The software program trade is collectively hallucinating a well-recognized fantasy. We visited variations of it within the 2000s with offshoring and once more within the 2010s with microservices. Every time, the dream was an identical: a silver bullet for developer productiveness, a lever managers can pull to make supply quicker, cheaper, and higher. At the moment, that lever is generative AI, and the pitch is seductively easy: If delivery is bottlenecked by writing code, and enormous language fashions can write code immediately, then utilizing an LLM means velocity ought to explode.
However software program growth has hardly ever been constrained by typing velocity. The bottleneck is nearly all the time the whole lot besides typing: deciding what to construct, aligning on an strategy, integrating it into an ecosystem that already exists, getting it by safety and compliance, after which working what you shipped.
AI might help with syntax, scaffolding, and the drudgery of boilerplate. It might additionally make a unique drawback a lot worse: It makes complexity low-cost. So how will we sort out that drawback? The reply is platforms. Or paved roads. Or golden paths. Regardless of the time period, the influence is similar: by giving builders guardrails, we are able to dramatically enhance their productiveness throughout the enterprise.
Manufacturing versus productiveness
The proof to date is beneficial exactly as a result of it refuses to inform a single comforting story. A randomized managed trial from METR, for instance, discovered that skilled open supply builders, working in advanced repositories they already knew, took about 19% longer to finish duties when utilizing AI instruments, even whereas believing they’d be quicker. In a really completely different setting, GitHub reported that builders utilizing Copilot accomplished a particular, remoted programming process considerably quicker in a managed experiment and in addition felt higher concerning the expertise.
So which is it? Is AI a turbocharger or an anchor? The reply is sure, and that ambiguity is the purpose. Put AI right into a wholesome system and it will possibly compound velocity. Put AI right into a fragmented system and it will possibly compound chaos. The end result relies upon much less on which mannequin you picked and extra on the atmosphere you enable that mannequin to function in. “AI makes builders productive” shouldn’t be a device declare—or it shouldn’t be. It’s a techniques declare.
That atmosphere drawback shouldn’t be new. Years earlier than immediate engineering grew to become a job title, I argued that unfettered developer freedom was already colliding with enterprise actuality. Freedom seems like agility till it turns into sprawl, fragmentation, and an integration tax no person budgeted for. Generative AI doesn’t reverse that dynamic. It accelerates it as a result of it removes the friction that used to decelerate dangerous selections.
That is the place management groups hold making the identical elementary error: They confuse manufacturing with productiveness. Should you outline productiveness as “delivery extra code,” AI is the best invention in our lifetime. However in manufacturing, code shouldn’t be an asset in isolation. Code is a legal responsibility you will need to safe, observe, preserve, and combine. Each new service, dependency, framework, and intelligent abstraction provides floor space, and floor space turns velocity into fragility.
AI lowers the price of creating that floor space to close zero. Previously, dangerous architectural selections had been restricted by how lengthy it took to implement them. Now a junior engineer can generate a sprawling set of providers and glue them along with believable code they don’t absolutely perceive as a result of the assistant dealt with the implementation particulars. The staff will likely be happy with their velocity proper up till the primary time the system must be audited, patched, scaled, or handed to a unique staff.
At that time, the supposed productiveness win exhibits up as an working price.
If you wish to speak about developer productiveness within the AI period, it’s important to speak about supply efficiency. The DORA metrics stay a cussed actuality examine as a result of they measure throughput and stability fairly than quantity: lead time for modifications, deployment frequency, change failure fee, and time to revive. The SPACE framework can also be helpful as a result of it reminds us that productiveness is multidimensional, and “feels quicker” shouldn’t be the identical as “is quicker.” AI typically boosts satisfaction early as a result of it removes drudgery. That issues. However satisfaction can coexist with worse efficiency if groups spend their time validating, debugging, and transforming AI-generated code that’s verbose, subtly incorrect, or inconsistent with inner requirements. In order for you one manager-friendly measure that forces honesty, observe the time to compliant deployment: the elapsed time from work being “prepared” to precise software program operating in manufacturing with the required safety controls, observability, and coverage checks.
That is the half the trade nonetheless tries to bop round: AI makes the liberty drawback worse. Gergely Orosz argues that as AI writes extra of the code, engineers transfer up the abstraction ladder. The job shifts from writing to reviewing, integrating, and making architectural decisions. That feels like a promotion. Hurray, proper? Perhaps. In observe, it may be a burden as a result of it assumes a stage of techniques understanding that’s inconsistently distributed throughout a staff.
Compounding the issue, when creation turns into low-cost, coordination turns into costly. Should you let each staff use AI to generate bespoke options, you find yourself with a patchwork quilt of stacks, frameworks, and operational assumptions. It might all look effective in pull requests and unit checks, however what occurs when somebody has to combine, safe, and function it? At that time, the group slows down, not as a result of builders can’t kind, however as a result of the system can’t cohere.
Paved roads and platforms
Forrester’s current analysis hits this nail on the top. They argue that structure communities are the “hidden engine of enterprise agility.” This isn’t about re-establishing the ivory tower architects of the service-oriented structure period who drew diagrams no person learn. It’s about stopping the large tax of integration workarounds. Forrester means that with out coordination, architects spend as much as 60% of their time simply making an attempt to connect disparate techniques collectively fairly than innovating. AI, left unchecked, will push that quantity to 90%.
The answer is to not ban AI, neither is it to let it run wild. The answer is to pave the highway. I’ve written extensively concerning the want for golden paths. A golden path, or “paved highway” in Netflix parlance, is an opinionated, supported path to manufacturing. It’s a set of composable providers, templates, and guardrails that make the precise method of constructing software program additionally the simplest method.
Within the AI period, the golden path is non-negotiable. The cognitive load on builders is already too excessive; asking them to decide on libraries, fashions, prompting methods, and RAG architectures is a recipe for burnout. Your platform staff should standardize the boring components.
Think about two situations. Within the first, a developer asks an AI to construct a microservice. The AI scans the web, picks a random framework, and writes code that complies with zero of your organization’s safety insurance policies. The developer feels quick for 10 minutes, then spends every week combating the safety overview.
Within the second state of affairs, the developer is on a golden path. The AI is constrained to make use of the interior templates. It generates a service that comes pre-wired with the corporate’s authentication, logging sidecars, and deployment manifests. The code it writes is boring. It’s compliant. And it deploys in 10 minutes. On this mannequin, the productiveness win didn’t come from the AI’s capacity to write down code. It got here from the platform’s capacity to constrain the AI inside helpful boundaries.
The best builders of the subsequent decade received’t be those with essentially the most freedom. They would be the ones with the perfect constraints, to allow them to cease worrying concerning the plumbing and deal with the issue. Should you’re a growth lead, your job is to assist create constraints that allow, fairly than stifle, productiveness.
