Forge

Every development organisation can now learn at AI speed

Chapters I and II were about what AI can do for the services themselves. But every development organisation, whatever programme it runs, also has an enormous opportunity to move faster with AI: to run its core operations better, and to learn faster what works.

Unfortunately, very few people in development have been close to how fast things are now moving, and I think that's why our level of ambition is lower than it could be.

The opportunity

In development, execution matters just as much as intervention. The same teaching method can do a great deal or almost nothing, depending on who runs it and how.

So the best programmes keep getting better the way good products do. They try versions, measure each, and keep what works. I've been very aware when writing these essays of the (true) adage: technology is not a silver bullet. One day, walking through the park, I realised: technology is not a silver bullet, but experimentation is a forge.

AI is making that forge much, much cheaper and faster, and I think that within a year it'll run much of the loop itself: building the versions, measuring them and keeping what works. We know that teams that build with it can ship far more, with fewer mistakes, because a few development organisations already work this way.

This isn't only true for digital products. It's just as true for the ordinary systems that run a programme, and at most organisations that's where the impact is, rather than a flashy AI strategy: proven programmes going bigger and more effective by doing the boring stuff better and faster. As one example, a highly effective education programme is commissioning software to assemble its lesson plans. Its RFP follows what used to be a normal timeline for a build like that, and the gap vs. what's now possible leaves half a year of impact (and, I suspect, around $100k) on the table.

Our ambition is lower than it could be

I think we could raise our ambition for how fast we improve our programmes by a lot. Today, our pace of change leaves a lot on the table. AI tools are allowing startup engineers to build 3x as fast; at digital-first development organisations, the pace has barely moved.

This is partly because fewer of our engineers visibly build with AI, though those who do have doubled their output.

But what seems like the biggest missed opportunity is that we're not trying hard enough to speed up: far fewer development organisations are asking their engineers to build with AI than in the private sector.

Few of us have been close to it

I think the reason is that very few people in development have direct experience of what's now possible. I looked at 40 effective development implementers, drawn at random from the grantees of five large funders.

Unfortunately, the sector's tech advisers mostly don't have this experience either. To be clear, this is not a failing! The world is just changing really quickly (most of this happened in the last three years) and most of the sector simply hasn't been close to it.

If you haven't seen how fast things now move, it's hard to know what to ask for, even from the people you hire or contract to build technology.

How do we bring more frontier ambition to the sector?

I think that the way we capture this opportunity is mostly by expecting more. That looks like leaders asking if we can build in weeks rather than quarters, or test many versions of an idea rather than one. At the frontier, raising that expectation is now a big part of senior leadership's job.

This isn't to say that every organisation should hire a big team of expensive AI engineers. The labs are building the models and many of the products, as Chapter II argues. What matters is just increasing the density of people who know what the frontier makes possible and who can help leaders raise their ambition to match.

I can see three sketches of ways to do that. I'd love to chat about these ideas!

Next in the series: Method →