From AI exploration to AI efficiency

Table of Contents

For the last 3 years, I have been deep in AI exploration. Not in the casual "I tried a few tools" sense, but properly in it: models, agents, coding workflows, subscriptions, prompt techniques, autonomous setups, and all the messy operational details that come with trying to make these systems useful.

It gave me a clear sense of what AI is actually good for in a work context, what is mostly theatre, and where the real bottlenecks are. I come out of it feeling less like an enthusiast and more like an operator. More specifically, deciding where AI belongs in a workflow, understanding the economics behind it, and being honest about whether a tool creates value or just activity.

The shift

That matters more by the week, because the conversation is changing. The question is no longer just what AI can do. The harder and more useful question is what it should do, under what constraints, and at what cost. Time saved matters. Cost matters. Attention matters. The operating model matters.

I recently moderated a roundtable at a finance event about AI spend and the need to monitor subscriptions, model usage, and token consumption more carefully. It is one thing to talk about that from the outside. It is another to feel the constraint hit your own workflow in real time.

This weekend, that shift stopped being theoretical for me.

Anthropic shut down the use of Claude Max subscriptions for OpenClaw. It was always something of a grey zone, so I cannot say I was surprised. But it changed the economics immediately. Once that usage goes to API billing, even trivial interactions become very expensive.

In some ways, this move simply accelerated a shift I had already started.

What exploration taught me

My exploration phase taught me a lot. I gradually pushed Lovable up to the $400 per month plan while exploring. Next month it will be down to the minimum $5 tier. At the same time, I was running OpenClaw with Claude Code Max, and at the peak I even took a second Max subscription because I did not want to wait a few days through cooldowns. Those are just the big spenders. There are plenty of other tools and subscriptions.

The same thing happened with agents. I built a small fellowship of OpenClaw specialists, each with a role. That was useful because it taught me what delegation should look like, and also where delegation is unnecessary.

Last month, when Lovable was free without limits, I even had a couple of OpenClaw agents using my Lovable account non-stop, creating whatever ideas they could think of. The Lovable tokens were free, Claude was a fixed subscription, but the mental load to follow up was definitely not free.

The lesson: a system can be productive and still be wasteful.

What I am cutting

That is why I am now becoming much more ruthless about cutting down rather than adding. I want fewer tools, fewer overlapping subscriptions, fewer unnecessary workflows, and tighter control over what actually stays in the stack. The bar is simple: does this save meaningful time, improve quality, or create clear value? If not, it probably should not survive.

I do not need proactive agents with multiple heartbeats per day for everything. Some work is better handled reactively, some is better done ad hoc, and some is simply faster to do directly with the right tool. Exploration gave me breadth. Efficiency now demands selectivity.

What is coming

While one door closed, others opened. Google released Gemma 4, a free local model that could change everything, again. Meanwhile, I have gone back to using OpenAI's models and the Codex app, both for OpenClaw and for coding more generally.

Claude Code has been excellent, but compared to Codex it's starting to feel a lot like a personality hire.

Even the prompting culture is changing. One of the funniest examples I have seen recently is "caveman prompting" — intentionally stripping out all the polite, flowery language and reducing prompts and outputs to the bare functional core to save tokens.

A normal AI answer might say: "Review which overlapping tools save time and cancel the rest."

Caveman AI says: "Too many same tool. Keep ones save time. Cut rest."

It is absurd on the surface, but it is also a very honest signal of where the market is heading. People are starting to care less about the pleasantness of the interaction and more about efficiency, cost, and throughput.

What this phase requires

We are moving out of the phase where AI was mostly about abundance and possibility, and into one where it has to justify itself operationally.

The people and teams who will benefit most are not necessarily the ones trying the most tools. They are the ones building the clearest systems: choosing the right models, measuring spend, reducing waste, and keeping only what genuinely helps them do better work.

That is where I am now. Still very deep in AI, but less interested in novelty for its own sake. More interested in leverage, reliability, and discipline. Exploration was necessary. It made me fluent. But fluency is only useful if you can turn it into judgement.

You could say I have moved from experimenting AI tools to managing an AI operating model.

And yes, in the process, I may have fired the personality hire.