TL;DR
"How much faster are we with AI?" has no answer, because only a particular kind of work can be measured. Whoever measures the first attempts measures beginners: agentic engineering is learned over months and years. So pick work that suits an agent, measure it, and keep measuring. Hard numbers exist only where work repeats; elsewhere, follow the trend and ask the people doing it. Use what you find to decide where to invest next: training, a better setup, or support.
"How much faster are we with AI?" is usually the first question asked, and it assumes the speed arrives with the tool.
Someone who opens Excel for the first time builds a slower, worse budget than the colleague who has used it for ten years. Nobody takes that first spreadsheet as the measure of what Excel does for a finance department. asks for more than a new tool does: it changes how work is specified, handed over and checked.
The first number measures the learning curve.
#An early reading shows where people stand this month. It says little about what the same work gains once agents are in daily use, once teams know which of their work suits an agent, and once the setup is tailored to that work. That climb takes months and years, and it keeps going, because the tools keep changing underneath it.
- The tool arrives
The reading everyone quotes, taken at the lowest point on the curve.
- It is in daily use
People hand work over every day, and get it wrong often enough to learn from.
- The fitting work is found
Teams know which of their work an agent is good at, and stop pointing it at the rest.
- The tool is tailored to that work
It is set up for the job in front of it rather than used as it arrived.
A number that means little alone still means something as a trend.
#A clean before-and-after exists only where work repeats: a migration you have run before, a monthly report, one class of tickets. Most engineering work is done once, and nobody can say how long it would have taken the other way. Counts such as tickets closed per month are relative too, because no two tickets cost the same effort. Track them anyway, per kind of work, and read the direction. The same imperfect number, taken the same way each quarter, shows whether a team is climbing or stuck, as long as the mix of work stays comparable.
What people report can be wrong, and it still shows who will get good at this.
#Ask the people doing the work what they gain, about one named kind of work. In a controlled trial in 2025, experienced developers took 19% longer with AI tools and believed they had been 20% faster. Most of them had spent only a few dozen hours with the tool: a first reading again. Their answer still tells you what no throughput figure does, which is whether they accept working this way. Only someone open to it keeps practising, and practice is what moves the curve. Where people report no gain, find out which it is: the work does not suit an agent, the practice never became daily, or the setup is wrong.
Whether to use agents on work that suits them is settled. The numbers are for steering.
Keep exploring