I’ve implemented some of the first operational AI use cases at three different companies.

One of the things I keep noticing is how much the experience of using AI has distorted our expectations of implementing it.

You can open ChatGPT, type a question, and get a reasonably useful answer in seconds. You don't need to understand the underlying technology. You don't need to know how to write code. You don't even need to be particularly good at explaining what you want.

That accessibility is extraordinary.

It's also misleading.

Because making AI useful inside an organization is a very different problem from making AI useful to an individual sitting in front of a chat window.

And I'm beginning to think that even the organizations solving that problem well may be asking a relatively small question.

We're spending an enormous amount of energy figuring out how AI can do the work we already do.

I'm not sure we're spending nearly as much time asking why we do the work that way in the first place.

The Interface Is Not the System

Large language models are trained, in part, to predict what comes next in a sequence of tokens. That sounds almost absurdly simple compared with what they're capable of producing.

The capabilities that emerge from that training are much more sophisticated than autocomplete. But the distinction between generating a plausible response and establishing that a response is correct matters.

A model can explain a company's revenue model without knowing which of its thirteen revenue fields is authoritative.

It can summarize a customer relationship without knowing that the customer exists under four different account records.

It can recommend a next action without understanding which person has the authority to take it, which contractual obligation it might affect, or what happened the last time someone tried.

These aren't necessarily problems with the model.

They're problems with the information and organizational architecture surrounding it.

I've worked with AI engineers who could get models to perform remarkably sophisticated tasks but struggled to connect those capabilities to the organization's actual data.

And I don't mean connecting an API.

I mean understanding what the data represents.

Which records describe the same customer? Which fields are reliable? Which definitions changed over time? What information can a particular person access? What happens when two systems disagree? Which exceptions are meaningful?

Those questions aren't unique to AI. They're the questions I've spent much of my career working through.

AI just makes the consequences of not answering them more interesting.

When a person encounters contradictory information, they may recognize the ambiguity. They may know whom to ask. They may remember the meeting where the definition changed, or the customer who required an exception.

Some of that judgment is documented. Much of it isn't.

An AI system can retrieve information, reason over it, and use tools to take action. But giving it access to more data doesn't automatically give it the context required to interpret that data correctly.

Knowledge graphs, retrieval systems, tool connections, and shared data models can help. They aren't interchangeable, and none of them eliminates the need to decide what the information means.

If the organization itself doesn't understand its data well enough to explain how it should be used, configuring AI to use it reliably becomes considerably harder.

There's a possibility here that I find both exciting and uncomfortable.

AI can make an organization's systems much easier to interact with without making those systems any more coherent.

It can also make the incoherence less visible.

A conversational interface can produce one clean answer from information that was never designed to support one.

And when that answer is wrong, the error may look less like a broken integration and more like a confident explanation.

We're Automating the Organization We Already Have

Most of the conversations I encounter about implementing AI begin with an existing activity.

Can we automate this research?

Can we summarize these meetings?

Can we eliminate these handoffs?

Can we reduce the time it takes to produce this report?

Can we do the same work with fewer people?

I don't think these are bad questions. I've built systems to do exactly those things, and the results can be meaningful.

Reducing repetitive work matters. So does giving people better information, removing unnecessary administrative burden, and making services more accessible.

But all of these questions begin with an assumption: that the work we've identified is work that still needs to exist in roughly its current form.

Consider a fairly ordinary organizational workflow.

Sales collects information about a customer. Operations cleans it up and reconciles it with existing records. Finance validates the commercial details. Someone builds a report. Leadership reviews the report and decides what to do.

There may be several applications, multiple teams, and a collection of meetings holding the whole thing together.

AI could make every one of those steps faster.

It could research the customer, reconcile records, validate information, generate the report, and summarize its findings.

That's useful.

But why were those activities separated in the first place?

Some of the separation exists for good reasons. Financial controls, independent review, privacy, accountability, and specialized judgment aren't simply inefficiencies to eliminate.

Other boundaries may exist because information was historically difficult to collect, expensive to move, or impossible for one person to interpret without help.

We created departments, roles, reporting structures, and processes partly in response to those constraints.

Over time, those arrangements became the organization.

Now we're introducing a technology that changes some of the constraints.

And much of our implementation strategy is to make the existing arrangements operate faster.

I wonder how often we're automating the cost of an organizational decision without reconsidering the decision itself.

What would that workflow look like if the information were available at the point of decision?

Which responsibilities would still need to be separate?

