Eighteen Megabytes

There was a file. It was an Excel model, it was about eighteen megabytes, and it produced the rent and renovation report for a US residential property portfolio worth a couple of hundred million dollars.

Opening it took a while. Recalculating took longer. One person understood it properly and a second person understood it well enough to be dangerous. It was updated monthly, and the update was a full working day that someone senior lost every month, plus the day after that, when the errors surfaced.

We automated it. The saving was low six figures a year.

I've been thinking about that file again this year, because the enterprise AI conversation has developed a blind spot in exactly its shape.

The thing nobody puts a number on

Here's the shape of most AI business cases I see in 2026. There's a technology. There's a workflow. There's an estimate of time saved, usually expressed as a percentage, usually derived from a vendor's benchmark. Multiply the percentage by a loaded salary cost, annualise it, and you have your return.

The trouble is that the eighteen-megabyte file doesn't appear anywhere in that calculation, and the eighteen-megabyte file is where the money is.

What that spreadsheet actually cost wasn't the day a month. It was that decisions waited for it. A property that should have been sold in March got decided in May, because the numbers that would have triggered the decision arrived on a monthly cycle and the person who could interpret them was on leave. The cost of the file was the cost of every decision it delayed, and that number appears on no line of any P&L.

We measured the wrong thing for years because the wrong thing was the only thing that was measurable.

This is, I think, the single largest reason why enterprise AI returns look so thin in the aggregate data. McKinsey's survey this year found 89% of organisations using AI and 37% able to attribute any EBIT impact — a figure that's been flat for twelve months. NBER surveyed around 750 CFOs and found self-reported productivity gains of 3.0% expected for 2026 against revenue-based gains of 1.8% — running about seventy percent ahead of what shows up in the revenue line.

Both of those gaps have the same explanation available. The savings are real and they're landing somewhere that isn't measured.

What actually moved the number

When I look back at the work that produced defensible savings in that business, the automation wasn't the interesting part.

The first was putting the decision inputs in one place. We built a dashboard that showed ROI, IRR and gross margin for a property in a single click. That halved the time to a buy, sell or renovate decision. Not because the calculation got faster — the calculation was never the slow part. Because the assembly of the inputs stopped requiring a person to go and find them.

The second was removing hand-offs. Work moved between acquisition, construction, operations, renovation, leasing, property management, accounting and legal teams across four countries. Every hand-off was a queue, and every queue was a place where something waited for someone. Restructuring the assignment and tracking so work routed itself cut hand-off time by about 40%.

The third was renovation tracking, where the manual effort went from days to minutes and the annual saving ran to mid six figures.

None of those made anybody better at their job. Each one removed a reason to wait.

That's what these technologies actually do, in my experience, and it's why the time-saved business case underestimates them so badly. You can count an hour. You cannot count a queue that stopped forming.

Why I'm telling you about 2016 in 2026

Because the mechanism hasn't changed and the mistakes haven't either.

The organisations getting nothing from AI right now are, overwhelmingly, the ones that bought a capability and attached it to an existing process. The model drafts the email that a person was drafting. The agent summarises the document that a person was summarising. The work is faster. The decision still waits for the monthly cycle, the approval still sits in the same queue, the hand-off still happens at the same boundary. Nothing structural moved, so nothing shows up in the accounts.

McKinsey's own read is consistent with this. Nearly three-quarters of the organisations they classify as high performers report fundamentally redesigning workflows, against a quarter of everyone else. They're careful to present this as a distinguishing practice rather than proof of causation, and I'll be equally careful. But it matches what I've watched from inside, across four companies and two decades.

The redesign is the work. The technology is the occasion for it.

The practical version

If you're building a business case for AI right now, replace the time-saved calculation with these. They're harder to answer and they're worth more.

What decisions in this business currently wait for information, and how long do they wait? Not how long the report takes to produce. How long the decision sits.

Where does work change hands, and what happens to it in between? Every boundary between two teams is a queue. Price the queue, not the task.

What exists only because assembling information used to require a person? The eighteen-megabyte file existed because somebody had to pull numbers from six systems and reconcile them. The file was a monument to a constraint. Once the constraint was gone, the file was just an expensive habit.

If you run those three questions honestly, you'll usually find the return is larger than the time-saved case suggested and located somewhere entirely different.

The part that should worry you

There's a reason organisations keep reaching for the time-saved calculation despite it being the weakest available. It's the only one that survives contact with a finance function.

"We'll save 400 hours a year" is auditable. "Decisions that currently take nine weeks will take two" is not, until after you've done it. And the second claim is worth ten times the first.

So the instrument selects for the wrong project. Not deliberately, and not because anyone is being foolish. Because one number can be defended in a room and the other can't, and the work gets shaped around what can be defended.

I don't have a clean fix. What I'd say is that the eighteen-megabyte file is still out there in most organisations, wearing a different costume. It's a weekly reconciliation, a monthly pack, a handover checklist, a report somebody builds because a decision waits for it. Go and find yours. Then ask what it's really costing — not the day, the delay.

That question has never once given me the answer I expected.