By Nick Leeder, Owner, Nick Leeder & Co.

In my last article, I argued that your own diary tells you whether a digital programme is drifting. So, say you fix that, and say you are properly close to it, decisions land in days, and everyone knows exactly whose neck is on the block. 

You still have the harder meeting to get through. 

It is the benefits review, the CFO is in the room, and you can describe everything that is live. Sensors on the critical assets, dashboards in the shift huddle, two AI models running in production. What you cannot do, in a single sentence, is say what changed. 

So, the money reads as spent rather than invested, and the next request gets harder, which is the part that hurts 18 months later. 

Why we all reach for the wrong number 

Brian Marsh put the question plainly in this newsletter recently: can you measure success in a smart manufacturing initiative? Yes, he said, but only if you measure business outcomes rather than technology deployment. 

He is right, and I want to add why the wrong thing is so easy to reach for. 

Deployment is countable, because assets connected, users onboarded, dashboards built and models shipped are all things you can pull from a system on a Friday afternoon. Nobody argues with them either. Behaviour change is not countable in that way, and no one can hand you a query that returns “the night shift now acts on the alert instead of noting it”. 

So, we count what counts easily. That is not laziness, because it is the only number that survives contact with a slide deck. 

The evidence is blunter than the conversation 

The World Economic Forum’s latest Global Lighthouse Network work carries a figure worth sitting with for a moment. More than 80 percent of executives report that their firms have seen no measurable EBIT impact from generative AI. Not a slow return, and not a return that is hard to isolate. No measurable return at all. 

The same work tested 25 different transformation attributes to see which of them predicted whether the impact turned up. And the attribute with the largest effect by some distance was redesigning the workflow. 

Read those two findings together and the picture is fairly stark. Firms are buying the tool, leaving the work exactly as it was, and then measuring the tool. 

There is a name for where that ends. The Lighthouse research calls it the scaling slump, which is the point where early momentum stalls because pilots never replicate past the first few sites. You know it when you see it, because the pilot site loves it, the second site is polite about it, and the third never really starts. 

Name the behaviour before you sign 

Here is the fix, and it costs nothing except one slightly awkward conversation before the money moves. 

Before you approve the spend, write down the specific thing that someone will do differently, and where they will be doing it. Not “improved visibility”, but something closer to this: the maintenance planner moves two jobs a week out of the breakdown queue and into planned work. That sentence is testable, because somebody can go and look. 

Then put three questions to it. 

• Who changes what they do, and on which shift?
  If you cannot name the role, then the benefit does not have an owner and it will not arrive. 

• What do they stop doing?
  This is the one everybody skips, and if nothing comes off the plate, the new thing becomes extra work and gets dropped in the first busy week. 

• What number moves as a result, and how long before it should move?
  Put a date against it, because a benefit with no date is just a hope in a nicer font. 

Get those three onto paper before the purchase order and the benefits review more or less writes itself. Leave them until afterward, and you are reverse-engineering a story out of whatever the system happened to log. 

The mistake I see most often 

Measuring at the wrong level. A site proves the thing works, the number is real, and then the group case gets built on an assumption that it repeats twenty times over. It does not, because site two counts downtime differently and site three sits under a different maintenance contract. 

The Lighthouse work is useful here in a quieter way. Around 60 percent of what those sites run is foundational, meaning the connectivity and dashboard layer. Only about a tenth sits in the genuinely autonomous bracket. Even the best in the world is mostly doing the unglamorous part well. 

There is no shame in that. Pretending otherwise in a board paper has a real cost. 

The distinction worth holding onto 

A workshop with a full set of tools is not the same as a workshop that has finished any jobs. You can photograph the tool board, but you cannot photograph the work, and the tool board is what most benefits reviews end up putting on the slide. 

The point is not that deployment numbers are worthless, because they do tell you the thing exists. They just cannot tell you it is being used, and those two facts get conflated constantly, usually by people who are not being remotely dishonest about it. 

So, try this before your next review. Take one initiative that is already live and answer the three questions retrospectively: who changed what they do, what did they stop doing, and which number moved. If you cannot answer all three, you have found something considerably more useful than a green status. 

That is an observation from the field rather than a finding, by the way. 

One question, and I would much rather have a real answer than a tidy one. What is the last operational number a digital investment actually moved for you, and how long was it before you could tell? 

Sources 

Global Lighthouse Network: Rewiring Operations for Resilience and Impact at Scale, World Economic Forum, 2026.
The generative AI impact figure and the workflow redesign finding are reported there from McKinsey’s March 2025 State of AI survey.
Brian Marsh, MX.0 Media newsletter, July 2026.