Where would human judgment become more important, rather than less?

What would happen to the relationship between the person serving the customer and the people responsible for fulfilling the promise?

I don't think the answer is necessarily fewer departments or flatter organizations. Those are structural preferences, not conclusions.

I think the answer begins with understanding what the current structure is doing, including the things it does that we don't measure.

Then we can ask which parts are still necessary under different conditions.

We could become extraordinarily efficient at doing work we no longer need to organize that way.

The World Outside the Organization Isn't Standing Still

There's another part of this that I've been thinking about.

While companies are figuring out how to insert AI into their existing operations, AI is already changing the environment those operations depend on.

I was reading about how AI systems generate product recommendations.

The interesting question wasn't whether the models could compare blenders or running shoes. It was which information they had access to in the first place.

Some publishers invest heavily in independent testing and restrict access to their content by AI crawlers. Other websites produce enormous quantities of commercially motivated recommendation content, often structured specifically to appear in search results and AI-generated answers.

That means an AI assistant's answer may reflect not only what information is accurate or useful, but what information is accessible and prominent in the systems supplying it.

A highly regarded independent review may be absent. A commercially optimized list of the ten best products may be readily available.

The assistant can still produce a coherent recommendation.

The user may never see the information that wasn't considered.

Of course, this isn't entirely new.

Search engines have influenced how people publish information for decades. Advertising, affiliate relationships, and search optimization have long shaped what we encounter online.

But AI changes the interface between the person seeking information and the sources supplying it.

Instead of presenting a collection of links that a person evaluates, an assistant can synthesize information into an answer, make a recommendation, or increasingly take an action.

That changes the incentives for everyone producing the underlying information.

Publishers have to decide whether to allow AI systems access to their work.

Businesses have to consider how their products and services become visible to automated agents.

Content producers have new incentives to create information that machines can retrieve, summarize, and recommend.

And the people using these systems may become less aware of which sources were excluded, which claims were commercially motivated, or which alternatives were never presented.

The information environment starts adapting to the technology interpreting it.

Then the technology interprets the environment it helped reshape.

I'm interested in that loop.

Not just because of the potential for misinformation, although that's certainly part of it. I'm interested because the mechanisms through which we discover, evaluate, and trust information are part of the infrastructure of society.

They influence what gets purchased, which businesses survive, which institutions retain authority, and whose knowledge remains accessible.

If automated agents increasingly mediate those decisions, the design of the information environment becomes consequential in ways that extend far beyond a better search experience.

What Are We Actually Designing?

I think there's a strange contradiction in the way we're approaching AI.

Inside organizations, we're trying to make it conform to the systems we've already built.

Outside organizations, it's already changing the conditions under which those systems operate.

We're teaching AI to navigate existing workflows while the economics of information, discovery, and decision-making begin to shift around us.

And we're doing this with technology whose behavior depends heavily on the information, instructions, incentives, and tools surrounding it.

That's worth paying attention to.

A conversational assistant may be designed to be helpful, agreeable, and responsive. Those qualities make it easier to use, but they can also create problems when a user needs contradiction rather than affirmation.

An enterprise agent may be instructed to complete a task efficiently. But efficiency relative to what? The immediate request? The organization's policies? The customer's interests? The long-term consequences of the action?

The model doesn't independently resolve all of those questions simply because it can generate an answer.

Someone is still making choices about objectives, constraints, access, evaluation, and acceptable behavior.

Those choices become part of the system.

And as AI moves from producing information to taking action, they become more consequential.

I don't think we should abandon the practical work of implementing AI well. Quite the opposite.

Data architecture matters. Context matters. Integration matters. Governance matters. Human judgment matters.

I've seen enough implementations to know that these are not trivial problems, and that getting them right can create substantial value.

But I also don't want us to mistake solving those problems for understanding the larger change.

There are at least two different things happening.

We're building AI into the organizations and institutions we already have.

And AI is beginning to change the conditions those organizations and institutions were designed to operate within.

The first requires good engineering and organizational design.

The second requires imagination, observation, and a willingness to question arrangements we've come to treat as inevitable.

I don't know what the organization of tomorrow looks like.

I don't know which boundaries will become unnecessary, which new ones we'll need, or how much of the information environment we're already changing without fully understanding the consequences.

But I suspect the most consequential questions about AI won't be answered by measuring how much faster it can complete today's work.

They'll come from noticing what changes when the assumptions underneath that work no longer hold.

And asking whether we're willing to redesign the system, rather than simply making it faster.