LLMs and potatoes: stop admiring the flowers

From a commercial point of view, software is not primarily code. It is a scalable, repeatable mechanism for delivering value at near-zero marginal cost.

Customers do not buy code. They buy:

• Time saved
• Revenue increased
• Complexity abstracted
• Compliance ensured
• Insight generated

Good software turns raw transactional data into actionable decisions. It is table stakes for a bookkeeping solution to provide an overview of accounts. It becomes valuable when it starts suggesting, for example, that increasing spending in a certain category could maximise tax benefits.

In that sense, software is leverage.

• Operational leverage: A business can serve more customers
• Labour leverage: Human work is reduced
• Cognitive leverage: Expertise is embedded directly into the system

Every feature built into a software product should increase one or more of these forms of leverage, at minimal cost.

You cannot eliminate complexity, only shift it

The world is complex. As every experienced software engineer knows, complexity cannot be eliminated; it can only be shifted.

The question is: where should it live?

Instead of placing the mental burden of remembering procedures on the user, software can encode workflows that quietly absorb that burden. A well-designed system compensates for distraction, stress, or lack of expertise. It covers the angry customer calling at the worst possible moment. It ensures that the right steps happen in the right order, even when attention is fragmented.

Chat-based systems are a step back for most of this. To use them effectively, the user has to:

• Know that a problem exists
• Formulate the right question
• Ask the system effectively

This puts cognitive load on the user. It also assumes that the user has time and capacity for this sort of interaction during a regular workday.

Chat with me, maybe?

Many software companies are currently racing to integrate large language models into their products. The focus is on a particular interaction style - chat - rather than asking a more fundamental question:

What form of leverage is actually being increased?

When the potato was introduced to Europe around 1570, it was initially grown as an ornamental plant. People simply did not know what to do with it. Being part of the nightshade family, improper consumption was toxic. It took time before people realised that the true value of the potato was not in its flowers, but in its tubers.

LLM chatbots today are in a similar phase. They are often more ornamental than useful, and more effective behind the scenes than in the spotlight. In the worst case, they increase operational costs while adding little real leverage. Just because they are text-based does not mean text is the most valuable interaction pattern through which to leverage them.

The true leverage of LLMs

The true value of LLMs is not in exposing them directly to users. It is in using them behind the scenes.

LLMs should be:

• Given clear instructions and goals by domain experts
• Embedded into workflows
• Tasked with analysing data continuously
• Evaluated against a usefulness threshold before surfacing results

Only once an insight has been identified, verified, and deemed valuable should it be presented to the user.

At that point, chat may become useful for inspecting the reasoning or exploring alternatives. But the ultimate leverage is a proactive suggestion that can simply be applied, without requiring the user to construct it from scratch.

Toward proactive software

Chat should not be seen as the end goal. It is, at best, an intermediate step.

The real shift is toward software that is no longer passive - sitting idly on a cloud server waiting for user input. Instead, it actively analyses data, identifies risks, spots opportunities, and proposes improvements. Some of this is classic data analytics. Some of it can be augmented with LLMs.

If that is the future, then debates about chat windows and UI consistency are largely secondary.

The more important question is this:

How should users interact with software that thinks ahead?