I've now talked to the third company that expects to make a profit because AI writes their code faster.
On the surface it makes sense. Models keep getting better and quicker at writing production-level code, and I don't see that stopping. Give it a while and anything you hand an agent will come back as the perfect equivalent of what you asked for, in whatever language you asked for it.
That last part hides the issue. The perfect equivalent of a bad task is bad code. If you give it shitty tasks, it produces incorrect results, only faster.
So when a company tells me they want to go AI-native, I point them at product principles and operations. Code is getting heavily commodified, and those two areas of AI work still need a lot of judgment.
The product part: verify the need
An enterprise client comes to you with a need. Before anything gets built, verify it: will the thing they want actually bring them something?
By the time an enterprise request reaches you, it has usually been through 30 meetings, a few escalations and whatnot, and it's most likely missing any KPI or success definition. So ask the questions those meetings skipped:
- How will we know what we're about to build is a success?
- Which KPI tells us so?
- How will we know we should delete it after two weeks?
This is where I use AI, with my own judgment in charge. I use a model as a sparring partner to poke holes in my thinking and to do the research. I decide which holes matter and what the answer is.
Once the product part is nailed down, you can split the work into pieces and run them in parallel, so your agents turn the perfect definition of each task into a perfect version of your code. That only works if the definitions really are right, which is why the product part comes first.
The operations part: check access the moment it arrives
Say the same client needs you on their VPN. Every question on the list is a yes or no:
- Do you have VPN access?
- Once you have it, can you connect? Do you get through?
- Do you have credentials?
- Is two-factor mandatory?
- Do you have access to every internal system the task requires?
Usually the email or Slack message lands, you come back to it two hours later, try to connect and find you can't. Now it's "Okay, go get the VPN." Checking the internal systems by hand is slower still: connecting, poking at endpoints in Postman, looking at what shapes and what errors come back. Done manually, that takes hours and hours.
An agent can run those checks automatically the moment the message reaches you. It connects, hits the endpoints and records what came back and what failed. I read the result, decide what's actually missing and write to the ops person: "I can get through this. I can't get through this. Here's what I need." All of that is work AI can handle, if it's properly supervised.
Why 30 seconds beats an hour
This doesn't shave off a trivial amount of time. If you get back to someone 10 minutes later, or an hour later, they're probably in a different meeting. If you get back to them 30 seconds later, there's a good chance they're still sitting at their computer, maybe about to close the windows they were supposed to close.
This applies to smaller clients too
If you don't work with enterprise clients, this section is for you. Any customer who outsources features or development to you likely has the same gaps, and they're often missing things that enterprise clients have for enterprise reasons. The product questions and the access check matter at least as much there.
Where to start
Before your next client kickoff, write down how you'll know the build worked and when you'd delete it, and set up an agent to run your access checklist the moment the request lands.