Yes, AI tooling makes software cheaper and faster to build. On recent projects we have watched build phases compress by roughly a quarter compared with our 2024 baselines. What has not moved at all is the time it takes to work out what to build, for whom, and why anyone would pay for it. Discovery in 2026 takes about as long as discovery in 2019.
That gap matters because founders increasingly budget as if AI discounts the whole project. It discounts one slice of it: the typing.
The speed gains are real, but they live in the typing
The evidence for faster building is solid where the work is well defined. In GitHub's controlled experiment, developers using Copilot finished a scoped HTTP server task 55.8% faster than the control group. McKinsey's lab study found generative AI helped developers document code 45 to 50 percent faster and refactor it 20 to 30 percent faster, with gains shrinking below 10 percent on tasks developers rated highly complex.
Notice the pattern. The clearer the specification, the bigger the gain. AI is superb at producing code once a human has decided precisely what the code should do.
The counter-evidence points the same way. A randomised trial by METR found experienced open-source developers actually took 19% longer on real issues in mature codebases when using AI tools, even though they believed AI had sped them up by 20%. Messy context, ambiguous requirements, high judgement content: the gains evaporate, and perception stops matching reality.
Discovery did not get faster, and that is the point
The chart above, built from McKinsey's developer productivity study, shows where the savings concentrate: routine, well specified work. Nobody has published a study showing AI compresses customer interviews, problem validation, or the argument about which workflow deserves to exist. Those activities are constrained by how fast humans can learn, decide, and agree, not by how fast anyone can type.
In practice, that means the shape of a project budget has changed. Two years ago, discovery might have been 15% of the spend and build 85%. Compress the build and discovery becomes a larger share of both the cost and the risk. Skimping on it now damages a bigger proportion of the outcome than it used to. That is why structured product discovery and strategy work has become more valuable as AI has made everything downstream cheaper, not less.
Faster output raises the cost of building the wrong thing
Speed without direction just gets you to the wrong place sooner. The 2025 DORA report found AI adoption among software professionals hit 90% and is now associated with higher delivery throughput, but it continues to correlate with worse delivery stability. Teams ship more, and more of what they ship wobbles. DORA's own year in review puts it plainly: AI amplifies whatever system it lands in, strong or weak.
The same logic applies one level up. If your discovery is sharp, AI amplifies a good decision. If your discovery is thin, AI helps you produce three months of polished, tested, wrong product in six weeks. The thinking was always the constraint; AI has simply made that visible by removing the other one.
FAQ
Does AI actually reduce software development costs?
For the build phase, yes, and meaningfully. Controlled studies show 20 to 56 percent time savings on well specified coding tasks. Total project cost falls less than headlines suggest, because discovery, integration, review, and stabilisation do not compress at the same rate.
Should we spend less on discovery now that building is cheaper?
The opposite. When building was expensive, a wrong bet was throttled by slow output. Now a wrong bet ships at full speed. Discovery is the cheapest point in the project to be wrong, and cheaper builds make that asymmetry stronger.
Why did the METR study find AI made developers slower?
Because the tasks were real issues in large, mature codebases the developers knew deeply. Reviewing and correcting AI output cost more than it saved. AI gains concentrate in greenfield, well specified work, which is exactly what a good discovery phase produces.
A checklist before you bank the AI discount
1. Split your budget into thinking and typing, and only apply AI savings to the typing.
2. Timebox a proper discovery phase and treat its outputs as the specification AI tools need to perform well.
3. Validate the core workflow with real users before generating a line of production code.
4. Invest in tests and review capacity, since throughput gains without them show up as instability.
5. Re-estimate quarterly, because tool capability and the evidence base are both moving fast.