If you spend any time around AI people, you'll hear "recursive self-improvement" within about ten minutes. Usually said in a hushed voice, like it's the moment the machines take over.
Strip the drama out and it's a simple idea. And it's one of the most useful ideas a business owner can borrow right now.
What recursive self-improvement means
Recursive self-improvement is a system that gets better at its job, and then uses what it's learned to get better at getting better.
The "recursive" bit is the important part. It isn't just improving. It's improving the thing that does the improving, so every round starts from a higher point than the last.
This has stopped being theory. In May 2025 Google DeepMind announced AlphaEvolve, an AI coding agent built on its Gemini models. It found a faster way to run part of Gemini's own training, speeding up one key step by 23% and cutting overall training time by 1%. It also found a better way to schedule work across Google's data centres that recovers, on average, 0.7% of Google's worldwide computing power.
Read that again. The AI made the training of the AI underneath it faster. That's the loop the tech world is excited about.
Why it matters to a business that doesn't build AI
You don't need to be training models for this to matter. Point the same idea at your own business.
Your marketing is a process. Your sales follow-up is a process. Every one of them has a version that's slightly better than the one you're running today. The problem has never been knowing that. The problem is that nobody has the time to keep looking.
Someone tweaks the email subject lines once, sees a small lift, and moves on. Six months later nobody remembers what was tested or why. The process freezes at "good enough".
AI changes that because it doesn't get bored and it doesn't forget. It can look at every enquiry, every email and every lost deal, spot what's dragging, change one thing, measure it, and then do it all again next week on the improved version.
What it looks like in practice
Here's a simple example. An AI agent reviewing how a business handles new enquiries, week after week.
- Week 1
- Looked at
- Read 240 enquiries and what happened to each one
- Found
- Enquiries answered after two hours booked half as many calls
- Changed
- Every new enquiry now gets a reply within five minutes
- Result
- Calls booked: 14% to 19%
- Week 2
- Looked at
- Read the replies to the faster first message
- Found
- People asked about price before agreeing to a call
- Changed
- Added a starting price to the first reply
- Result
- Calls booked: 19% to 23%
- Note to self
- Removing a question beats rewording the message. Try that first.
- Week 3
- Looked at
- Checked which questions still came up before booking
- Found
- Most people wanted to know how long a project takes
- Changed
- Added typical timescales to the first reply
- Result
- Calls booked: 23% to 26%
- Note to self
- Confirmed. Answer the next question before it's asked.
Look at week two. It didn't just make a change. It wrote itself a note about which kind of change works. In week three it used that note and got to the answer faster.
That's the recursive part. The process gets better, and so does the way it gets better.
The same works for sales. An agent reads the notes from every deal you lost last quarter and finds that most went cold after the quote, not before it. So it tries a check-in two days after every quote goes out, measures it, and keeps going. Then it looks at which follow-ups get replies and which get ignored, and tunes those. Each fix makes the next problem easier to see.
Small gains, every week
None of those changes is dramatic on its own. That's the point. Recursive self-improvement isn't one big breakthrough. It's lots of small ones that never stop.
And small gains stack. A process that gets 2% better every week is around 2.8 times better after a year. A process nobody touches is exactly as good as it was on the day it was set up, while your competitors' get better around it.
Most businesses already know their marketing and sales could be better. What they lack is a way to keep improving them that doesn't depend on someone remembering to.
What has to be in place first
This isn't magic, and it isn't something you switch on. A few things have to be true.
One clear number per process. Calls booked, quotes accepted, repeat orders. If you can't say what "better" means, the AI can't either.
Data you trust, in one place. The AI can only learn from what it can see. If your enquiries live in three inboxes and your deals live in someone's head, fix that first. I wrote about this in Do you trust the data?
A person signing off the changes. Even the research versions need watching. Sakana AI's self-improving coding agent more than doubled its score on a coding test by rewriting its own code, from 20% to 50%. It also, at one point, faked a log to make it look as though it had passed tests it never ran. Let the AI find and test the changes. Keep a human approving what goes live, at least until it has earned your trust.
A process that's worth improving. AI will make a decent process better, round after round. It won't rescue a broken one. That's a theme I keep coming back to, including in why AI won't fix a weak sales process.
How we build it
This is the work we do at LeadDigital. We connect a business's marketing and sales data into one place, usually a CRM. We build AI agents that review each process on a regular cycle. And we set up the approval step, so the owner sees every proposed change, with the evidence, before it goes live.
We start with one process, normally new enquiries or sales follow-up, because those are where small gains turn into money fastest. Once that loop is running and trusted, we add the next one.
If you want to know more about how agents differ from the chatbots you already use, start with Assistants talk. Agents work.
The real shift
For most of business history, improvement happened in bursts. A new manager, a consultant, an away day. Then things settled back down.
Recursive self-improvement means the improving never has to stop. Your marketing and sales can be a little better every week, and better at getting better.
The businesses that win won't be the ones with the best process today. They'll be the ones whose process improves itself.




