Chapter 18
Doing Less
The Utopia of the Agentic Enterprise, Part IV. Doing Less
The British government decided to spend £60 billion on a high-speed railway from London to Birmingham and Manchester, which would take about an hour off the trip to Manchester by the end of the 2020s. Rory Sutherland offered to take 40 minutes off it within six months, for around £250,000. Passengers on advance tickets can only travel on the train they booked, he pointed out, so they get to Euston about 45 minutes early in case they miss it, and in that time two earlier trains leave with empty seats. An app that let them take an earlier train when it had room, for a small payment, would cut the journey without making any train faster.1 As far as he knew, nobody took the idea seriously.
The government’s own trial of an AI assistant was measured like the railway, in minutes saved. From the end of September to the end of December 2024, 20,000 civil servants in 12 organisations used Microsoft 365 Copilot. Asked to estimate how much time it saved them, they put it at 26 minutes a day, with the largest savings in drafting documents, and the report noted that it couldn’t say how the time saved was spent.2 Sutherland asked what the passenger spends the time on, and from that side a good part of the journey is waiting for something that didn’t need to happen.
Some of the work in an organisation exists only because something used to be expensive. A summary of a long report was worth writing when reading the report took an afternoon, and once anyone can ask an agent what the report says about their own question, the summary is a step nobody needs. A developer needed a specification because building a small tool took weeks of someone else’s time; once the requester can build it in an afternoon, after a conversation about which systems it may touch, the specification is a step nobody needs. Stopping work like that costs a decision. Automating it costs a licence and the people who check what the agent produced, and a business case written to automate has no row for the first.
In January 2023 Shopify tried the bluntest version. It wrote code that went through every calendar in the company and deleted the recurring meetings with three or more people. Staff were told to wait two weeks before adding anything back, and to be “really, really critical” about what they brought back. A month later the chief operating officer said people had started declining invitations from senior managers: “People have been saying ‘no’ to meetings from me, and I’m the COO of the company.” Steven Rogelberg, an organisational psychologist who has studied meetings for decades, wasn’t a fan of the purge. Meetings weren’t the problem in themselves, he argued. Bad meetings were.3
Neither of them was wrong. Deleting everything and waiting to see what comes back finds out who needed a meeting by making them ask for it, which is quick. It can’t tell a meeting nobody misses from one that is missed three months later, in a decision taken without the people who used to be in the room. The slower way takes one piece of work at a time, and it used to be expensive, because someone had to read every ticket and calendar entry to do it. That reading is what agents are cheap at. Used this way, an agent doesn’t run the process. It reads it.
Start with a piece of work small enough to stop on its own: one report, one approval, one meeting, one manual step. Write down what starts it and where it ends, and what it’s supposed to be for, as a claim to check rather than a fact. Then list who consumes what it produces. Beside that list goes everyone else who touches the work, including the ones who make up for it without being asked. The agent can draft both lists from the documents and the tickets, and the people who do the work correct them. The same lists can be drawn for an agent, and more easily, since its instructions say what it’s for and its log says who received what it produced.
The question that decides it is the one Shopify asked of everything at once: what happens if this stops for a month? Asked of one item, it has answers. Who would notice, and how soon? What would fail, and who would pick up whatever is still needed, whether or not anyone tells them to? An approval that hasn’t rejected anything in a year looks like an easy stop. It may be deterring the requests nobody sends because they know it’s there, and the month is how to find out, as long as someone is watching for them.
The hardest thing to see is the judgement inside the work. A rule can fix the outcome and still need someone to decide when it applies. Stopping or changing the work moves that decision somewhere, to another person or into a prompt or an exception queue, and someone has to handle the cases the new arrangement can’t, with the information and the authority to do it. If nobody can say who that is, the judgement has gone nowhere, and the first hard case will show it.
Two rules stop it doing harm. A workaround stays until its cause is fixed, because removing the patch while the fault remains passes the fault to whoever is downstream. And it assesses work, never people. It asks whether a report needs to exist, and a version that starts asking whether its author does will stop getting true answers.
Each item comes out with one of four answers: keep it, stop it, change it, or find out the one fact that would decide. A stop needs an owner, and a date on which someone checks whether the work should come back. A change says where the judgement goes. None of it is the agent’s decision. The agent reads the work and drafts the case on each side, and a person decides and puts their name to it.
Stopping things is the easy half. Work comes back unless someone has the authority to keep it stopped, and the people who did it decide whether to help by watching what happens to their time. If the hours go straight into more throughput, they learn to keep the next report going. And if the stopped work was how juniors learned, the organisation has paid for this year’s hours with the next decade’s expertise. Those are decisions about who keeps the gain and who gets to decide, and none of them comes with the agent.
Checklist
- Which steps in your agent rollout exist only because reading or building used to be expensive?
- Pick one report, approval, meeting or manual step. What starts it, where does it end, and what is it supposed to be for?
- Who consumes what it produces, and who makes up for it without being asked?
- If it stopped for a month, who would notice, and who would pick up whatever is still needed?
- What judgement does the work need, and where would that judgement go if it stopped?
- If it’s a workaround, what fault is it covering, and who owns fixing it?
- For each stop, who owns it, and on what date does someone check whether the work should come back?