AI sped up development. Now the rest of the company is the bottleneck

An AI pilot that turned out too fast

On 6 October 2026, Mark Ajzenstadt, founder of Limestone Digital, described in a post on X an AI pilot at a hardware manufacturer where software is a big part of the product’s value. On the client’s side: more than a hundred engineers, a codebase dating back to 2017 and a plan to rebuild most of the system next year with the existing team.

As the author tells it, the CTO was mostly worried about the people on the product side. If AI made the Scrum teams faster, who would write their specifications? He didn’t have ten product owners, and he expected the problem to surface only once ten teams were using AI.

It surfaced in week four, with two engineers. The pilot covered one self-contained mobile app. AI agents wrote the code while engineers directed and reviewed the work. The bar was high, since the goal was production-quality software the client’s team could maintain without inspecting every line. When the code was ready, it turned out the product managers still assumed they had more time. They were planning to work through the next phase’s requirements by hand, in Word and spreadsheets.

“We could build faster than the business could decide what came next.”

Nobody suggested slowing development down. The team moved the AI agents earlier in the process instead. The company was merging three mobile apps into one, so rather than documenting everything from scratch, agents would map the features, overlaps and gaps across all three in a day or two. The business got something concrete to review, with the goal of signing off the scope within a week.

The author points to one more shift. A year earlier, the CTO had said he didn’t trust AI and wasn’t ready for it. Now he is asking how to get the rest of the organisation ready for the new pace. Ajzenstadt ends the post by saying an AI pilot should also test whether the part of the process where decisions are made can keep up.

Why the bottleneck always moves

What happened to the CTO has a name. In the 1980s, Eliyahu Goldratt described it in his book “The Goal” as the theory of constraints: the throughput of a whole process is set by its slowest step. Speeding up a step that isn’t the bottleneck doesn’t speed up the whole. It just makes the queue in front of the next step longer.

AI makes this visible faster than ever, because it can speed up individual tasks many times over. In its 2025 report, based on a survey of nearly 5,000 technology professionals, Google Cloud’s DORA team describes AI as an amplifier. It helps strong teams and magnifies the problems struggling teams already have. The authors also write that faster development can expose weaknesses further downstream, and that the biggest return comes from clear workflows and aligned teams rather than from the tools themselves.

Andrew Ng made the same point in his talk “Building Faster with AI” at Y Combinator’s AI Startup School in 2025. His teams build prototypes up to ten times faster, so deciding what to build increasingly takes the longest. He even mentioned a team that proposed having more product managers than engineers.

The numbers in Figma’s AI report (8,403 survey responses and 639 interviews over three years) point the same way. 90% of respondents say design is at least as important as it was before AI, and 65% of developers say it matters more today. So when one team gets AI, it pays to ask straight away who will be the next bottleneck.

Where your company will get stuck

In the story above, the queue formed in front of product requirements. Companies that build digital products and run campaigns have more of these points, and any of them can hold up the team that got AI first.

Developers working with agents finish tasks faster than the company can decide what comes next. Faster development also needs complete designs sooner, and when it gets incomplete mockups, questions and post-release fixes eat the time AI was supposed to save. Marketing writes campaign copy with AI in minutes, yet every new format, size and language version still waits in the designers’ queue. Sales and HR have a different problem. Decks and job ads made with AI on their own are quick to produce, but they don’t look like your brand, so consistency becomes the bottleneck.

The design process sits in the middle of this chain. It takes decisions from the business and hands finished assets to development, marketing and sales, which is why speed gained on one side turns into a queue on the other faster there than anywhere else.

The lesson from the pilot: bring AI in earlier

What interests us most is what the team did once it saw the problem. It didn’t tell developers to slow down or add people to write documentation. It moved AI to the step that was holding everything up and used it to prepare material for decisions. People still made the decisions, only faster, because they had something to assess.

In the design process this looks very similar. AI can prepare campaign adaptations for new formats and languages before a designer sits down to work. It can add the descriptions of states and components that developers need, check design and code against the design system, or describe the brand so generative tools reproduce it faithfully. The decision stays with people every time; AI shortens the way to it.

How to check whether your process can keep up with AI

Before you roll an AI pilot out to more teams, ask yourself a few questions:

1. Where does work wait today? Look for “almost done” tasks sitting in front of the next step.
2. How long does a decision take, from idea to approved spec and from design to sign-off?
3. Do developers get complete designs, or do they come back with questions?
4. How long does a campaign take to prepare across formats and languages?
5. Do AI-made materials created outside the design team look like your brand?
6. What does the pilot actually measure? If only one team’s speed, you won’t see where the rest gets stuck.

If the answer to any of these is “it depends” or “we don’t know”, that’s where to start.

Where we start at CAASUAL

For the same reason, our AI design workflows service starts with a process map, not with choosing tools. First we look at the stage where projects stall and between which teams. Then we show where AI will give your teams time back fastest and what’s better left unautomated. We pick the tools last and build them into the daily work of marketing, sales and development, not just the design team.