For most of the history of software, writing code was the slow, expensive part. It took skilled people a long time, so development organizations were set up to write as much code as they could. Requirements, tests and design notes were treated as lesser concerns.
AI coding tools changed that. Give an AI agent a clear description and it can write a working feature in minutes. Two things follow.
- First, the code is no longer the most valuable thing you own. The description is, along with the tests and the architecture rules the system has to follow. If you lost the code and kept those, you could generate the code again. If you kept the code and lost those, you would have software that nobody can safely change. I call this the asset inversion.
- Second, the slow, expensive part of the work is now somewhere else. Deciding what to build takes as long as it ever did. So does checking that what was built is correct. Those two now take most of the time and cause most of the delays.
Most engineering leaders I talk to think of this as something that happened to software. It also happened in marketing.
The same change in marketing
In August, Wharton marketing professor Americus Reed II described the same change in his own field, in an article for Knowledge at Wharton. AI lets any company produce marketing content quickly and at low cost. Because every competitor can do the same, producing more content no longer sets a company apart. What sets a company apart, he argues, is whether customers see the brand as part of who they are. Only customers can decide that. AI cannot produce it.
Reed is writing about brands, and only about brands. But compare the two:
- Software: AI made code fast and cheap to write. People still have to decide what the software should do, and check that it does it.
- Marketing: AI made content fast and cheap to write. People still have to give customers a reason to see the brand as part of who they are.
In both fields, AI took over producing the thing. It did not take over deciding what is worth producing. Two fields with little in common changed in the same way within a few months of each other. I expect the same in any work where AI can now write the first draft: legal documents, analysis, design, reports.
What it does to roles
Almost every role in a delivery organization does three kinds of work:
- Deciding what to produce.
- Producing it.
- Checking that what was produced is right.
AI made the second kind much faster and much cheaper. It has not done the same for the first and third. Here is how that looks in four roles.
Business analyst
- Deciding: finding out what the stakeholder needs, which is often different from what they asked for.
- Producing: writing the requirements document.
- Checking: confirming that what was built meets the need.
Developer
- Deciding: choosing how the feature should work within the architecture.
- Producing: writing the code.
- Checking: confirming that generated code is correct, is safe, and follows the architecture rules.
Scrum Master
- Deciding: working out what the team most needs to improve.
- Producing: running the meetings.
- Checking: confirming that the team is improving.
Anyone who leads
- Deciding: choosing what the teams work on and what "finished" means.
- Producing: collecting and reporting status.
- Checking: confirming that finished work produced the result you expected.
A person who spent most of the day producing has less to do than before. A person who is good at deciding and checking is needed more than before, because a team can now build the wrong thing in minutes. In my client work, I rarely see organizations look at roles this way. They choose whom to keep by job title, and they let go of the people who were doing the deciding and the checking.
Three questions for an engineering or product leader
- Where does your budget go? If most of it pays for writing code (more developers, more tool licences) and little of it pays for clearer requirements, better tests and faster decisions, most of your money is going to the work AI already made cheap.
- On each team, who decides what to build, and who checks that it is right? If the answer is "everyone," find out whether anyone has the time to do it.
- The last time you reorganized, did you choose by job title, or by who does the deciding and the checking?
This is why I say agile isn't obsolete. Its core practices are these: the team agrees on what to build, gets feedback often, and decides together whether the work is finished. Those are the deciding and the checking. The teams I have seen release more software with AI tools are the ones that kept doing them.