The most important AI may be the AI you stop noticing

Right now, using AI usually feels like using AI. You open something, ask it a question and wait for the answer. I don't think that's where the most useful business AI ends up.

When most people think about using AI at work, they probably still imagine going somewhere to use it. Open ChatGPT. Open Claude.

Open Copilot. Type something. Get something back.

Copy it. Change it. Put it somewhere else.

That's enormously useful. But it's still a very visible interaction with AI. You decide you need AI.

You go to the AI. You give it some work. Then you bring the result back into whatever you were doing.

I think a lot of business AI will eventually feel much less like that. It will simply become part of how the work happens.

Think about how we use technology already

You probably used dozens of pieces of technology before lunchtime today without thinking particularly hard about most of them. An email arrived. A payment was processed.

A calendar invitation appeared. A customer record updated. A website loaded.

A file synchronised. A notification was triggered. A password was authenticated.

A database returned information. You didn't stop each time and think: I'm using cloud computing now.

Or: What an exciting database interaction. The technology is underneath the experience.

It becomes infrastructure. That's what happens when technology matures.

Good infrastructure tends to disappear

Earlier in my career, I worked in infrastructure environments. One of the things about infrastructure is that when it's doing its job properly, most people aren't particularly interested in it. They want:

the system to be available, the information to be there, the process to work,

the connection to connect, and the thing they need to happen to happen. They tend to become very interested in the infrastructure when it stops working.

I think AI may follow a similar path. Right now, we're fascinated by the AI itself. Eventually, in many situations, we may care much more about the outcome it helps produce.

The AI button may be temporary

We've gone through a phase where putting an AI button into software is itself a feature. Generate with AI. Ask AI.

Summarise with AI. Write with AI. That makes sense while people are learning what the technology can do.

But imagine a CRM where AI is genuinely embedded into the workflow. Perhaps you don't click: Ask AI to summarise this customer.

The useful information is already there when you open the record. You don't click: Ask AI what happened since I last looked.

The changes that matter are already highlighted. You don't click: Ask AI to draft a follow-up.

A suitable draft is waiting when the process reaches that point. The AI hasn't disappeared technically. It's disappeared experientially.

You're just doing the work.

That's a bigger shift than adding a chatbot

The first wave of business AI has often looked like this: existing process + AI assistant You do the job much as you did before, but AI helps at certain points.

That's useful. The next shift looks more like: redesigned process with AI inside it

Now the employee doesn't necessarily have to decide when to use AI. The system already knows where AI belongs. That's a much more interesting form of implementation.

The human shouldn't have to be the integration

Imagine a salesperson receives an enquiry. They copy it into an AI tool. Ask for a summary.

Copy the summary into the CRM. Ask AI to draft a response. Copy that into email.

Send it. Create a follow-up task. That's AI-assisted work.

But the human is still moving everything around. The human is the integration. A better system might:

receive the enquiry, understand what it's about, find the existing customer information,

prepare the relevant context, draft an appropriate response, create the CRM record,

and prepare the next action. The salesperson sees what matters and steps in where their judgement is useful. They don't necessarily need to think:

Time to use AI. The workflow already contains it.

This is what infrastructure means to me

Calling AI "infrastructure" can sound grander than it needs to. I don't mean every company needs its own AI platform, enormous technical team or complicated architecture. I mean something simpler.

AI becomes a capability that other systems can call upon when they need it. Need to understand this message? Use AI.

Need to extract information from this document? Use AI. Need to compare these records?

Use AI. Need to prepare a response? Use AI.

Need to decide which workflow should run? Potentially use AI. The person using the business system doesn't necessarily need to know which model performed which part.

They need the process to work.

We already expect layers of technology to disappear

When you buy something online, you don't normally care which database stores the order. You care that the order exists. When you make a card payment, you don't want to understand every service involved in processing it.

You care that the payment is secure and successful. When you search a website, you don't want a lesson in its search infrastructure. You want the right result.

