Where is the value?

This could trigger some people. What if there is no value to adopting AI? Now, for the engineering folks, you might want to go read something else, but I like to push back on ideas and really think about things differently, just for the hell of it.

So let’s get started.

We discuss it constantly: the bottleneck has shifted. Code is cheap now, so the constraint has moved to review, or to verification, or to who can take part. I’ve said versions of it myself, and for a lot of software teams I still think it’s true.

But has it? The bottleneck I mean; I’ve been turning that over for a while, and I’m not sure the question even makes sense for a lot of the organisations adopting AI right now.

Was there ever a bottleneck?

A bottleneck shifting assumes a line. Ideas come in one end, software goes out the other, and somewhere in the middle the work piles up. For a long time that pile was engineering capacity; there were always more things to build than people to build them, so the backlog grew and the roadmap was a negotiation over what got left out. AI loosened that, and the pile moved further down the line.

Right?

That’s a fair story if you had the line. And because so much of the conversation happens among people who did, we like to assume that’s how companies produced value: a somewhat relevant forward setup of SaaS, with product teams, sprints, a CI pipeline and a backlog longer than anyone can build.

Not everyone came from that or even has a fully functional model of it. Plenty of organisations never had a delivery pipeline with a queue to clear. Their value came from somewhere else entirely: relationships, operations, a service done well by people who know the work. The software they ran was bought, configured once and left alone. For them, engineering was never the thing holding them back, because engineering was never the thing they did.

So when we tell them the bottleneck has shifted, what are they meant to hear? There’s nothing to shift from.

Starting from somewhere else

So who are we talking about? Picture a business that runs on spreadsheets, email and a phone. It does well. Customers stay for years. The real system is one person who has been there twenty years and knows where everything is, plus a shared drive that only makes sense to her. Ask where the backlog is and you get a blank look, because nobody has ever written down a list of software they wished they had. They just worked around it.

Is that business behind? I don’t think so. Its value is the customers who stay and the twenty years in her head. A faster way to ship could still be a prize, and there’s no discounting the speed of AI. But will it really change anything for her customers?

Not having a pipeline might even help. Think of the countries that skipped landlines and went straight to mobile. They didn’t upgrade a phone network, because there wasn’t one to upgrade. They started somewhere else, and in a few places ended up ahead of the countries that had to unpick the old network first.

I think some organisations are in that position with AI. They don’t have a decade of SDLC to rework. They’re starting fresh. Is that good or bad? Honestly, I don’t know.

That doesn’t stop the pressure, though. There’s a fear of irrelevance, that if you don’t adopt AI now you’ll be the business that got left behind. So it gets pushed into things that would eventually benefit from it but don’t need it, because being seen to adopt feels safer than asking whether you should. And that feeds the beast. Every rushed rollout becomes another story of someone who adopted, and the pressure on everyone else goes up.

And some take that fear to the extreme. In April 2026, Allbirds, the shoe company, announced it was selling its footwear business to become an AI compute provider, renting out GPUs. By June it had a new name, Smartbird, and a new CEO. The shares jumped on the news, then gave some of it back. I don’t know if it’ll work. But a shoe company deciding the way to stay relevant is to stop making shoes is about as clear a picture of that fear as you’ll get.

And even when it isn’t that extreme, adopting comes with a bill. Back in the business that runs on spreadsheets, the spreadsheet gets made faster, but now someone has to keep the plugins and skills up to date so the agent doesn’t start making things up. That’s close to a full-time engineer, in a business that never had one. Helpful today, maybe not so much tomorrow.

So what are they doing with AI?

The organisations starting fresh are adopting AI too, but not all of it is rushed. Some of the most interesting adoption is happening outside the typical companies, the ones that don’t look like a software business and never wanted to.

And I don’t think they’re going from what they were before to a faster version of it. They’re grappling with new ways of working right at the start of AI, and a lot of it is about finding ways to be relevant (or so adoption asks of them). Doing something they couldn’t do before, offering something they couldn’t offer, rather than doing the old thing quicker.

What does that look like?

1. Knowledge out of one head

   Who: The person who knows

   The person who knows where everything is writes it down for the first time, as a skill an agent can use, so the knowledge stops living in one head.

2. A tool nobody would have hired for

   Who: A small team

   A small team builds a tool they could never have justified hiring developers for.

3. Help at two in the morning

   Who: The service team

   A service business turns what it knows into something a customer can use at two in the morning, without anyone picking up the phone.

Essentially the path to the second user.

It’s happening in research too. When Anthropic launched Claude for Life Sciences, one of the examples was Paper2Agent from James Zou at Stanford, which turns research papers into interactive AI agents. A paper stops being something you read and becomes something you can work with. Nobody was going to get that from a faster pipeline.

You won’t find any of those on a chart of where the bottleneck went. They’re new, and they were made by the people who own the problem.

The trap is copying the playbook

Which is why it worries me when the SaaS playbook gets handed over wholesale. Here’s how you set up an SDLC. Here’s how you measure cycle time. Here’s where the bottleneck will move to next. It’s good advice for a software team, and I’ve written plenty of it. But for an organisation that never had a line, it’s advice about a problem they don’t have.

And I think there’s a bigger ask hiding in that playbook. Companies that never had engineering are being asked to become SaaS companies, just because AI turned up. Build your own tools, run your own agents, ship something a customer logs into. Nobody stopped to ask whether that’s the kind of company they want to be.

Worse, it can take away the thing they’ve got going for them. In a software company, the people who own the problem and the people who build the solution are usually different people, and closing that gap is hard. In a lot of these organisations, there is no gap. The person with the problem is the one sitting with the agent. Wrap that in a process built for handoffs between product and engineering and you’ve invented a handoff that wasn’t there.

They still need some process, though. It just has to start from what they actually do.

Where is the value, then?

My opinion? It’s not faster that’s the success. It’s what we do with the technology that’s not been done before, and honestly, that’s the exciting bit.

The bottleneck framing hides that. It treats AI as an accelerant on an existing line, and measures success by how much faster the line runs. And I don’t think that only goes for the organisations that never had a line. A software team that ships the same roadmap twice as fast has still just done the old thing quicker. That’s useful, but it’s not why I get excited about this technology, and I doubt it’s why their customers will.

So speed isn’t the metric. Did this let us do something we couldn’t do before? Did a customer get something they couldn’t get from us last year? Is the business still needed in a world where anyone can ask an AI the question we used to answer?

Those are much harder to measure, which is probably why we reach for the bottleneck story instead. It has a queue you can count. Answering them needs someone to say what the value is meant to be before the work starts, which is the outcome problem again. Without it, you end up measuring the thing that’s easiest to see.

Final thought

I think the bottleneck story is comfortable for people like me, because it keeps the pipeline at the centre of everything. If the constraint moved, the pipeline still matters, and so does everyone who knows how to run one.

The organisations with no pipeline don’t need that story. Some of them are adopting out of fear, but some are already asking where the value is, and they’re answering it with whoever is closest to the work. Has the bottleneck shifted? For a lot of software teams, yes. But for these organisations it was never there.

So is there no value to adopting AI? If faster is all you’re after, I’m not sure there’s much, because faster was never going to be the success. The ones worth watching aren’t the teams shipping faster. They’re the ones doing something with this technology that they couldn’t do before, and a lot of them weren’t shipping at all until now. That’s where the value is, and it’s the most exciting part of all of this.