Last week I wrote about how AI is unbundling middle management by automating coordination work. The response split cleanly down the middle.

Half the replies: "This is already happening at our company."

The other half: "But we're using AI to help our managers write those reports faster."

That second group has fundamentally misunderstood what AI is for. And they're about to get lapped.

The £20-a-month conversation partner

Here's the uncomfortable pattern I see in almost every company I talk to.

They've bought ChatGPT Enterprise or Claude Pro for their teams. Usage is high. Leadership is delighted – look at all that AI adoption! Then you dig into what people are actually doing with it.

The answer is always some variation of the same thing: they ask it questions, get answers, and then manually complete the rest of the workflow.

The AI drafts the email – they send it. The AI summarises the data – they paste it into the deck. The AI generates the report – they format and distribute it. The AI writes the code – they review and deploy it.

Every. Single. Time.

You've automated the thinking but kept all the busywork. You're not using AI – you're having expensive conversations with it.

Two modes, completely different outcomes

There are really only two ways to use large language models in business, and the gap between them is enormous.

Prompt mode: You identify a task, ask the AI to help, review the output, and then manually complete whatever comes next. The AI handles maybe 20% of the total work – the creative or analytical bit – while you handle the other 80%: gathering inputs, formatting outputs, moving information between systems, notifying people, updating trackers.

Agent mode: A trigger fires – a calendar event, a new file, a ticket status change – and the AI executes the entire workflow autonomously. It pulls the data it needs, processes it according to rules you've defined, delivers the output where it needs to go, and notifies the right people. You only get involved when something genuinely requires human judgement.

The difference isn't incremental. It's categorical.

When you prompt AI, your people get slightly more productive at the work they already do. When you deploy agents, your organisation gains parallel capacity – work that happens whether anyone's watching or not.

What this actually looks like

The management coordination example from last week wasn't about using ChatGPT to draft better status reports. It was about status reports happening automatically while the manager did something else entirely.

The work moved from "task the manager does with AI help" to "task that completes whether the manager is there or not." Friday afternoon, the agent pulls data from Jira, Salesforce, and the finance dashboard. It applies the format and commentary rules it's learned. It drops the deck in SharePoint and posts a summary in Slack. It flags three anomalies that need Monday review. The manager glances at the anomalies over coffee on Monday morning and moves on.

Total human time: five minutes. Total coordination work completed: what used to take three hours.

That's not productivity improvement. That's multiplication.

And it's not special. The same pattern works for onboarding task coordination, compliance checking, pipeline reviews, incident triage, procurement routing, customer research synthesis – any workflow that happens repeatedly, follows a pattern, and involves moving information between systems.

The question isn't whether the work can be automated. It's whether your organisation is ready to ask.

The economics everyone's missing

Let me be precise about what this means in business terms, because this is where the gap becomes obvious to finance directors.

Most companies are measuring AI value like this: 50 knowledge workers each save 30 minutes daily using ChatGPT. That's 25 hours saved per day, 6,250 hours annually. At £40 per hour loaded cost, you've created £250,000 in value. Brilliant ROI on a £12,000 annual ChatGPT Enterprise spend.

But here's what that calculation misses.

Those 25 hours saved? Your people still work full days. You've made them more efficient at their existing work, which is valuable, but you still have a 50-person team doing what 50 people can do. You haven't changed your capacity.

Now run the agent mode calculation. Same 50-person team, but now ten workflows run autonomously. Weekly reporting, onboarding coordination, compliance checking, pipeline updates – things that previously consumed human hours. Each agent does work equivalent to eight human hours per week. That's 80 hours weekly, 4,160 hours annually. But agents work nights, weekends, and holidays, so the real effective capacity is closer to 6,000 hours.

Your 50-person team now has the output capacity of 53 or 54 people. You haven't hired anyone. You've created parallel capacity that exists alongside your human workforce.

The gap compounds over time. Year one: ten agents give you the output of 54 people. Year two: 25 agents give you the output of 60 people. Year three: 50 agents give you the output of 68 people.

Meanwhile your competitor who stuck with prompt mode has 50 people who are each 15% more productive. Still 50 people worth of output.

When you both try to scale to 100-person teams, you need 88 people. They need 100. That's not a rounding error – that's £900,000 in annual salary difference, compounding every year after.

This is why the companies that figure out agent mode first will be operating on a fundamentally different cost structure within three years. It's not that they're more productive. It's that they're running on less friction.

Why companies stay stuck

It's not a technical problem. Current tools – Claude, ChatGPT, Make.com, Zapier, plain API connections – are more than capable of building autonomous workflows. The barrier is conceptual.

Prompting fits how organisations already work. You're still in control. You ask, the AI answers, you decide what to do with the output. It's comfortable. Incremental. Safe.

Building agents requires letting go of control, and most companies aren't institutionally ready for that conversation.

Here's what I mean. When you prompt AI, you never have to answer:

  • What exactly needs to happen here?
  • Who approves what, and when?
  • What's a genuine exception versus routine work?
  • When should a human actually intervene?

Most workflows have never been properly articulated. They exist as "how Sarah handles it" or "we figure it out in the moment" or "just escalate anything weird." That ambiguity is fine when humans are running things because humans are brilliant at handling ambiguity on the fly.

