We use AI everywhere in our build process. Lovable for AI-augmented frontend code, Cursor for backend logic, Claude and GPT for specs, ADRs and content, custom agents for QA and integrations. The honest accounting of what AI accelerates inside our studio is somewhere around 60%. The remaining 40% — the part that decides whether your project actually ships and works in production — is still human.
That ratio matters because it determines what you can expect from any agency that calls itself AI-native. If you understand which 60% the AI is doing and which 40% it isn't, you can tell the difference between a studio that uses AI as leverage and a studio that uses AI as marketing.
What AI does well in our build process
AI shines at scaffolding, repetition, and translation. Generating a CRUD interface from a schema, writing the boilerplate for a Supabase RLS policy, drafting copy for an empty state, converting a Loom recording into a written spec, translating English UI strings into German, summarizing a 90-minute Discovery call into action items — these tasks used to take hours. They now take minutes.
The pattern is consistent: anything that has been done thousands of times in public training data, anything that follows known conventions, anything that translates between formats. AI hits 80–95% quality on the first pass. We review, correct the edges, and ship. The leverage is real and it compounds across a project.
What this looks like in practice
On a typical Custom Build engagement, AI saves us roughly two days per week of work that would have been mechanical. We don't pass the savings on as a discount — we pass them on as more design iterations, more integration tests, and more time spent on the parts of the build that actually need a human.
What AI does poorly — and why we don't ship it unsupervised
AI is bad at judgment, context, and consequences. It will confidently propose a database schema that fails the second you try to filter on a multi-tenant column. It will write a Stripe webhook handler that looks correct and silently double-charges customers when retries fire. It will translate a marketing page into German that is grammatically perfect and culturally tone-deaf.
The failure mode is consistent too: AI is good at the average case and bad at the specific case. Your business is the specific case. Every Discovery we run surfaces three or four constraints that no model could have inferred from public data — the way your team handles handoffs, the regulator your industry answers to, the unwritten contract you have with your largest client. AI cannot encode that. Humans have to.
Where the 40% sits
Architecture decisions, security review, integration design, copy that has to land for your specific buyer, anything that touches money or compliance, the conversation in week one of Discovery that decides what we're actually building. We don't let AI ship any of those without a human signing off.
How we structure human review
Every deliverable in our process passes through at least one human review before it reaches a client. Code is reviewed by a senior engineer before it goes into a pull request. Copy is reviewed by an editor before it lands on a page. Schemas and RLS policies are reviewed against a security checklist before any data goes live. Integration tests are designed by a human, even when the test code is AI-generated.
The review isn't a rubber stamp. Roughly 30% of AI-generated work gets meaningful corrections — sometimes a small fix, sometimes a complete rewrite. We track this internally because it tells us where our prompts and our agents are weakest, and it keeps us honest about what AI can and cannot ship without supervision.
What to ask any AI-native agency
Ask them what their human review process looks like. Ask who specifically reviews AI output before it reaches you. Ask what happens when a senior engineer disagrees with what the model produced. If they don't have a clean answer, you're paying for AI to be your developer — not for an agency to be accountable for the result.
Why "vibe coding" doesn't ship production systems
There is a popular pattern right now of building entire apps by chatting with an AI in plain English. It works well for prototypes, demos, weekend projects, and internal tools where the cost of bugs is low. We use it for that ourselves — Lovable is a fantastic prototyping environment and we ship internal tools with it routinely.
But production systems for businesses serving real customers fail differently. They have to handle edge cases the prompt didn't think of, scale beyond the data the model was trained on, integrate with APIs whose documentation lies, and survive a year of changes by a team that wasn't in the original conversation. "Vibe coding" handles none of that. It produces code that works on the happy path and breaks the moment reality intrudes.
Where the line is
Use AI to generate a working draft. Have humans take that draft and make it production-grade — write the tests, harden the edge cases, document the architecture, plan for change. The hand-off from prototype to production is exactly where most AI-built projects die. Treating it as a serious engineering step is most of what "AI-augmented" actually means.
What you should expect from us
Concretely: faster delivery than a traditional agency, similar or better quality, and a clear answer when you ask what AI did versus what we did. Our Discovery week is mostly human — we listen, we map your operations, we ask questions no model could ask. Our Build weeks are AI-leveraged — we ship two to three times faster than we did before AI tools existed. Our QA and security review are mostly human, with AI assistance for repetitive checks.
The pricing reflects the leverage. We don't charge what an AI bot would charge, because what you're buying is judgment, accountability, and a team that will own the outcome. We don't charge what a traditional agency would charge for the same scope, because we don't need to. The 60% acceleration is yours.
If you want to see this in action
Book a Discovery Call. We'll walk you through a recent build, show you exactly which parts of it AI accelerated and which parts a human had to do, and explain why we structured the engagement that way. It's the easiest way to tell whether "AI-augmented" means something real to us, or whether it's marketing.
The honest summary
AI is a leverage multiplier on talented humans. It is not a replacement for them. The agencies that are getting good results from AI are the ones that already had good engineers, good process, and good taste — AI is making them faster. The agencies that are using AI to paper over weak fundamentals are producing the kind of work that looks fine in a demo and breaks in production.
Sixty percent AI is honest about both halves: the leverage is real and worth paying for, and the human 40% is what makes the leverage safe to use. If anyone tells you they're 100% AI, they're either selling you a prototype or hiding the humans who are quietly fixing the model's mistakes behind the scenes.
If you want to see what this looks like applied to a real engagement, book a Discovery Call. We'll show you the actual workflow, name the tools, and walk through where humans intervene. It's a 30-minute conversation, no obligation, and you'll come away with a much clearer view of what AI-augmented development can — and cannot — do for your project.