Every AI coding tool demo ends with some version of the same pitch: build in minutes what used to take weeks, at a fraction of the cost. It's a compelling story, and I hear it from prospective clients more every month — "can't I just have AI build this for way less?" I build software for a living, so I have a stake in the answer. But the actual research on AI-assisted development complicates the pitch a lot more than the sales decks let on, so let's actually test the theory instead of just reacting to it: will AI application development bring the cost of software down, or does it just move the cost somewhere less visible?
The Case For: AI Is Already Cutting Costs
Start with the part that's genuinely true. AI has made a real dent in how long it takes to produce working code, and that shows up in real numbers, not just vendor talking points.
Individual Developers Are Shipping Faster
Controlled experiments have found developers completing well-defined programming tasks up to 55% faster with AI assistance, and broader industry data points to roughly 3.6 hours saved per developer per week. Adoption backs that up: 84% of developers now say they use or plan to use AI tools in their workflow, and telemetry from real codebases shows about 22% of merged code is now AI-authored. That's not a fringe trend anymore — it's most of the industry.
The catch, even in the optimistic research, is that the gains aren't automatic. A systematic review of the productivity research found that AI adoption alone doesn't improve delivery performance — it only pays off for teams that already have strong engineering habits: small batch sizes, automated testing, mature CI/CD. Hand the same tools to a team without that discipline and the results get a lot less predictable. Keep that caveat in your back pocket — it matters a lot for the second half of this article.
AI App Builders Are Replacing Entire Dev Cycles for Simple Projects
The bigger cost story isn't developers typing faster — it's tools like Bolt.new, Lovable, v0, and Replit Agent letting non-developers skip the build entirely for simple projects. The savings being reported are real: AppDirect's marketing team built 11 internal projects with AI tools and documented over $120,000 in software cost savings. Delivery Hero cut feature-validation time by 66%, building prototypes in an hour that used to take three weeks. Gartner projects 75% of new applications will be built using low-code tools by 2026, up from just 40% in 2021, and 87% of enterprise developers already have their hands on these platforms in some form.
I've made this exact case before about websites specifically — AI-powered drag-and-drop builders really have changed what a simple build costs, and pretending otherwise doesn't help anyone. For a landing page, an internal tool, or a genuine minimum-viable-product, the cost curve has bent down and it's not bending back.
Commodity Work Is Getting Commodity Pricing
I'll add one observation from my own work that isn't from a study, just experience: the kind of build that used to justify a week of a contractor's time — a simple booking form, a basic CRUD internal tool — now genuinely takes an afternoon with AI-assisted scaffolding. Clients have started to expect that, and honestly, they should. When the work really is commodity work, AI has commoditized the price along with it.
The Case Against: Why the Bill Often Comes Due Later
None of that is fake. But it's also not the whole picture — and the research that looks past "how fast did the first version ship" tells a very different story.
Experienced Developers Doing Real Work Are Actually Slower
The most important study on this topic isn't a vendor case study — it's a randomized controlled trial from METR, an independent AI research nonprofit. They had 16 experienced open-source developers complete 246 real issues in codebases they already knew well, half with AI assistance and half without. Going in, the developers expected AI to make them about 24% faster. The actual result: AI-assisted work took about 19-20% longer — and even after seeing their own timed results, the developers still walked away believing they'd been faster.
METR's explanation isn't "AI is bad." It's more specific than that: weak prompting habits, unfamiliarity with the tooling, a quality bar on real production code that AI's suggestions often don't clear, and the plain cognitive overhead of reviewing someone (or something) else's output line by line. To be fair to the "case for" side, METR is explicit that this finding doesn't claim to generalize to beginners, to greenfield toy projects, or to boilerplate generation — exactly the categories the cost-savings stats above are drawn from. That's not a contradiction. It's the same finding from two different angles.