AI will increasingly become another layer inside systems like these. Not everywhere. Not for everything.

But wherever interpretation and intelligence improve the process.

This changes how businesses should think about adoption

If AI is becoming part of infrastructure, then "How many employees use ChatGPT?" becomes a fairly weak measure of AI adoption. A company could have hundreds of employees opening an AI chatbot every day and still have very little AI embedded into its actual operations. Another company might have relatively few employees directly interacting with a standalone AI tool while AI quietly supports:

customer enquiries, document processing, internal search,

reporting, software development, quality checks,

sales workflows, and administration. Which company is "using more AI"?

I'm not sure the question matters. The better question is: What has become possible or improved because AI is there?

The interface matters less than the capability

We're currently very focused on AI products. Which chatbot? Which model?

Which subscription? Which new feature? Businesses obviously need to make those choices.

But over time, I think the model itself becomes less visible in many implementations. A customer-service system might use one model for one task and another for something else. A workflow may switch providers.

A model may be upgraded. A specialised system may handle part of the process. The employee shouldn't necessarily need to care.

That's an architectural decision underneath the work. This is one reason I wouldn't build a business AI strategy entirely around one AI product. Products change.

The business requirement is more durable.

Start with the capability you need

Instead of saying: We need ChatGPT in this department. try:

We need to understand incoming enquiries more quickly. Instead of: We need an AI agent.

try: We need this process to continue without somebody manually moving it through six systems. Instead of:

We need an AI knowledge bot. try: Staff need to find the correct information without searching for twenty minutes.

Now you've defined something useful. AI might be part of the answer. And if the underlying AI changes later, the business requirement remains.

Invisible doesn't mean uncontrolled

This is important. If AI becomes less visible to the user, the controls around it become more important, not less. Someone still needs to know:

what information the AI can access, what it can change, what actions it can take,

which model or service is being used, where data goes, how outputs are checked,

what happens when it fails, and who is responsible for the process. The employee may not see the AI.

The business still needs to understand it.

There is a danger in AI becoming too invisible

Imagine an employee sees: Recommended action: refund customer £450. Where did that recommendation come from?

A fixed business rule? A person? An AI model?

A combination? What information did it use? How confident should we be?

For low-consequence tasks, perhaps that distinction doesn't matter much. For higher-consequence decisions, it matters enormously. Invisible AI should not become unaccountable AI.

Sometimes the user needs to know that AI contributed to the result. Sometimes they need to review it. Sometimes they need to be able to challenge it.

The interface should reflect the consequence.

The boring parts become really important

Once AI becomes infrastructure, the glamorous demo matters less. The questions become: Does it work reliably?

What happens when it doesn't? How much does it cost at scale? How quickly does it respond?

What data does it need? Which systems can it access? How do we monitor it?

How do we change providers? What gets logged? How do we know it hasn't quietly started performing worse?

Who owns it? These aren't particularly exciting questions. They're the questions that turn a capability into infrastructure.

Dependence changes the standard

There's a big difference between asking AI to help draft a LinkedIn post and embedding AI into a process your customers depend on. If the LinkedIn draft is poor, you rewrite it. If an operational system silently misroutes customer enquiries for three days, that's a different problem.

As AI becomes more deeply embedded, the required standard changes. Reliability. Monitoring.

Fallbacks. Permissions. Ownership.

Those things matter more as dependence increases.

You need to know what happens when the AI isn't there

This is a useful test. If the AI service is unavailable for an hour, what happens? Does the entire process stop?

Does work queue? Does a simpler rule take over? Does it go to a person?

Can the process continue without the AI? There isn't one correct answer. But there should be an answer.

Infrastructure needs failure paths.

The same applies when the AI is uncertain

A mature AI workflow shouldn't need to pretend the AI knows everything. Sometimes the correct behaviour is: I don't have enough information.

Or: This doesn't match the normal cases. Or:

A person should look at this. That isn't a failure of automation. It's good system design.

The aim isn't maximum autonomy. It's appropriate autonomy.