But you can't hand a workflow to an agent without defining it. The agent needs to know: when this happens, do that. If you see this pattern, escalate. If the value exceeds this threshold, get approval. Otherwise proceed.

Prompting lets you avoid that entire conversation. Agents demand it.

This is why the companies pulling ahead aren't technically superior. They're institutionally braver. They're willing to ask "could this just happen automatically?" about workflows everyone else assumes need constant human oversight.

And increasingly, the answer is yes.

What this looks like beyond management

The management coordination example is just one workflow where this pattern applies. You can see the same shift happening everywhere:

Stop pasting release notes into ChatGPT asking "does this meet our compliance requirements?" and then manually filing the answer. Start building a system where the pull request triggers the agent, which checks against your compliance rules, approves or flags exceptions, and posts the result automatically.

Stop asking AI to "summarise these customer interviews" and then copying insights into a research document. Start building a system where a new interview gets uploaded, the agent extracts themes, updates your research database, and flags divergent patterns automatically.

Stop copying deal data into ChatGPT every Monday for your weekly pipeline summary. Start building a system where Monday at 9am the agent pulls the data, generates the executive summary, identifies deals slipping, and posts everything with relevant mentions automatically.

Same pattern everywhere: AI stops being a tool you use and becomes a process that runs.

The real divide that's forming

In two years, there will be two types of companies, and the difference will be stark.

Type A companies: Everyone knows how to write excellent prompts. They've rolled out training. Usage is high. They're seeing 10–20% productivity gains across the board. Leadership talks proudly about their "AI-first culture."

Type B companies: Almost nobody writes prompts anymore, because the work happens automatically. Weekly reports generate themselves. Compliance checks run on every deployment. Customer research aggregates continuously. Pipeline reviews appear every Monday whether anyone asks for them or not. They're operating with 20–30% fewer people than their headcount would suggest, because a significant chunk of coordination work has moved out of human hands entirely.

Type A companies think they're winning because their people are more productive.

Type B companies are winning because their organisations are more productive.

The gap isn't visible yet because both look like "AI adoption" on the surface. But it's widening every quarter, and it'll be obvious and unbridgeable within 18 months.

The question that changes everything

Next time someone on your team asks ChatGPT for something – a summary, a draft, an analysis, a data pull – ask them one question:

"Could this just happen automatically?"

If they're doing this task weekly, pulling similar data each time, following roughly the same format, delivering to the same people – then the answer is probably yes.

And if the answer is yes, then prompting isn't the solution. It's the problem you're working around.

Every time someone copies and pastes from ChatGPT, they're doing work that shouldn't exist. The question isn't "how do I do this faster?" It's "why am I doing this at all?"

This is the shift most companies are missing. They're optimising the prompt when they should be eliminating it.

Where this goes

The pattern is already clear if you look at how the leading companies are deploying AI.

They started with prompting – everyone did – but they didn't stop there. They identified the repetitive workflows, the ones that happened weekly or daily, the ones that followed a pattern. And they quietly moved them out of human hands.

Now they're building what I'd call agent factories: systems for spinning up new autonomous workflows quickly. The first agent takes weeks to build properly. The tenth takes days. By the time you've automated 20 workflows, you're not building agents anymore – you're just configuring them.

And when agents start triggering other agents – when the weekly report agent automatically kicks off the follow-up briefing agent, which triggers the task-allocation agent – the organisation begins to hum with its own quiet intelligence. Work happens faster because it's not waiting for people to remember to do it.

This is where the economic advantage becomes structural. It's not that Type B companies have better AI. It's that they've fundamentally reduced the friction coefficient of getting work done.

The window is narrow

Here's what concerns me about the companies still stuck in prompt mode: they don't realise they're falling behind.

Their dashboards show high AI engagement. Their people are enthusiastic. Leadership sees the productivity gains and thinks they're winning. But they're optimising for individual efficiency while their competitors are redesigning organisational capacity.

The shift from prompts to agents isn't obvious from the inside because both feel like "using AI." But one is asking questions and the other is building systems, and those are completely different activities with completely different outcomes.

The companies that figure this out in the next 12 months will build an advantage that's very hard to close. Not because the technology is complicated – it isn't – but because changing how an organisation thinks about work is hard, and it gets harder as the gap widens.

Starting to think differently

You don't need to automate everything immediately. You need to stop thinking of AI as something you talk to and start thinking of it as something that runs.

Pick one workflow your team prompts AI for regularly. Weekly reporting, data collection, routine analysis, compliance checking – something tedious that follows a pattern.

Then ask: what would need to be true for this to just happen? What data would the agent need access to? What rules would it follow? What would trigger it? When would it escalate to a human?

If you can answer those questions, you can build the agent. And once you've built one, the next one is easier.

This isn't about technology. It's about whether your organisation is ready to articulate what work actually needs doing versus what work exists because nobody's questioned it yet.

The bottom line

The prompt was training wheels. It showed people what was possible. But training wheels aren't the destination.

Every time you copy and paste from ChatGPT, you're doing work that shouldn't exist. Every workflow you prompt AI to help with could probably run autonomously. The only question is whether you're ready to let it.

The businesses that win the next phase of AI won't be the ones where everyone knows how to prompt.

They'll be the ones where nobody needs to.