We're the ones who have to adapt, not the software. Again.

AI in the enterprise isn't a ready-made meal. The model is an intermediary: what it really takes behind it: harness, context, and a make AND buy mix of standard and bespoke tools.

I was hoping, along with many of you, that generative AI on its own would solve quite a few of humanity's problems (😏). At least the ones you solve in the office by pressing keys. I was wrong. Obviously.

Harness, context engineering, workflows, loops, agentic environments, MCP, skills: if you've heard even half of these words in the past year, you know what I'm talking about. And if you've never heard them, don't worry: they're all different names for one thing, the scaffolding of data and software you have to build around the model so it pulls off something useful.

They're the biggest proof of how far a model, left to act on its own inside a computer, is from pulling off anything reliable. And I'm not saying it's useless (millions use a "bare" chatbot and find value in it every day) but on its own it's unreliable the moment you also decide to have it do things.

To really put these AI models to work, we have to spoon-feed them, look after them, tuck them in, put them at the center of an agentic system. And I'm afraid that soon we'll have to walk them a couple of times a day too.

The hard part hasn't disappeared: it has just moved. It ended up behind the model, where someone will have to sort it out, as always.

The model is an intermediary

Models are "a new interface between the idea and the action". But "interface", the way I mean it, stands for something precise: an intermediary. The model isn't where the substance lives; it's an enabler that 'makes things happen'. For some it's a machine that gives answers; for me, it's a system that asks interesting questions and has to help me get things done.

Generative AI models are nothing more than intermediaries between our ideas and the actions carried out in a complex computing environment. AI in this sense is an enabler. Let me try to explain with a little diagram.

Standard applications

Last year the schema they were selling us was clean:

The promise of AI agents

The model had slipped in there in the middle to act as an intelligent go-between, and that was that.

Everyday reality

The interface to AI is necessarily new: those outcome-oriented interfaces I've been talking about for over a year. And the "chaos", both before and after the model, is all the harness people talk about and everything that surrounds the model: context engineering, workflows, loops, permissions, connectors, guardrails. Extensive, and growing. Built by you or by the model's makers. It's there, in the two pockets of chaos on either side of the intermediary, that the effort is spent.

I don't know whether it's the natural human tendency toward obfuscation and complexity, or because knowing how to delegate to others has always been a craft for the few. The fact is that in putting together our outcomes, the actions, the models and the architectures of corporate information systems, we're realizing all over again that it's all damned complicated. And you can read it in the tons of posts, articles and books that explain a different thing to do every month.

The fact is that "AI will just handle it" is a phrase to treat with growing suspicion, and one that makes you smile, because making a model act inside real processes is structurally hard.

My little garden and the real systems

I speak from experience, from my own backyard, my miniMe, the individual context: my folder that knows a good chunk of my digital life and tailors the answers to me. And from a few dozen similar systems, more or less shared, that I've had the chance to see over these years.

The personal level is still easy, or rather, it's the level where I decide the effort and I pay for it. A walled garden: if it grows crooked, it grows crooked for me, and so be it.

On this, a year ago, I had a firm conviction: that it was enough to build your own bespoke tool, cut to fit your process perfectly. I even built several tools, including LocalAgentViewer, which shows me every interaction with every AI I work with, or AI, Max Viewer to extend the graphical interfaces in my agentic environment. All open source.

But building it is exactly what changed my mind. Because the real problem starts when you step outside the fence of personal activities and walk into a company. An organization that wants to keep control of what its agents do, see and touch is not a bigger garden: it's a different category of project. What works (A LOT) for you can't be extended to everyone. It would be like forcing everyone to keep their desktop as messy as your own. It's a personal workspace and everyone should build it as they see fit (within company limits, of course).

But when we talk about extending it to a work team, default choices and a binary answer to the Make OR Buy question are definitively over.

Emporium and workshop: make AND buy

You need a mix of solutions. Standard tools, the ones the market already gives you ready-made, tested, maintained by someone else, and custom tools only where the process is truly yours and no standard covers it. Custom everywhere is the engineering version of the meal you cook yourself every single night: noble, but unsustainable the moment the kitchen has to serve too many covers. And, I'm sorry, but the personal-meal world is exactly where most of the 'second brains' of recent months sit.

A year ago I said that, among the various people on a staff, you needed an agent shepherd (IT): someone to lead them, call them back, decide where to send them to graze and keep them contained. I stand by it, but it's no longer enough.

The shepherd and their colleagues need some support: a well-stocked emporium, where you buy ready-made products. And a workshop, where they or others can build the bespoke. It's not make or buy. It's make and buy.

And if you're not handy (and almost nobody is at everything) the consequence is an old one: you have to go to the craftspeople, the experts, and head back to the old workshops. Precisely the ones AI was supposed to retire, only to discover that to hire an AI in your company you'll need to hire new figures and new people.

In a company AI touches core processes, entrenched habits, patterns that work and that often aren't codified but are also, and above all, intellectual property that has to be protected. Custom tools and the experience that comes from containing the various models are what keep this crucial knowledge inside organizations.

The broken promise

And here it is, the broken promise of this wave: that it would be easy to do everything yourself by delegating any goal to AI.

It isn't. Not at home, not in the company. The easy part, the intermediary that answers, we got it. The hard part, making it do what we need and the way we want, requires navigating the two cores of chaos on the sides, the emporium to keep stocked, the workshop to keep running. And it's all still on our shoulders.

There's a shortcut, of course: surrendering to a strong lock-in with one vendor. Using the various ChatGPT Work, Claude Cowork and the like. Local Agents (IT) which in my view are the only real answer to AI in the enterprise.

Relying on a single vendor and taking whatever the house serves. But "whatever the house serves", today, has a problem: they're the features the vendor pushes onto you at their pace, not yours. You don't choose them: they arrive, and you have to manage them all the same, and there are dozens of them every week. What's more, hyper-advertised, so your own users clamor for them. And chasing what someone else decided for you becomes a job in itself. It's a legitimate option, and one I suggest for getting started. But it's the exact opposite of owning your own context and governing your own processes. And it's the reason why, sooner or later, you'll have to go to the workshop anyway.

So what?

The model remains an intermediary. Behind it there's no genius tool to buy once and be done: there's an emporium to keep stocked, a workshop to keep running, constant maintenance and capable people. Forget the ready-made meal.

On the contrary: once again it's up to us to throw ourselves in to prepare it.

Total AI-do-it-yourself was the ad. This is the real work, and it takes care. And, once again, competent people who enjoy doing it.

Enjoy Artificial Intelligence Responsibly!

Max