AI Writes Code. Engineering Ships Software.

A board is reviewing a technology proposal. The company is a private-equity-backed services operator that needs a software platform to support its exit, and the vendor's quote is $350,000 over six months of AI-assisted engineering. Two years ago that same platform ran closer to $800,000 and eighteen months. The price dropped because the tooling is real. A board member asks the question everyone asks now: "How much of this is AI?" "All of it," the CEO says. The board approves.

Three weeks in, a technical advisor opens the architecture documents. The data isolation design will not hold two enterprise clients who compete with each other. The security model assumes a single tenant. The portability work, the part that lets a buyer take ownership of the code at exit, is simply not there. None of it showed up in the proposal, because none of it shows up in a demo.

The board funded the part of the platform AI makes cheap. The part that sets what the platform is worth was never on the page, and nobody in the room knew to ask for it.

AI was not oversold to that board. The tools are genuinely changing how software gets built. The problem is narrower: the people approving the capital usually cannot tell which part of the work the AI is doing, and that blind spot is enough to fund the wrong build. A simple framework fixes it, and it holds against every proposal a board will see in 2026.

What AI Actually Accelerates

Give the productivity case its due, because it is real. GitHub Copilot crossed 20 million users in 2025 and runs inside 90 percent of the Fortune 100.¹ In Stack Overflow's 2025 survey, 84 percent of developers said they use or plan to use AI tools, and 42 percent of committed code is already AI-assisted.² McKinsey found that teams which rebuilt their process around AI cut pull request cycle time from 9.6 days to 2.4 days.³ Gartner expects 75 percent of enterprise software engineers to be using AI code assistants by 2028, up from under 10 percent in early 2023.⁴ Building software has gotten faster and cheaper, and AI is the reason.

What matters is where those gains land. In GitHub's own controlled study, developers using Copilot finished the task 55 percent faster, in 1 hour 11 minutes against 2 hours 41 minutes.⁵ The task was to build an HTTP server, a well-scoped job with one right answer. That is the work AI is best at, and most of a platform's visible layer is built from it: the login screen, the dashboard that reads from a table everyone already agreed on. It is real, and it is the least valuable layer of the platform.

The most rigorous independent study points the same direction from the other side. In a 2025 randomized controlled trial, METR had experienced developers work on the mature codebases they knew best. With AI allowed, they took 19 percent longer, even after predicting a 24 percent speedup and then, afterward, swearing they had gotten one.⁶ The slowdown clustered where the work was hard and the quality bar was implicit. METR's 2026 follow-up suggests newer tools have likely closed that gap, though the measurement got messy enough that the team is rebuilding the study.⁷ Read carefully, the finding is narrower than the headline. AI helps most on well-scoped work. It helps least, and sometimes hurts, on the architectural decisions that ripple through everything built after them.

The Two Layers of Every Platform

Every software platform has two layers, and the gap between them is the whole decision.

The first is the one you can see: the screens and workflows a user touches and a demo shows off. This is the layer AI compresses best, and it is the layer a board looks at when it reviews a demo.

The second is the one nobody sees, and it is the one a buyer pays for. Five systems live down there, and each has to be built from the first sprint, because none of them can be bolted on cleanly later.

Data isolation keeps competing clients' records physically apart. Without it, the platform cannot hold two big customers at once, and the growth story behind a premium multiple evaporates. Security architecture is the hardened storage and authentication built against known attack patterns, and it is where AI-built code tends to break: Veracode's 2025 testing found roughly 45 percent of AI-generated code shipped with security flaws, almost three times the human rate.⁸ The bill for those flaws grows the longer they sit, and a benchmark long tied to IBM puts a fix in testing at about 15 times the cost of catching it in design.⁹ Portability is the clean, IP-owned codebase that moves to a buyer's infrastructure without a rebuild; when it is missing, that gap becomes a price cut at the closing table. Audit trails are transaction records built into the data model to survive legal and compliance review, not logging stapled on at the end. Scalability is the headroom to take ten or a hundred times the load without starting over, the same growth the buyer is underwriting.

None of the five show up in a demo. All of them get pulled apart in the first week of diligence, where the buyer's technical team usually opens the authentication code first and can tell within a day whether every endpoint was secured the same way or each one was improvised. AI can help build any of these systems, but it decides none of them on its own, and code shipped without an architect making those calls comes apart in the data room.

The Right Question

So the wrong question is the one the board asked. Does this use AI? They all use AI. The questions that separate a real build from a demo are more specific, and any non-technical director can ask them.

What, exactly, does the AI-assisted engineering cover? AI covers implementation, and architecture is not implementation, so if the answer is everything, ask again. Who is the named architect accountable for data isolation, security, and the portability that lets a buyer take the code? If that person is the tooling, you have your answer. If it is someone with a track record, you are having a different conversation. What will this platform look like after a buyer's technical team has been through it? A vendor who cannot walk you through that has not built the second layer. And which milestones track exit-readiness, not features? If every milestone is a feature, nobody is measuring the foundation.

Price the architect against the exit, not against the build. A technology-enabled business can exit near 8 to 12 times EBITDA where a pure-service operator exits near 3 to 6 times.¹⁰ On an illustrative $5 million of EBITDA, the midpoint gap between those two outcomes is about $25 million. A senior architect costs more than an unattended AI build. That difference is the premium on protecting the $25 million, and the second layer is what the acquirer is paying for.

The cheap AI-assembled platform is not a discount on the same outcome. It is a different outcome that happens to look identical in the room where the money gets approved.

Which brings it back to the board. The next proposal lands, and it sells AI the way the last one did: faster, cheaper, AI all the way down. This time someone has the framework. They ask what is in the 30 percent the AI is compressing, and who is accountable for the 70 percent it is not. The answer decides whether the board is buying the platform it thinks it is.

Prepared June 2026. All figures validated against the live sources cited below.