The cost of doing nothing: why AI called in twenty years of deferred decisions

The cost of doing nothing used to be a slide in a business case. It was the page you put after the price, the one that said here is what happens if you decline. This month it stopped being a slide and became a news cycle.
So here is my claim, and it is a bold one.
What we are watching is not a run of bad luck in enterprise technology. It is the bill arriving for twenty years of decisions to defer, and artificial intelligence is what brought the date forward.
The cost of doing nothing is the liability an organisation accumulates when it repeatedly declines to modernise the systems, data and processes its operations depend on. It does not appear on a balance sheet. It is not a line item anyone owns. It sits in slower release cycles, integrations held together by hand, and processes that only function because a small number of people remember how. It compounds quietly, and it stays invisible until something forces it into view.
Something has now forced it into view.
Three failures, one headline, and only one of them is this story
Between 12 and 17 November, several large systems failed in quick succession. Microsoft 365 went down across 74 countries. European supermarket refrigeration was driven into permanent defrost by people who should never have been able to reach it. American aviation came to a national stop, and more than 18,000 flights were cancelled over three days.
They have largely been reported together, as evidence that modern technology is fragile. That framing is comfortable and it is wrong, and the reason it is wrong matters.
The Microsoft failure was traced to a latent defect introduced in a prior networking gateway fix, compounded by an earlier capacity upgrade that left servers unable to cope when a further change was made. That is a change management failure on modern infrastructure at a company with effectively unlimited engineering resource. It is serious. It is not deferred modernisation.
The refrigeration attacks are a different problem again: connected operational equipment, installed across thousands of sites, that nobody was ever made responsible for defending. Also serious. Also not this.
If you put all three in the same sentence, you lose the one that actually explains something.
What an invoice from 1995 looks like
The aviation failure is the one.
A routine update was pushed to the Federal Aviation Administration's private communications network. Parts of the system were forced to reboot. Legacy analogue routers and switches, already operating with minimal memory buffers, were overwhelmed by the resulting data spike and began outputting inconsistent information. Three and a half thousand aircraft stopped moving. Close to five thousand flights were cancelled.
The detail that makes it an invoice rather than an accident: newer digital equipment, already installed at some facilities, continued working normally throughout.
The modernisation was not impossible. It was partially done. The parts that had been done held. The parts that had been deferred did not.
That is not a story about fragile technology. It is a story about which bills were paid.
The control group

Four days later, United Airlines suffered a company-wide systems failure. Several internal applications went down and the transmission of aircraft weight and balance data was interrupted, which meant crews could not obtain the performance figures they needed to depart. Aircraft returned to gates.
United Express flights departed as normal. Those regional carriers run their flight preparation and weight calculations on independent systems.
Same day. Same weather. Same regulator. Same brand on the tail. Different architectural decisions, made years earlier, by people who are mostly no longer in the building.
You rarely get a control group inside a single incident. Here you got one.
Why it kept getting deferred
For twenty years, anyone who did this work for a living found the same things. Integrations held together by hand. Data that cannot be trusted once it crosses between functions. A process that works because three people remember how. And a price for putting it right that came back in the tens of millions.
I have had that conversation more times than I can count. On one occasion, twenty five million to fix the service design integrations properly. The answer was almost always no, followed by something smaller, more visible and more presentable.
For a while I read that as a failure of nerve. It was not. It was a choice, and it was the same choice every time.
Look at the two things that were on the table. The real work has no launch date and nothing to announce. It produces no usable result for years. It is finished, if it is finished at all, under a different chief executive, a different finance director and a different set of targets. The smaller visible thing has a name, a launch, a slide and a result inside the current reporting period.
One of those can be taken to a board this year. The other gets taken to a board four years from now, by somebody who is not in the room today.

