The next shift in AI isn't better prompting. It's handing over the workflow.

For the first few years of generative AI, we learned how to ask it for things. The next phase is different: increasingly, we are giving AI the work itself.

For most businesses, using AI has meant opening a chat window. You ask it to write an email, summarise a document, generate some ideas, analyse a spreadsheet or help with some code. Then you take the answer and do the next bit yourself.

That is useful. I use AI like that every day. But I do not think that is where the biggest change is happening any more.

The more interesting shift is from "do this task for me" to "take this piece of work and get it done". That sounds like a small difference. It is not.

A prompt normally ends with an answer

Traditional generative AI is essentially a back-and-forth. You provide something, the AI responds, and you decide what happens next.

Suppose you want to research five competitors. You might ask AI to help identify them, then visit their websites, copy information into a document, ask AI to analyse it, put the results into a spreadsheet, write a summary, create a presentation, and email it to somebody.

AI may have helped at almost every stage. But you were still the one moving the information between systems, deciding what happened next, noticing when something was missing, and keeping the whole thing moving.

You were still the workflow.

That is beginning to change.

What happens when AI gets the workflow?

Imagine giving AI the outcome instead: "research our five closest competitors, compare their positioning and pricing, put the findings into our competitor spreadsheet, and prepare a summary of anything we should discuss at Friday's meeting."

Now there is not one task. There are several. The AI has to work out what needs doing, find the information, visit different sources, extract the relevant details, compare them, use another system, create the output, and potentially notice if something is not available and work around it. That is much closer to delegation than prompting.

That is where AI agents and the new generation of AI work tools become interesting.

ChatGPT Work is one example of the change

The latest generation of AI tools is increasingly designed around completing work rather than simply generating responses. ChatGPT Work is a good example. Instead of the normal question-and-answer pattern, the AI can be given a larger piece of work and use a browser, files, connected services and other tools to carry it through: researching across multiple websites, working through a collection of files, gathering information from different systems, analysing it, and creating a finished document.

Why businesses need to stop thinking only about prompts

Prompting still matters. Clear instructions matter. Context matters. But I would not make "getting better at prompts" the centre of an AI strategy. The bigger question is: what work are people repeatedly moving through your business?

A customer enquiry arrives. Somebody reads it, checks a system, finds some information, updates another system, writes a response, creates a task, tells somebody else, and schedules a follow-up. None of those actions takes very long on its own. Together they form a process, and businesses are full of them. That is where AI starts to get much more interesting.

The unit of AI is getting bigger

Here is how I think about the change. First, AI generated words. Then it became useful for tasks. Now it is beginning to handle workflows. Eventually, in some areas, we may give it responsibility for outcomes.

Those are not clean technical stages and they overlap enormously, but they are useful for understanding what is happening. Writing one email is a task. Handling an incoming enquiry, checking the customer's history, drafting the right response, updating the CRM and scheduling the next action is a workflow. Making sure every qualified enquiry receives the right response and follow-up is closer to an outcome. Each step requires the AI to understand more of the business around the work.

This is where the boring systems suddenly matter

AI demos tend to focus on the impressive bit: the model, the conversation, the generated answer. In a real business that is often only one piece. The useful information might be in HubSpot, an ecommerce platform, a database, Google Drive, a support system, an accounting platform, or a spreadsheet somebody has maintained for seven years.

The AI needs access to the right information. It needs to know what it is allowed to change. It needs to pass information between systems. It needs rules, permissions, checks, and somewhere to record what it has done. Which means the more capable AI becomes, the more important the infrastructure around it becomes too.

Why my old work suddenly feels very relevant

Before AI became the centre of what I do, I spent around 15 years working in web development and digital: websites, ecommerce, databases, APIs, CRMs, analytics, integrations. Before that, I worked in infrastructure roles where the job was often about understanding how a business actually operated and putting the processes and systems underneath it.

AI has not made that experience irrelevant. Quite the opposite. Because once you move beyond asking AI to write something, you very quickly arrive back at: where is the data? Which system owns it? What happens next? What can be automated? Who needs to approve it? What happens when something goes wrong? How does one system talk to another? The model may be the clever new part. The workflow around it is still a workflow.

AI agents do not remove the need for process

There is an appealing idea that sufficiently clever AI will simply work everything out. Sometimes it can. But businesses contain rules the AI cannot magically know. A refund above a certain amount needs approval. A particular customer must be handled differently. This information can be read but never changed. That email must come from this address. This system is the authoritative record. This action must be logged. A human must check this before it goes live.

