“The more advanced a control system is, the more crucial may be the contribution of the human operator.”
— Lisanne Bainbridge, Ironies of Automation

‍

‍

Last summer, I decided to install a rainwater diverter on my drain pipe. How hard could it be? Cut the downspout, insert the diverter, and connect the pipe to the reservoir—done in an hour, tops.

‍

Reality proved different. Our downspout is unusually large, square, and made of metal. Some screws holding the brackets were stripped. Sourcing the right diverter took a week, removing the stripped screws required more time (and pliers) and cutting the pipe with a hacksaw took an hour. The actual installation took only 20 minutes, but these complications turned a one-hour job into two frustrating weeks.

‍

Naturally, I thought of legacy modernization. Essential complexity—taking systems written in an older language, interacting with old data sources, moving them to a more modern language and so on—is challenging but describable. It’s installing the diverter and making it work.

‍

Incidental complexity stems from unforeseen circumstances peripherally related to the core challenge, like stripped screws. They could be security constraints, unclear dependencies, or obscure legacy tooling that only retirees know. Or they might be organizational inertia, grumpy DBAs, compliance hurdles. And because I love the Oxford comma, these are hard to articulate, predict, and measure, yet they can introduce significant delays, confusion, and risk.

‍

Most AI efforts around essential complexity focus on transpilation: use AI to translate from COBOL to Java. Super fast, right? But this doesn’t solve the long pole in the tent: what slows everything down is the testing and verification phase to make sure everything works okay.

‍

In fact, letting AI translate code from COBOL to Java exposes projects to AI’s imaginative, free-wheeling non-determinism. AI can’t grasp the nuances of business logic or maintaining context, especially with larger chunks of code. It can’t deal with having two implementations coexist for a while to determine which is more promising (something juvenile humans do naturally when they have two parents to play off of one another).

‍

What about incidental complexity? LLMs can certainly help abstract legacy systems so you can uncover hidden context and dependencies, or summarize what a certain chunk of code does. But they’re terrible at actually solving issues. In contrast, humans are ace at managing incidental complexity. We can fix a spaceship while hurtling through the vacuum of space with little more than a lithium hydroxide canister, a roll of duct tape, a few plastic bags, a cue card, a hose from an EVA suit, and a towel.

‍

A good craftsman knows how to use a tool; a great one knows when not to. Throwing AI into modernization without a solid sense of what it can and can’t do might make things worse, not better. And there’s still no substitute for actual humans solving hard problems with experience, judgment, and maybe some duct tape and a towel.

‍

News and Views

We know who could use a few more actual humans and less AI: King Features, a Hearst syndicate that churns out content. They created a Summer Reading List that features—yes, you guessed it—books that don’t exist, which was then republished in The Chicago Sun-Times and The Philadelphia Inquirer. Double-oops: The Sun-Times let go of 20 percent of its staff last year.

‍

The Economist writes entertainingly, at least for Premier League football fans, about Why so many IT projects go so horribly wrong (subscription required). Although we take some issue with one of their recommendations (a free Imogen tote bag for the first one to correctly identify which one), we do agree that abstraction and fear of failure remain strong culprits.

‍

Move fast and break things or move fast and make things? We prefer the latter. WSJ reports on the latest upheavals on the IT estate of the US federal government.

‍

Federal News Network recently interviewed MO's Mark Forsthoffer about how incremental modernization enables real progress inside high-stakes systems.

‍

From the Orchard

We’ve kicked off a joint blog series with Thoughtworks! The first one from Shodhan Sheth and me discusses how our two organizations are approaching modernization in a way that balances speed with control—and why that balance is what makes it safer. Our second will talk about rethinking the “R”s of modernization.

‍

Our CEO Rob Mee and Thoughtworks’ CTO Rachel Laycock discussed how speed, quality, and accountability matter most in modernization, and where Imogen fits in on the Thoughtworks’ Technology Podcast.

‍

In our latest episode of On a Limb, author and interdisciplinary thinker Samuel Arbesman discusses his new book, The Magic of Code, and what it means to bring wonder, history, and responsibility back into our relationship with software.

‍

Heading to the Gartner Applications Innovation and Business Solutions Summit in Las Vegas? Come find us at Booth #121 and catch our CCO, Edward Hieatt, on June 4th at 12:50 PM PDT, where he’ll be talking about how to change the risk equation around legacy modernization.

‍

Curious to learn more? Say hello@mechanical-orchard.com.

‍

——————

Issue first published on May 28th, 2025.

‍