AI Agents Are the New Z Programs

Last Thursday I stayed up late to go see Terminator 2 at the cinema. Well, the screening was at 21:10 and for me that already counts as staying up late.

They re-released it in 3D for the 35th anniversary, right around August 29, which is Judgment Day in the film. I'm sure I've seen it at least 10 times in my life, but never in a theater, because when it came out in '91 I had neither the age nor the chance. And the room was full. I wasn't the only lunatic there.

Beyond the 3D effects Cameron added, what struck me was seeing a room full of people who knew exactly how it ends. Nobody moved the same, and we all clapped with a bit of sadness when the T-800 sinks into the lava.

Cameron used the anniversary to drop a line that stayed with me: after thirty-five years you're allowed to spoil it, he said, so spoiler alert, the good guys beat the superintelligence.

And walking out of the cinema it made me laugh to connect it to another film I've already seen. That one wasn't directed by anybody, has no special effects, and I've seen it at almost every company I've worked at.

This week SAP published the report that describes it from start to finish, without realizing it.

The numbers everyone will quote

As is now routine, SAP commissioned Oxford Economics for its second report on the value of AI. They surveyed 2,600 companies in thirteen countries, all with more than five hundred employees, with responses from director level upward.

Skimming it, the three data points that caught my attention were these:

1. 30% of the tasks at an average company are already supported by AI, and they expect to reach 48% within two years.
2. The return went up: from 4.7 million dollars last year to 6.3 million this year, and they project 15.9 million two years out.
3. 69% are satisfied with what they're getting.

Those are the numbers you'll see on LinkedIn this week, with a nice AI-generated chart on top.

I'll give you some others, which probably won't make those pictures.

The numbers nobody will quote

Look at this picture of companies deploying AI agents.

44% have a registry of agents. Which means 56% have no list of what agents they're running.

32% have cost or usage controls on those agents.

A third have engineering guardrails for AI-generated code.

12% consider themselves prepared to govern AI effectively. Only 3% feel completely prepared to deploy and scale agents.

And the one that made me put down my coffee and read it again:

Companies admit that different teams are building overlapping agents, for the same needs.

On top of that, half don't even have a shared definition of what an agent is and what it can do.

That's where I recognized the film. And probably now you do too.

I've seen this one already

For anyone not from the SAP world: a Z program is custom code.

The Z comes from custom objects starting with that letter, to tell them apart from the standard. Every company has its own, written over decades to solve something the base system didn't do.

Now, what happens with those programs.

They get written fast, for a specific need, and there they stay. Whoever wrote them changes project or leaves the company. Nobody decommissions them, because nobody knows for certain who uses them. Three different departments end up with three reports that do almost the same thing, each convinced theirs is the good one. And the real cost of all that doesn't show up on any spreadsheet until much later.

When does it show up? When it's time to migrate. Or when Clean Core arrives.

Anyone who's been on an S/4 migration project knows the moment. You run the custom code analysis and the result is always similar: a huge share of the objects hasn't executed in years, and nobody deletes them because nobody can state what they do. The company ends up paying to analyze, remediate or carry code that's no longer useful to anyone.

But go back and read the report's numbers. No registry. No clear owner. No cost control. With teams duplicating effort on the same thing. And without a shared definition of what the thing they're building even is.

The agent film is the same one as the Z programs.

The difference is the speed: a Z program required somebody to write it, test it and transport it. An agent can be stood up by anyone in an afternoon, and now there are platforms so that people who don't code can do it.

The ecosystem spent two decades learning, at a cost of millions, that code without governance turns into debt. The lesson is still warm and we're already repeating it, faster and with less control than we had in 2005.

The most honest thing in the report is what cuts against whoever paid for it

There's something here I want to say carefully, because it isn't a swipe at SAP.

SAP commissioned this report. And even so they published that only 3% of companies feel prepared to scale agents, that half don't share a definition of what an agent is, and that 56% keep no registry of the ones they're running.

Then they close with their Chief AI Strategy Officer saying that connecting AI to data and processes with context and governance is "what we call the Autonomous Enterprise".

So the same document selling the autonomous enterprise contains the evidence that almost nobody is in a position to sustain one. At least that's what you get if you look at the numbers past the headline. I don't think it's cynicism. I think it's what happens when you publish real data and then tie a commercial bow on top. What's striking is that almost nobody will read past the ROI.

Before you agree with me

Three things that cut against what I just said.

1. On the report: it's a vendor-commissioned study, with return figures self-reported by directors. Nobody audits those numbers, and if we get picky about the sample you can distrust even that. And there's a fairly well-known bias here: people tend to report good returns on things they defended internally to get the budget. I'd handle the ROI with tongs. The governance data strikes me as far more credible, precisely because it cuts against whoever paid for the study.
2. On me: I am not neutral. I work in enterprise technology, I push AI adoption in my area, and when I say this needs governance I'm describing work that people like me get paid to do. Keep that in mind when you read me.
3. The one that makes me most uncomfortable: governance also gets used as an excuse to do nothing. I've already written about companies where approving a twenty-dollar license takes three months. I'm not asking for that. An agent registry and a cost control don't slow anyone down, it isn't rocket science.

Back to Terminator

In the film, Judgment Day has an exact date: August 29, 1997. When that date passed without an apocalypse, the sequels moved it. And moved it again. And again.

The same thing is happening with AGI, except the dates arrange themselves around the IPOs coming up: always close enough that you have to pay attention now, and far enough away that nobody can hold you to them when it doesn't happen. I've already written about who benefits from those dates existing.

But here's the point, and it's where the metaphor breaks in our favor.

Your company's judgment day won't be a date on a calendar or a machine becoming conscious. We're far from that, even if two weeks ago a humanoid robot ran a hundred meters in 9.39 seconds in Beijing, against Usain Bolt's 9.58 in Berlin, and even if back in April another one won that city's half marathon faster than the human world record. Which is impressive, and which still doesn't mean the machine knows what to do with your financial close.

It'll be any given Tuesday, like today, when somebody runs the equivalent of a custom code analysis over the agents you have running and discovers nobody knows what half of them do, who approved them, or what they're costing.

That isn't decided by any superintelligence. It's decided by whether this week somebody opens a spreadsheet and starts writing down which agents exist, who asked for them and what they can touch.

Cameron was right about the spoiler. The good guys beat the superintelligence. What he didn't say is that they beat it by doing the boring work.

Pablo / En Producción

P.S. If someone at your company is building agents right now, ask them where the list is. If there's no list, there's your task for the week, and it needs no budget.

References:

• SAP and Oxford Economics. "The Value of AI", second annual report, 2026.