Building is cheap now. Judgement isn't.

For most of the history of software, the hard part was the building. An idea was cheap; turning it into something people could use took a team, a budget and the better part of a year. That gap shaped everything, from how startups raised money to how big companies decided what not to try.

That gap is closing fast. A working prototype that once took a quarter can now take a week. Interfaces, back ends, integrations and even first drafts of copy arrive in hours. If you've built anything this year, you've felt it.

So here's the question we keep coming back to at Humano: if building is no longer the bottleneck, what is?

What the numbers say

The clearest signal comes from how new companies are growing. AWS's 2026 Engines of Growth study surveyed more than 3,400 founders and senior leaders across 20 countries. It found that AI-native startups — companies under five years old with AI at the core of the product — are reaching billion-dollar valuations in about 3.5 years, roughly half the time and with half the staff earlier cohorts needed. They reported 156% average annual revenue growth against 65% for startups overall, and 55% of them clear $400,000 in revenue per employee.¹

It's a survey commissioned by a cloud provider, so it deserves a raised eyebrow. But the direction matches what anyone shipping product can see: small teams are doing what used to take large ones.

The more careful evidence is humbler, and more interesting. METR, which runs controlled studies of AI and developer productivity, found in 2025 that experienced open-source developers were actually 19% slower with the AI tools of the time, while believing they were 20% faster.² By early 2026 the picture had shifted: developers increasingly refused to take part in studies that asked them to work without AI, and METR concluded the real speed-up was probably larger than its measurements could show.³ In a separate note, it found that many AI-written changes that pass automated tests still wouldn't be accepted by the people who maintain the code.⁴

Read together, the message is simple. AI makes producing software dramatically cheaper. It doesn't make producing the right software automatic.

What got cheap, and what didn't

When something becomes abundant, value moves to whatever is still scarce. Here's how we see the split.

Cheap now:

• Prototypes, landing pages and first versions of a product
• Boilerplate code, integrations and internal tools
• Variations — ten versions of a screen, a flow or a pitch

Still scarce:

• Knowing which problem is worth solving. The market doesn't reward volume. It rewards being right about a real, painful problem.
• Taste. When everyone can generate a passable interface, the difference is in the details people feel but can't name.
• Trust. Customers, especially enterprises, buy from people they believe will still be there next year.
• Distribution. A product nobody hears about is a very well-built secret.
• The discipline to stop. Cheap building makes it easy to keep polishing something nobody wants.

That last one matters more than it sounds. When a prototype costs a week instead of a year, the expensive mistake isn't building the wrong thing. It's continuing to build it.

The enterprise lesson

Large organisations have run into the same wall from the other side. MIT's Project NANDA studied enterprise generative AI in the first half of 2025 and found that around 95% of organisations saw no measurable return from their pilots, despite an estimated $30–40 billion of spending.⁵ The study's own definition of success was narrow, and the headline number has been argued over since.

What's worth keeping is the authors' diagnosis. Most failures weren't about model quality. They came from how the tools were deployed: systems that didn't learn from feedback, didn't fit the workflow and weren't owned by anyone who could change course. The technology was ready. The judgement around it wasn't.

Why we build the way we do

This is the reason Humano is a venture studio that builds, rather than a fund that writes cheques.

When the cost of trying an idea falls, the best strategy is to try more ideas, learn faster and be ruthless about what you keep. That works only if the people judging the idea are the same people building it. Founders, designers and engineers sit on one team, so what we learn from a prototype on Tuesday changes the plan on Wednesday.

In practice that means a few things:

1. We start from the problem, not the model. Every venture begins with a specific person and a specific moment that's broken. The AI comes in because it changes what's possible there, not because it's fashionable.
2. We ship to learn. Early versions go in front of real users in weeks. Their behaviour, not our enthusiasm, decides what happens next.
3. We design it to be trusted. Taste and clarity are not a final coat of paint. They're how a product earns the right to be used every day.
4. We kill ideas on purpose. A studio's edge is its willingness to stop. Every venture we don't continue frees the team for one that deserves it.
5. Ventures that work get their own team. Once something proves itself, it spins out with dedicated people and funding to grow.

You can see the shape of this in what we've built so far: LiveTwin, an AI twin that prepares people for reviews and customer moments inside the tools they already use; Bümi and BumiPay, bringing fair financing and simpler payments to Malaysian people and businesses; and Wildchild AI OS, applied AI built around how a business actually runs. Different markets, same approach: find the real problem, build quickly, keep only what earns its place.

The new bottleneck

The founders who win this decade won't be the ones who can build the most. Almost everyone can now. They'll be the ones with the best judgement about what deserves to exist, and the nerve to act on it quickly.

That's the work we care about. If you're holding an idea that sounds a little too wild, we'd like to hear it.

---

Footnotes

1. AWS Startups, Engines of Growth global startup trends report, June 2026. Press release. ↩

2. METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025. Paper. ↩

3. METR, We are changing our developer productivity experiment design, February 2026. Post. ↩

4. METR, Many SWE-bench-passing PRs would not be merged into main, March 2026. Note. ↩

5. MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. Coverage. ↩