An orchestra without a conductor

Picture an orchestra. Every musician is talented. Every instrument works. But there is no conductor, and nobody is keeping time. What you hear is just noise.

That is how most companies use AI today.

We have been here before

Around 1900, most factories ran on a single steam engine. It turned a long shaft running the length of the building, and every machine was connected to that shaft by a belt. Machines were placed in a line to fit along the shaft.

When electric motors arrived, most factories replaced the steam engine with one big electric motor, but they kept the shaft. Productivity barely moved at all because the machines were in the same place.

The gains came when each machine got its own small motor. Freed from the shaft, machines could be placed in the order the work flows. The real gains came only when factories were redesigned for the new technology.

It took about twenty years for US manufacturing to go from 5% to 50% electric power.

Most companies today are at the second step. They have the new power, but use it like the old one: writing emails in ChatGPT, inside a process that has not changed. They are paying for AI but getting little in return. ROI only materialises at the third step, when processes are redesigned for AI.

The musicians are excellent

Today’s models are far more capable than the way most companies use them. In a blind test, professionals with about 14 years of experience each did a real task from their own job. On average it took them about seven hours. The AI did the same task about 90 times faster, in a single attempt. Then experts compared the results without knowing which was which.

But nobody is conducting

Use is broad, shallow and mostly unplanned.

Almost every large company claims it uses AI. But four in five businesses do not. And very few use it well. Most use is chat: drafting, summarising, tidying up emails. The most powerful tool is being used for the smallest, most insignificant tasks.

Five things hold it back

The evidence points to five major factors holding businesses back. None of them is the model, and the biggest is the organisation itself.

The organisation. A five person team can change how it works in a week. A company of three hundred has layers of approval, processes that other processes depend on, and roles built around how the work is done today. And larger companies have layers upon layers of invisible organisational friction stemming from office politics, culture and competing priorities. Larger companies are more likely to buy AI, but every extra layer makes changing the work itself harder. The technology is the easy part. People and structures are the project. This is why I advocate for fixing low hanging fruit first, the highest friction workflows, and working horizontally across the organisation from there. Changing an entire organisation in one hit is almost impossible. Incremental changes that travel from department to department are not.

Data. AI is only as useful as the data it is given. After studying a million business uses of its own model, Anthropic concluded that complex work is held back more by access to the right information than by what the model can do.

Process. Most companies add AI to the work instead of redesigning the work. Large and older companies find this hardest. Their processes run on systems built years or decades ago, and each system depends on the next. They are prisoners of their own legacy. Newer and smaller companies can build around AI from the start. That is the electric motor on the old shaft.

Know how. AI is excellent at some tasks and unreliable at others, often side by side. Using it well means building checks around it, so it is reliable where it matters. Most people have never been shown how, so they either trust it too much or give up on it.

Permission. People follow their manager. Employees whose manager clearly backs AI use are 8.7 times as likely to say it has transformed their work. In other words, sometimes the biggest source of friction comes from the company's own misaligned leadership.

The saving is in the design

A recent test gave two interesting results: If you use an AI’s response without checking it at all, it is 474 times cheaper than paying an expert to do the same task. However, when an expert then checked and fixed every answer, the AI was only about 1.2 to 1.6 times cheaper.

That's the result of a poorly designed process. Build the workflow so most of it is checked in code, and people only handle the cases that need judgement. Then the saving moves closer to 474.

This is also one reason so many enterprises using AI cannot prove an ROI. They pay for the AI, and then get a human to redo the work.

What closes the gap

The measured wins have one thing in common. The work itself changes, and someone counts the output.

In an experiment at Procter & Gamble, one person working with AI matched the performance of a two person team working without AI. That is not faster email. It is a different way of working.

In one consultancy’s survey, the companies getting real ROI from AI were nearly three times as likely to have fundamentally redesigned how the work is done.

Compare that with handing everyone a chat assistant. When 20,000 UK civil servants got one, they said it saved them 26 minutes a day. A separate department that checked found no evidence productivity improved. That is the second step in the diagram: faster drafts, same process, nothing to show for it.

Saving time is one thing. Deciding what to do with saved time is another. Twenty six minutes a day, spread across a team with nothing new to do, disappears into longer meetings and slower email. If you don't have a plan for how to spend saved time, it will go nowhere. If people simply work shorter hours or take longer breaks, ROI is unaffected. However, if saved time is allocated towards increased output, that will show up as ROI.

Minutes saved feel good. Output is what shows up in the P&L.

Where to start

1. Pick one process that costs real money and runs every week. Ideally it's a boring admin process that staff dislike. Measure how it performs today.

2. Rebuild it completely around AI instead of just adding AI to it. Give the AI the data it needs, and decide where a person signs off.

3. Give it a small team with the authority to change how the work is done, outside the usual layers of approval.

4. Decide in advance how the saved time will be spent. Measure the output, not the minutes. Then move to the next process.

What happens if you wait

Every few months new and improved models are released. A company that has rebuilt a process properly around AI can simply plug the new model straight in and take the gain. A company still at the chat window phase gets slightly better emails.

That gap compounds. In one consultancy’s survey, the small group of companies furthest ahead on AI grew revenue 1.7 times faster than the laggards.

And size works against you. A ten person firm can rebuild a process in a month. A company of three hundred can take a year of meetings and committees before even starting. Electricity showed the same pattern. The first factories to benefit were new ones, built for electricity from the start. Older plants held back, because rebuilding machinery that still worked did not seem worth it yet.

The bigger the orchestra, the longer it takes to play together. The ones that start late may never catch up. This presents an opportunity for smaller companies to compete with larger legacy businesses across most industries. The slower they are to adapt, the bigger the chance a smaller and more nimble business will eat their lunch.

The musicians are already in the room. Someone just needs to write the score and pick up the baton.

Sources

McKinsey, The State of AI in 2025, November 2025

OpenAI, GDPval, arXiv, October 2025

METR, Time Horizons, 2025 to 2026

BCG, Are You Generating Value from AI? The Widening Gap, September 2025

Gallup, AI in the Workplace indicator, May 2026

US Census Bureau, Business Trends and Outlook Survey, May 2026

Chatterji, Deming and others, How People Use ChatGPT, NBER, 2025

Anthropic Economic Index, September 2025

Atkeson and Kehoe, NBER working paper on the electrification of manufacturing, 2001

Dell’Acqua and others, The Cybernetic Teammate, NBER, 2025

UK Government, Microsoft 365 Copilot cross government findings, June 2025

UK Department for Business and Trade, Copilot evaluation, August 2025