Most firms that measured their AI savings found less than they had planned
Bain asked 951 large companies what they expected automation and AI to save and what it did save, and the two numbers came apart. A team moving 800,000 store sensors to a new network kept a table of every default it had to replace.
How we Organise
Most firms that measured their AI savings found less than they had planned
Bain surveyed 951 companies around the world about automation and AI. The commonest target for savings from automation and AI was 11 to 20 per cent, set by 37 per cent of them. Among firms that measured the result, 40 per cent landed at 10 per cent or less. Only 7 per cent run agents with no person approving or watching, and Bain says many business cases were priced as if every agent ran that way. The largest group, 44 per cent, plans to pay for its next round of agents out of savings from the last round. Nine in ten are raising their budgets again.
We tell a finance team not to count a saving until the accounts show the cost has gone.
A saving is operational cost that has left the profit and loss account.
A firm paying for agents with savings it projected, rather than savings its ledger records, is spending money it may never see. Every figure is what a firm reported about itself to a consultancy that sells this work.
How we Build
A sensor fleet team rebuilt its defaults from what each experiment showed
Kuldeep Singh describes moving about 800,000 battery-powered sensors in 3,000 stores to a new network without replacing the installed devices. His team arrived with web and cloud habits. They changed a habit whenever real hardware contradicted it. A farm of physical sensors became part of the development environment, and when large stores hit network limits an emulator stood in for hundreds of devices. Firmware went out in stages. The team saw what an update did before sending it everywhere. The article ends with a table of thirteen assumptions and the default that replaced each. Its last row says to record discoveries as well as decisions.
Our position is that each piece of work ships and is checked before the next is planned, and that the check decides what comes next.
Work moves in increments thin enough to be shipped and evaluated before the next is scoped, and what an increment proves changes what the next one is.
That table is the record we ask for, kept by a team nobody asked to keep it. It is one team’s account of its own project, with no figures on rework or delivery time, and the author says an AI assistant helped edit it. A September study of 3,553 coding-agent sessions found that a requirement arriving after the first edit was followed by about twice as much discarded code, and that warning in advance made no detected difference. This account does not measure rework, so it does not speak to that finding.