Build vs Buy in the AI Era: Why Cheap Code Still Costs You
Build vs Buy in the AI Era: Why Cheap Code Still Costs You
AI made the first bill cheap. Owning the tool is the second bill, and it lands on your engineers. 4 questions to ask before you build.
Two engineers with an AI coding assistant, for one week. That’s roughly $7,700 of salary to rebuild a tool your company pays $30,000 a year for (illustrative numbers, the full model is in Exhibit 1).
On paper, the math looks obvious. And more teams are running it, because software bills keep climbing. The average organisation in Zylo’s 2026 SaaS Management Index spends $55.7 million a year on SaaS, up 8% in a year.
Traditionally, when a company wanted a new tool, say a CRM or a ticketing system, it would license a ready-made solution. Now that AI tools get better every month, many teams are asking, what if we build this ourselves?
This used to be a debate for large organisations with large engineering teams. Today, it’s a common conversation between a startup’s CEO and CTO.
When everything is so cheap to build, what is the case for buying anymore?
Sponsored by Hyphenate
Finance is where “we’ll build it ourselves” gets expensive fast, because every reconciliation needs someone accountable for it.
Hyphenate just raised $45 million in Series A funding to scale autonomous finance for the office of the CFO.
▫️ Does the work: reconciles cash, runs invoicing, maintains subledgers and moves the close forward
▫️ Keeps people in charge: learns from every exception your team resolves and leaves the judgment calls to them
▫️ Fits your stack: layers onto your ERP and banking systems and keeps evidence behind every action
Key Takeaways
Software has 2 costs: making it and owning it. AI coding tools cut the first one, so owning the tool is now the bigger bill.
Owning code eats engineering time. In Stripe’s Developer Coefficient survey, developers estimated that 17.3 of a 41.1-hour week goes to maintenance, about 42%.
Internal builds look cheap because their upkeep hides inside engineers’ salaries. Give each one a shadow budget line.
Vendors get the same AI discount and spread upkeep across their whole base. In MIT NANDA’s 2025 research, AI tools bought from specialised vendors succeeded about 67% of the time, against about a third of the time for internal builds.
AI-written code creates comprehension debt. In Anthropic’s trial, developers who used AI scored 50% on a quiz about the code they had just written, against 67% for those who coded by hand. Build small, low-risk tools with a written exit, and buy anything whose failure would be catastrophic.
Table of Contents
Is Building Software With AI Cheaper Than Buying It?
Why Does Building Look Cheaper on the Books?
Will SaaS Get Cheaper as AI Cuts the Cost of Code?
What Is Comprehension Debt in AI-Written Code?
When Should a Startup Build Instead of Buy?
FAQ
1. Is Building Software With AI Cheaper Than Buying It?
AI has made building software cheap, but it also made the job sound simpler than it is.
That’s because building a piece of software comes with two different costs. The first is the cost of making it: everything it takes to reach a usable, working version.
The second is the cost of owning it, which starts the day the tool ships and runs for as long as anyone uses it. It’s also the cost that’s hard to estimate in advance, because owning a tool means 6 recurring jobs:
Applying security patches when a new vulnerability lands
Upgrading it to keep up with its dependencies
Fixing it when it breaks at 3 a.m.
Adding the features users keep asking for
Getting it through audits
Refreshing the UI every few years
Add the two together (making & owning) and you get the total cost of ownership.
For most business software, the second bill is now far larger than the first, because AI tools have cut the first one sharply.
And the second bill was big to begin with. In Stripe’s Developer Coefficient survey with Harris Poll, developers estimated that their teams spend 17.3 hours of a 41.1-hour week on maintenance work like debugging and refactoring. That’s about 42% of their time, and the survey dates from 2018, so it’s the baseline AI coding tools started from.
AI helps with that second bill only at the margins. A tool can flag a newly announced vulnerability, but a person still has to decide whether it matters to your business, and a person has to take responsibility when something goes wrong.
That work needs a human, and a person’s time is the most expensive thing a software company has.
AI cut the cost of making software. The cost of owning it still lands on your best engineers.
The time AI saves is also easy to overestimate. In METR’s 2025 randomised trial, 16 experienced open-source developers took 19% longer on real tasks when they used AI tools, while believing AI had made them 20% faster. METR has since said tools have likely improved and those early-2025 results are dated. The gap between how fast it feels and how fast it is still deserves a line in your spreadsheet.
2. Why Does Building Look Cheaper on the Books?
The main reason building feels more attractive now (besides AI) is that it also looks different on the books.
A software licence is a visible budget line that the finance team reviews at every renewal.
And right now, that line is under pressure. In Zylo’s 2026 index, 78% of IT leaders reported unexpected charges tied to AI features or usage-based pricing, and 61% cut projects because of unplanned SaaS cost increases. When finance needs room for AI spend, the SaaS line is the easiest target.
The maintenance of an internally built tool, on the other hand, gets buried inside other budget lines. It stays invisible until something breaks.
That cost is spread across engineers’ salaries and buried in timesheets and tickets that look unrelated to the tool.
So the real cost of the tool rarely reaches the budget meeting.
That’s why building tends to win: its costs are harder to see. And an invisible cost rarely gets questioned.
So if you decide to build, give every internal tool that replaces a paid product a shadow budget line. Set it at a reasonable estimate of its maintenance cost, and review it on the date the licence would have renewed.
3. Will SaaS Get Cheaper as AI Cuts the Cost of Code?
That build-vs-buy spreadsheet usually assumes the vendor’s price stays put while the cost of building falls.
