has the same problem, except at an extraordinary scale.
A frontier model can arrive with a remarkable amount of general knowledge.
But general knowledge and organisational knowledge are different things.
One can increasingly be bought through an API or accessed through a subscription.
The other has accumulated inside your business over years.
The same model can produce very different value#
Imagine two competing companies use exactly the same frontier AI model.
Same model. Same underlying capabilities. Same access to general intelligence.
Company A gives employees access to it and encourages them to experiment.
People ask questions, produce documents, analyse information and save some time.
Useful things happen.
Company B does the same, but also begins connecting the model carefully to the information that makes its business work.
Product information.
Customer history.
Commercial policies.
Previous proposals.
Service records.
Operational procedures.
Contracts.
Meeting notes.
Internal research.
Performance data.
Previous decisions.
Relevant communications.
The accumulated knowledge of experienced employees.
Now imagine an employee in each company asks:
How should we approach this customer?
The underlying model might be identical.
The answer shouldn’t be.
In Company A, the model has to infer much of its response from general knowledge and whatever information the employee provides manually.
In Company B, it may be able to understand the customer’s history, previous interactions, products owned, outstanding issues, commercial importance, relevant policies and similar situations the organisation has encountered before.
Same artificial intelligence.
Very different organisational capability.
The model provides the intelligence. The organisation provides the context that makes the intelligence useful.
This is why connecting AI to information matters#
The technology industry is already moving rapidly in this direction.
OpenAI’s Responses API includes file search, allowing developers to retrieve relevant information from collections of documents and bring it into the model’s context.
In May, OpenAI expanded those capabilities and added support for remote Model Context Protocol servers, creating more ways for models to connect with external tools and information.
Anthropic has been moving in a similar direction.
It introduced the Model Context Protocol in late 2024 as an open standard for connecting AI applications to tools and data. In May, Anthropic launched Integrations that allow Claude to connect to remote services, while its expanded Research capability could search across the web, Google Workspace and connected applications.
The technical approaches will continue to evolve.
But the direction is more important than any particular implementation.
The frontier isn’t only about making the model more intelligent.
It’s also about connecting that intelligence to the information and tools required to make it useful in a particular environment.
For businesses, that’s a crucial distinction.
The model may understand the world.
The implementation has to help it understand the situation.
More data isn’t the same as better context#
There is an obvious response to this problem.
Give AI everything.
Every document. Every email. Every customer record. Every meeting transcript. Every policy. Every spreadsheet. Every report. Every message.
Then surely it will understand the business.
I’m not convinced it’s that simple.
Businesses contain enormous amounts of information.
They also contain enormous amounts of rubbish.
Outdated documents.
Duplicated documents.
Abandoned plans.
Incorrect CRM records.
Old pricing.
Contradictory policies.
Half-finished presentations.
Unlabelled spreadsheets.
Meeting notes without decisions.
Reports nobody trusts.
Documents that describe how a process is supposed to work rather than how it actually works.
The problem isn’t merely accessing information.
It’s identifying which information should matter in a particular situation.
Suppose an AI system is helping an account manager prepare for a customer renewal.
Does it need every document the company has ever produced?
Obviously not.
It needs the information relevant to that customer, that relationship and that decision.
Contract history.
Usage.
Service issues.
Previous conversations.
Commercial terms.
Relevant products.
Potential risks.
Perhaps comparable renewals.
Perhaps current company strategy.
Perhaps knowledge held by the account manager that exists nowhere else.
Context isn’t simply a larger pile of data.
Context is the right information made available at the right moment for the right problem.
That is a much harder problem.
Some of your most valuable knowledge isn’t written down#
There is another complication.
Not everything an organisation knows exists in a database.
Think about an employee who has worked in a business for 15 years.
They know which customers require careful handling.
They know which supplier always promises a date it can’t meet.
They know that a particular number in a report is frequently misleading.
They know which product combinations cause problems.
They know who to call when something unusual happens.
They know why a process contains an apparently unnecessary step.
They know the difference between an unusual situation and a genuinely concerning one.
Ask them how they know and the answer may simply be:
Experience.
This is tacit knowledge.
It exists in people’s heads, habits, relationships and judgement rather than neatly structured corporate systems.
Businesses have always had this problem.
People leave.
Knowledge disappears.
New employees take years to acquire experience.
Important decisions depend disproportionately on a relatively small number of people who understand how things really work.
AI makes the problem newly interesting because there is potentially much greater economic value in making organisational knowledge accessible.
If an AI system can help an employee access not only information but some of the accumulated reasoning, patterns and lessons of experienced colleagues, expertise may travel further through the organisation.
But first, the organisation has to recognise that knowledge exists.
And somehow capture enough of it to be useful.
The best data may be the decisions you’ve already made#
There is one category of organisational knowledge I think could become particularly valuable.
Previous decisions.
Businesses make thousands of decisions and preserve surprisingly little about why they made them.
A customer received a particular commercial offer.
Why?
A supplier was rejected.
Why?
A product feature was prioritised.
Why?
An employee escalated a problem.
Why?
A forecast was adjusted.
Why?
An investment was approved.
Why?