That is not irrationality. It is a precise response to how we measure the people making the decision. It is also how the liability got this large. Every year of deferral made the following quote bigger, because the estate grew more tangled and the people who understood how any of it worked kept leaving. Two thirds of legacy systems have no reliable documentation, and that number did not start at two thirds.
So the organisation bought the quick thing. Then the next quick thing. Then the one after that.
AI was not the escape. It was the audit.

Then came the biggest one of all. AI was the largest silver bullet anybody had ever been sold. Put the intelligence on top. Leave the foundations where they are. Twenty years of avoided work made irrelevant by something you could buy this quarter and announce at the next results call.
It has done the opposite.
Gartner now expects more than 70% of mainframe exit projects started in 2026 to fail to deliver their intended benefits, because buyers overestimated what generative tools could actually do. Sixty one per cent of organisations report that legacy platforms are actively blocking AI adoption, and 84% say integrating AI into their existing stack is difficult.
The mechanism is simple. AI needs data it can trust, processes it can follow, and systems it can reach. Every one of those is precisely what gets degraded when modernisation is deferred. So an AI programme does not route around the debt. It walks straight into it, and reports back.
AI did not clear the debt. It called it in. Every stalled pilot is a survey report on a foundation somebody decided not to open in 2014.
And rushed AI deployment does not leave the position unchanged. Fragile integrations and new dependencies get layered on top of the old ones, which is a second mortgage on a property that was already in arrears.
The number a board should look at
Not the outage figures. This one: only 36% of organisations say their modernisation plan is fully funded.
The plans exist. They have been written, presented and in many cases approved in principle. They are simply not paid for. Which means this is not a knowledge gap. Everybody knows. The gap is between knowing and committing, and it stays open for exactly as long as the reporting period is shorter than the work.
What is actually still open
The bill was always coming. Anybody who has spent time inside these estates knew the date was theoretical rather than optional.
What is still open is whether the next decision follows the same pattern. The obvious move for a leadership team under pressure to show AI results is another visible thing with a name and a launch date, and the last twenty years are a complete account of how that ends.
The alternative is not another enormous programme. It is finding out, precisely and quickly, where the organisation actually stands, which parts of the liability are genuinely yours rather than inherited, and what has to change before anything built on top of it will hold. That is a different proposition from a four-year bet, and it is available to any leadership team willing to look honestly at its own operation first.
The cost of doing nothing was never a hypothetical. It was only ever unbilled.
---
Sources
- [How a Router Update Crashed the FAA's Archaic Communications Network, Grounding All Flights](https://emeraldbook.org/news/nov-1326-3/)
- [United Airlines says it has resolved a 'widespread system outage', NBC News](https://www.nbcnews.com/news/us-news/united-airlines-system-outage-rcna609712)
- [System failure at United Airlines: weight calculation interruption, aviation.direct](https://aviation.direct/en/systemausfall-bei-united-airlines-hunderte-abfluege-nach-unterbrechung-bei-der-berechnung-von-flugzeuggewichten-verspaetet)
- [Microsoft says affected services fully recovered following worldwide outage, Al Jazeera](https://www.aljazeera.com/news/2026/11/13/microsoft-says-services-still-returning-amid-widespread-outage)
- [Supermarket fridges hit with cyberattacks across Europe, CNBC](https://www.cnbc.com/2026/11/03/supermarket-fridges-hack-cyberattack-europe.html)
- [Legacy IT is stalling the UK's AI ambitions: State of Technical Debt](https://natlawreview.com/press-releases/legacy-it-stalling-uks-ai-ambitions-61-organisations-say-windows-technical)
- [AI in Legacy Modernization: What Works in 2026 (Gartner forecast on mainframe exits)](https://abtosoftware.com/blog/ai-in-legacy-modernization)
- [The $1.52 Trillion Anchor: Why Technical Debt Is Quietly Killing Enterprise AI Ambitions, Digital CxO](https://digitalcxo.com/article/the-1-52-trillion-anchor-why-technical-debt-is-quietly-killing-enterprise-ai-ambitions/)