AI could become a shared capability across the business

There's another shift I think we'll see. Right now, different departments may independently buy AI tools. Marketing gets one.

Sales gets another. Customer service gets another. Developers have several more.

That may be appropriate in some cases. But businesses may increasingly think about AI more like a shared capability. Different processes can use it for different jobs while common rules govern:

access, security, data,

approved services, monitoring, cost,

and permissions. That's much closer to infrastructure thinking.

This is where governance stops being a policy document

If AI is something an employee occasionally opens, AI governance can feel like a list of rules: Don't paste this in. Don't use it for that.

Check the output. But if AI is built into workflows, governance can increasingly be built into the system itself. The AI only receives the information it needs.

It only has the permissions required for that task. Certain actions always require approval. Sensitive information follows different routes.

Activity is logged. Exceptions are escalated. The rules become part of the architecture.

That's much stronger than hoping everybody remembers a policy.

The most mature AI implementation may look quite boring

I think this is worth saying. The impressive demo might be: Look, our AI agent can operate the whole system!

The mature implementation might look like: A customer sends something. The right information appears in the right place.

The obvious administrative work happens. A person is shown the one thing that genuinely needs their attention. They make the decision.

Everything else continues. Nobody applauds. Nobody says:

Wow, AI. The work just moved. That's probably a good sign.

We may stop counting AI interactions

Businesses currently talk about AI usage in terms of: users, prompts,

licenses, sessions, and adoption.

Those are useful while organisations are learning. But eventually, the better measures may be ordinary business measures. Response time.

Conversion. Resolution time. Error rate.

Cost per transaction. Employee capacity. Customer satisfaction.

Processing time. Revenue. The AI becomes part of how the outcome is produced rather than the outcome itself.

This is why the AI strategy should connect to the business strategy

If AI becomes infrastructure, it can't sit permanently as a separate innovation project. It needs to connect to: how customers are served,

how information moves, how work gets done, how software is built,

how decisions are made, and how the company operates. That doesn't mean every business needs a grand AI transformation programme.

Quite often I'd start much smaller. One process. One problem.

One measurable improvement. Then another. Infrastructure grows because it proves useful.

There will still be visible AI

None of this means chatbots disappear. There are plenty of situations where a direct conversation with AI is exactly the right interface. Research.

Thinking. Writing. Analysis.

Exploration. Coding. Learning.

Sometimes you want to talk directly to the intelligence. The point is that this won't be the only way we use it. A great deal of AI may sit underneath the applications we already use.

The important AI may be the AI nobody talks about

We're in a stage where AI itself is the story. Every launch gets announced. Every feature gets an AI label.

Every company wants to tell you it has AI. That probably won't last forever. Eventually, some of the most useful implementations won't feel like AI products.

They'll feel like: a faster process, a better search,

a useful recommendation, less administration, a system that understands what you meant,

a customer getting an answer sooner, or an employee not having to copy the same information for the hundredth time. The AI will still be there.

It just won't be the thing anybody cares about.

That's when it starts becoming infrastructure

I spent part of my earlier career around infrastructure, and I think there's something familiar about this transition. Important technology often becomes less interesting as it becomes more essential. We stop marvelling at the component.

We build around it. We expect it to work. We notice when it doesn't.

And we judge it by what the system allows us to do. AI is still very visible today. But I suspect some of the most important business uses will eventually be almost invisible.

Not because AI became less powerful. Because it became part of how the business works.

Where to go next

• Don't start with AI. Start with the work. If AI is becoming infrastructure, the starting point isn't the model. It's the problem the business needs to solve.
• AI vs automation: what's the difference, and when do you need each? How AI, ordinary automation and people fit together inside a workflow.
• The important question isn't what AI can do. It's what you should let it do. Invisible AI still needs visible boundaries around data, permissions and actions.
• AI can now use the computer. That changes what businesses can automate. Why AI's ability to operate existing software could make it part of many more business processes.

Related: Don't start with AI. Start with the work..