A customer exception was granted.
Why?
Months later, the outcome may remain visible while the reasoning disappears.
Imagine if AI systems could interrogate not only what an organisation did, but the context around why it did it.
What information was available?
What assumptions were made?
What alternatives were considered?
What happened afterwards?
Was the decision successful?
Would we make the same decision again?
That begins to turn organisational history into something more useful than an archive.
It becomes material that can inform future judgement.
This doesn’t mean an AI system should blindly reproduce previous decisions.
Past decisions can be wrong.
Circumstances change.
Old assumptions become obsolete.
But organisational memory can still provide context that general intelligence doesn’t possess.
The model might know how businesses usually solve a problem.
Your organisation knows what happened the last three times you tried.
Context can be wrong#
This introduces an uncomfortable problem.
Giving AI access to organisational knowledge doesn’t mean the knowledge is correct.
Imagine an internal document says:
Customers with characteristic X are usually poor prospects.
Perhaps that’s based on careful analysis.
Perhaps somebody wrote it five years ago because three deals went badly.
AI doesn’t automatically know the difference.
The same problem exists with policies, processes, customer records, meeting notes, internal opinions and historical decisions.
Organisational knowledge contains biases, assumptions, errors and outdated beliefs just like human knowledge does.
Connecting AI to more information can therefore make it more informed.
It can also make it more confidently informed by the wrong things.
This is why provenance matters.
Where did this information come from?
When was it created?
Who created it?
Is it still valid?
Is it an official policy or someone’s opinion?
Has newer information superseded it?
How confident should we be?
The challenge isn’t merely helping AI retrieve information.
It’s helping the organisation establish which information deserves authority.
The right context depends on the task#
This becomes even more important when AI starts taking actions.
In May, OpenAI expanded the Responses API to allow reasoning models including o3 and o4-mini to use tools while working through problems.
That means the relationship between context and action is becoming increasingly important.
Consider an AI agent asked to resolve a customer complaint.
Before taking action, it might need to understand who the customer is, what happened, what has already been promised, what their contract allows, what the refund policy says, how valuable the relationship is, whether similar situations have occurred before and whether anything about this case requires human judgement.
Without sufficient context, an agent might still be technically capable of completing the workflow.
It might just complete the wrong workflow extremely efficiently.
That’s an important distinction.
Capability determines whether AI can act. Context helps determine whether the action makes sense.
As businesses give AI more responsibility, that distinction becomes increasingly consequential.
Your data isn’t your context#
This is also why I think businesses should be careful with the phrase “our data.”
It makes the challenge sound predominantly technical.
Where is the data?
What format is it in?
Can the AI access it?
Those questions matter.
But a business is more than its databases.
Context includes information.
It also includes relationships.
History.
Priorities.
Constraints.
Definitions.
Exceptions.
Judgement.
Objectives.
Permissions.
Risk appetite.
Ways of working.
Sometimes the most important piece of context is a number in a database.
Sometimes it’s the fact that everyone experienced enough to understand the number knows not to trust it.
An AI implementation that connects beautifully to every system but misunderstands how the organisation actually operates can still fail.
This is why the problem is larger than data integration.
The objective isn’t to give AI access to the business.
It’s to help AI understand enough of the business to perform a particular job well.
General intelligence may become the commodity#
There is a bigger strategic implication here.
Frontier models are becoming widely accessible.
A business doesn’t have to train a foundation model to access increasingly sophisticated intelligence.
It can use models built by OpenAI, Anthropic, Google and others.
Competitors can do the same.
That means access to general AI capability may become less differentiating over time.
But the context surrounding that intelligence is different.
Your competitors don’t have your customer history.
They don’t have your operating experience.
They don’t have your previous decisions.
They don’t have your internal knowledge.
They don’t have the accumulated lessons from your successes and mistakes.
They don’t have the relationships between your people, products, processes and customers.
They don’t know what your organisation knows.
That doesn’t automatically create an advantage.
Information sitting unused in a filing system isn’t an advantage.
Neither is knowledge trapped in someone’s head.
But if businesses become better at making their proprietary context usable by increasingly capable AI systems, something interesting happens.
The general intelligence may be shared.
What the intelligence has to work with isn’t.
This is the same competitive dynamic I explored in Your Competitors Don’t Need Better AI Than You: access to the technology is not necessarily where the advantage emerges.
Your AI doesn’t know your business#
There is a temptation to judge AI systems primarily by how intelligent the underlying model appears.
That will continue to matter.
Better models will solve harder problems.
Reasoning will improve.
Reliability will improve.
Agents will become more capable.
But for many business applications, model intelligence may only be one part of the equation.
The AI also needs to know enough about the environment in which that intelligence is being applied.
Which information matters.
Which information is current.
Which information can be trusted.
How the organisation works.
What the organisation is trying to achieve.
Where the exceptions are.
What has happened before.
When human judgement is required.
And what a good outcome actually looks like.
The frontier models may increasingly become something every business can access.
The context required to make them useful belongs to the business itself.
And that could become an increasingly important source of differentiation.
Because your competitors may have access to exactly the same artificial intelligence.
They don’t have access to everything your organisation knows.