Vendors use the same AI tools their customers do, so their costs fall too. And they spread the cost of maintaining their product across their whole customer base. HubSpot, for example, reported 306,446 customers at the end of June 2026. A company building its own CRM carries that same upkeep alone.
That is called amortization, and it’s why shared software usually beats a private build on cost.
The track record on AI projects points the same way. In MIT NANDA’s The GenAI Divide: State of AI in Business 2025, AI tools bought from specialised vendors succeeded about 67% of the time, while internal builds succeeded about a third of the time, as Fortune reported. The figures are self-reported, so treat them as a signal. Still, Aditya Challapally, the report’s lead author, told Fortune that “almost everywhere we went, enterprises were trying to build their own tool.”
Expect plenty of those builds to get shut down. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, with escalating costs and unclear business value among the reasons.
Writing Code vs Carrying Risk
Where a product is mostly code, cheap AI development will probably bring lower prices and new, cheaper competitors, which makes building your own even less attractive.
Where a product is mostly about carrying risk, such as login and identity systems or payment processing, cheap code changes very little.
The expensive part of those businesses is absorbing the consequences when something goes wrong.
Investors are watching the same split. As we argued in Don’t Believe Those Who Say that SaaS is Dying, weak horizontal tools that mostly organise information or route simple tasks are the ones facing a harsher buyer question. Software that sells convenience is exposed.
Software that takes on someone else’s liability is likely to become more valuable, because the more custom glue companies build, the more they need the dangerous parts to be someone else’s problem. Incumbents are also racing to add agents on top of the data they already hold, a fight we covered in SaaS Isn’t Dead, the Real Fight Is the Agentic Layer.
4. What Is Comprehension Debt in AI-Written Code?
We have talked about technical debt before, which is how engineers describe the shortcuts that save time now but cost time later, such as skipping tests or hard-coding a value that should be adjustable.
AI-assisted building adds a related problem that Addy Osmani calls comprehension debt: the growing gap between how much code exists in your system and how much of it any human genuinely understands.
When a model writes most of a tool during a long chat session, the person who prompted it often understands the result less well than an engineer who wrote every line would have.
That’s because the reasoning lives inside the model’s output, or inside a chat that’s long gone.
There’s now data on this. In a randomised controlled trial by Anthropic, 52 mostly junior engineers learned a new Python library, half of them with an AI assistant. The AI group scored 50% on a quiz about the code they had just written, and the hand-coding group scored 67%. The biggest gap was on debugging questions, which is the exact skill you need the day your internal tool breaks.
How people used the AI mattered a lot. Participants who handed the coding over to it averaged below 40%, while those who used it to ask conceptual questions or explain the code scored 65% or higher.
So what happens two years later, when the original author leaves the company, and someone new has to fix the mess?
Forget two years later. With AI-built tools this risk shows up almost immediately, because the code can be hard to follow even while its own author is still at the company.
5. When Should a Startup Build Instead of Buy?
Some tools are worth building in-house, and cheap development makes that list longer.
I believe cheap development makes sense when teams are building 4 kinds of tools:
Small internal tools whose failure poses little risk
Tools built around how a team actually works, where every commercial option forces awkward workarounds
Glue between systems that vendors leave unconnected
Temporary bridges while the market catches up
You could argue that engineers who figure out solutions to problems this big internally should just go and build their own startup.
And you’d be right. Plenty of great startups began exactly that way.
Build selectively: bridges and experiments with a written exit. Buy anything whose failure would be catastrophic.
4 questions to ask before writing code
How much breaks if it fails or gets hacked? Engineers call this the blast radius. A broken internal dashboard is an annoyance. A breached login system can end a company.
Who, by name, will maintain it three years from now?
At what point will the company switch to a commercial product? Write this down when you build, so the bridge stays a bridge.
Who has the authority to shut it down? This is the one teams forget most often. Internal tools rarely die on their own, and owning something includes the right to end it.
The companies that gain the most from cheap software will probably be the ones that build selectively.
That means using cheap development for bridges and experiments with clearly defined exits.
It also means buying anything whose failure would be catastrophic.
So keep engineers focused on what customers actually pay for.
FAQ
Is it cheaper to build software with AI than to buy SaaS?
Only for the first version, in most cases. Owning the tool is the bigger bill: developers in Stripe’s survey estimated that 17.3 of a 41.1-hour week, about 42%, goes to maintenance. AI cuts the cost of writing the code, while the patches and fixes keep landing on your team.
Do internal AI builds succeed as often as bought tools?
Less often. In MIT NANDA’s 2025 research, AI tools bought from specialised vendors succeeded about 67% of the time, and internal builds about a third of the time. The figures are self-reported, but the gap is wide.
What is comprehension debt?
It’s the gap between how much code you have and how much of it anyone understands, a term popularised by Addy Osmani. In Anthropic’s trial of 52 engineers, the AI-assisted group scored 50% on a quiz about their own code, against 67% for those who coded by hand.
Will SaaS prices fall because of AI?
So far, total spend is still rising. The average organisation in Zylo’s 2026 index spends $55.7 million a year on SaaS, up 8%, as vendors add AI tiers and usage pricing. The price pressure is landing on tools that mostly sell convenience, while software that carries risk for its customers holds its value.
When should a startup build its own tool?
When the blast radius is small and someone, by name, owns a written exit plan before the first line of code. Plan for the exit: Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027.
© 2026 The VC Corner
[email protected]