Those are not prompting problems. They are process-design problems, and the more autonomy we give AI, the more important they become.

The real question: what should the AI be allowed to do?

There is a big difference between "read this information" and "change this information". There is another jump between "prepare this email" and "send this email". And another between "tell me what you think we should do" and "do it".

Every useful AI workflow needs some thought about authority. What can the AI see? What can it change? What can it send? What can it spend? What requires approval? What happens if it gets something wrong? The future of business AI is not giving agents access to everything and hoping they are clever enough. The best systems give AI enough authority to be useful without giving it unnecessary authority to cause problems.

Human approval is not a failure of automation

There is sometimes an assumption that the perfect AI system has no humans in it. I do not find that useful. Take an AI that researches a prospect, checks the CRM, prepares a personalised proposal and creates the follow-up. Perhaps the human should still approve the proposal before it is sent. The AI may have removed 90% of the work; keeping a human at the point where judgement or accountability matters does not make the automation unsuccessful. It may make it better.

The goal is not "remove the human". It is "remove the work that did not need the human in the first place".

This changes how I look for AI opportunities

If I were looking at a business now, I would not start by asking "where could you use ChatGPT?" I would look for repeated movement. Where does somebody receive information and then put it somewhere else? Where are people repeatedly checking the same things? Where does one person copy information between systems? Where are documents created from information that already exists elsewhere? Where are people chasing things that could have been followed up automatically? Where does a process stop because somebody has to remember the next step? Where is a skilled person spending time on the administration surrounding the skilled part of their job? Those are often much better places to start.

Some workflows are much better candidates than others

I would not hand every business process to AI. Good early candidates tend to share a few things: they happen repeatedly, the information needed is available, the desired outcome is reasonably clear, mistakes can be detected, the consequences of a mistake are manageable, and there is an obvious point where a human can step in.

The worst place to start is usually a rare, badly understood, high-risk process where nobody agrees what "correct" looks like. AI does not magically fix a broken process. Sometimes it just lets the broken process happen faster.

Automation and AI are not the same thing

Businesses have automated processes for decades: if this happens, do that; move this information here; send this notification; create this record. Traditional automation is excellent when the rules are clear. AI becomes useful when part of the workflow needs interpretation: understanding what somebody means, extracting information from an unstructured document, deciding which category something belongs in, writing a response based on context, comparing several sources, handling variation that would otherwise need hundreds of rules.

The really interesting systems combine both. Automation handles what is predictable; AI handles what needs interpretation. That combination is far more powerful than either on its own.

This is not about replacing entire jobs

A job is normally a collection of many kinds of work: some repetitive, some administrative, some creative, some interpersonal, some based on experience, some requiring judgement. AI will be excellent at some pieces and poor at others. So "which jobs will AI replace?" is often the wrong first question. A more useful one is: which parts of this person's work should still require this person? If someone has years of experience, you probably do not want them spending half their day moving information between systems. Use the technology there, and keep the human where their judgement matters.

The businesses that benefit most may not have the fanciest AI

A company can have access to the world's best AI model and still get very little from it. Another can use the same model inside a well-designed process and gain an enormous advantage. The model is not the whole system. The surrounding data matters. The integrations matter. The workflow matters. The permissions matter. The checks matter. And understanding what the business is actually trying to achieve matters. AI capability is becoming easy to access. Knowing what to build around it is where much of the value sits.

So forget the perfect prompt

Not literally, good instructions still matter. But businesses have spent several years being told the secret to AI is learning how to talk to it. That was useful while AI mostly sat in a chat box. It is becoming less useful as a complete way of thinking about the technology.

The next question is not "what should I ask AI?" It is "what could I hand over?" A task? A research process? A piece of administration? A customer workflow? A reporting process? A sequence that currently passes through four systems and three people? That is where things get interesting, because once AI can work across the process, it stops being something you occasionally ask for help. It starts becoming part of how the business operates. For me, that is a much bigger change than learning how to write a better prompt.

Where to go next

• AI agents aren't digital employees. Here's what they actually are. What an agent actually is, and where it fits inside a process.
• AI can now use the computer. What changes when AI operates the same software your people use.
• The important question isn't what AI can do. It's what you should let it do. Permissions and authority as AI starts taking actions.

Related: AI agents aren't digital employees. Here's what they actually are..