The Bill Moves Downstream: Security and Technical Debt
This is where the "cheaper" argument gets expensive fast. Veracode tested over 100 language models on security-sensitive coding tasks and found 45% of AI-generated code samples introduced an OWASP Top 10 vulnerability — 86% failed to defend against cross-site scripting, 88% were vulnerable to log injection. Separately, Apiiro tracked security findings at Fortune 50 companies and found them jump 10x in six months — from about 1,000 to over 10,000 monthly — correlated directly with AI coding tool rollout.
The maintenance math tells the same story. Research on AI-heavy codebases has found maintenance costs balloon toward 300-400% of baseline within 18-24 months as unreviewed shortcuts compound, test coverage fall from an industry-typical ~68% down to as low as 12%, code duplication rise 48%, and refactoring activity — the work that actually pays down debt — drop 60%. For the first time on record, copy-pasted code has overtaken refactored code in real-world change data. None of that shows up in a launch-day cost comparison. It shows up 18 months later, as an invoice.
The real-world price tag is already visible: industry estimates put the number of startups that shipped AI-built production apps and now need a full or partial rebuild at over 8,000, at $50,000 to $500,000 each — somewhere between $400 million and $4 billion in cleanup cost, industry-wide, on work that was sold as the cheap option.
Enterprise Numbers Tell the Same Story
It's not just startups cutting corners. At the enterprise level, the total cost of ownership on a custom AI build typically runs 1.6 to 2.2 times the initial build cost over 24 months, with most projects clustering around 2x once you count ongoing monitoring, retraining, and infrastructure scaling — that alone adds another 15-30% of the build cost every year it stays in production. Moving a working pilot into real production typically costs 250-400% more than the pilot did, mostly on data pipeline work, security hardening, and integration nobody scoped for at the demo stage.
And the return isn't reliable even when the spending is. MIT's NANDA initiative found 95% of enterprise GenAI pilots produced no measurable effect on the P&L. Other surveys land all over the map depending who's asking and how success gets defined — Stanford's Digital Economy Lab found just 12% of deployments clearing 300%+ ROI against 88% at or below break-even; McKinsey put 23% of AI initiatives at real scale; Gartner found 28% fully paying for themselves. That spread isn't noise. It's the same message as the METR study, at enterprise scale: execution decides the outcome far more than the tool does.
My Take
Both sides of this argument are true — they're just true about two different kinds of software, and I think that's the actual answer the theory is missing. The cost curve genuinely bent down for one category: small, short-lived, low-stakes builds — prototypes, internal tools, landing pages, a simple booking form. That's exactly what the AppDirect, Delivery Hero, and Gartner numbers above are measuring, and I don't think that trend reverses.
The curve did not bend down — and by the maintenance and rebuild numbers, may have bent up — for the other category: complex software a business actually has to depend on working correctly for years, under real users and real edge cases. That's exactly what METR, Veracode, Apiiro, and the enterprise TCO numbers above are measuring. The uncomfortable part is that the tools and workflow that work great on category one don't automatically survive contact with category two just because someone used the same AI model for both.
I've landed here before, in a narrower context — AI website builders haven't made web designers obsolete, for basically this same reason. This is that argument playing out one level up, at the scale of full application development instead of just websites. It's also part of why I don't treat "runs cheap" and "was built cheap" as the same claim — the same logic is behind why I build client sites to run on close to $0/month in hosting: keeping ongoing costs low is a deliberate architecture decision, not a byproduct of skipping the labor that goes into getting the architecture right in the first place. And if you're curious why AI tools are priced the way they are in the first place, I've broken down how tokens and credits actually work in more detail.
So: will AI bring the cost of software down? For the easy stuff, yes — dramatically, and that part isn't going away. For anything a business actually has to depend on for the next five years, the honest answer right now is that AI shifts the cost forward in time rather than erasing it. Pretending otherwise is exactly how a company ends up as one of the roughly 8,000 currently budgeting for a rebuild of something that was supposed to be the cheap option.
Not sure whether your AI-assisted build is a genuine quick win or a five-year problem wearing a quick-win price tag? Reach out and I'll give you a straight read.