The quiet software

A prediction I would like to put on the record, so I can be wrong about it in public.

The most economically valuable AI work over the next decade will not happen in the products that most people associate with AI. It will not be in chat. It will not be in coding assistants. It will not be in image generators or video tools or any of the consumer-facing products that have absorbed most of the attention since 2023.

It will be in software that no one outside the company that built it will ever see.

I think of this as the quiet software. It is internal. It is unglamorous. It is specific to the operating model of one particular company. It is built by a small team, often with one foot in the business, and it sits between two or three systems that were never meant to talk to each other. It uses AI not as a feature but as a load-bearing component, the way a SaaS application from the 2010s used a database. The interface is rarely a chat box. It is usually a form, a queue, a dashboard, or a notification.

The reason I think the quiet software is where the value will accumulate is not because the noisy software is bad. The noisy software is great. It is the reason any of this is possible. But the noisy software, by its nature, is horizontal. Anyone can buy it. Your competitor has it too. It cannot, by construction, give you durable advantage. It gives everyone the same advantage, which is no advantage.

Durable advantage in this transition will come from operating-model-specific software that other companies cannot copy because they would have to copy your operating model first. The closer the software sits to the actual decisions a company makes, the more it has to know about how that company specifically works, and the harder it is to replicate.

I will give an example.

A specialty distributor I worked with last year had a routing problem. Inbound orders had to be assigned to one of fourteen warehouses based on a half-codified set of rules involving inventory, geography, customer tier, contract terms, and a handful of edge cases the warehouse managers carried in their heads. The decision was made twelve thousand times a week. It was being made by four people, inconsistently, against a partial spreadsheet that had not been updated in a year.

The horizontal AI option was to put a chat interface on top of the spreadsheet. We did not do that. The chat interface would have been an academic toy. The decision needed to be made in milliseconds, against live inventory, against live contract terms, with the edge cases encoded.

What we built instead was a small piece of internal software with one screen. It looked like a 1990s order management system. It received an inbound order, ran it against a model trained on the historical decisions, scored each warehouse for fit, and surfaced the top three with reasons. A human could click through to override. The override fed back into the training data the next week.

This software is, by any aesthetic standard, ugly. No one would write a TechCrunch article about it. It does not appear in any vendor's slide deck. It is also, as it turns out, the single most valuable piece of software the company owns, because it has compounded six months of operating advantage on a competitor that bought a horizontal AI tool and is trying to make it work against a similar problem.

This is the shape I think most of the next decade looks like. Internal. Specific. Unglamorous. Built around a load-bearing decision that the company already made twelve thousand times a week before AI was involved, made now twelve thousand times a week with AI as a component, and made better by some margin that compounds. The software is invisible from outside. The advantage shows up two years later in unit economics that everyone else cannot match.

There are a few implications of this view that I think are worth stating clearly.

The first implication is that buying AI tools is not the same as building AI software, and the two strategies will diverge in outcome. The tools are commodities. The software is operating-model-specific and durable. Companies that organize their AI strategy around tool procurement will find themselves with the same tools as their competitors. Companies that organize it around building the quiet software will find themselves with capabilities their competitors cannot replicate.

The second implication is that the right team to build the quiet software is rarely an AI team. It is a small group of generalists, ideally with one person who deeply understands the operating model and one person who deeply understands what models can do. A pure AI team will produce something technically interesting. A pure operations team will produce something that does not use AI well. The quiet software comes out of the intersection.

The third implication is that the right vendor model for this work is consulting more than software. Not management consulting. The kind of consulting where someone who understands models pairs with someone inside the company who understands the decision, and they ship a piece of internal software together over three to six months. Then they hand it off. Then they come back in a year to redesign it, because the model landscape has shifted under it.

I am willing to be wrong about all of this. Maybe the foundation models advance fast enough that the horizontal tools become so good that the quiet software is unnecessary. Maybe agent frameworks mature into a substrate that means most companies do not need to build anything. Both are possible. I have noticed that in every prior generation of enterprise software, the same prediction was made about whatever the latest horizontal layer was, and in every case the actual durable value ended up in the specific applications built on top of it. I expect the same here. I have not yet seen evidence that this generation breaks the pattern.

If I am right, the consequence is this. Most of the AI activity that gets press in 2026 is the noisy software. Most of the AI activity that creates durable enterprise value will, by 2030, be the quiet software. The two will not be reported on at the same volume. The one that gets reported on is the one that is easy to see.

The leaders who can resist the gravity of the noisy software, and instead invest patiently in the quiet kind, will compound. The leaders who chase the headline AI tools as a way to look modern will find themselves, in three years, with a stack that looks the same as their competitors and a P&L that looks the same too.