← Back to all articles
strategy

Lovable or a Custom App? When It Makes Sense to Move to a Professional Provider

Most founders who ask me whether it is time to leave Lovable are asking too early. A few are asking months too late, and those are the ones whose first paying customer has already found the problem for them.

I answer that question from the provider's chair, so weigh what follows with that in mind. At Conectia we take over apps built with Lovable, Bolt, Replit, Cursor and v0 and run them for real customers; before that I spent years in enterprise DevOps and incident response. I have already written about what happens after you vibe-code an app: the mornings on which a demo turns into an obligation, and the go/no-go that follows. This post sits one level above it. It is for the founder or business owner holding a working app who wants to know whether to move it to a professional provider, when, and how to do it without losing control of what they built.

My answer fits in one line. You move when the app takes on obligations the tool cannot answer for, and you stay for as long as it has none. How good the generated code looks comes second.

The tool stays the right choice for longer than providers like to admit

I will start with the case for hiring nobody, us included, because it is stronger than my industry pretends.

If the app has no paying users, holds no personal data beyond an email address, and you are still changing what it does every week, the tool that got you here is the right tool. A prototype has one job, to find out whether someone wants the product, and Lovable or Bolt does that job in days for the price of a subscription. A provider at this stage would spend most of your budget hardening screens you will throw away at the next pivot. An internal tool for five colleagues, a landing page with a form, a demo for a first investor meeting: none of them needs a contract with anyone.

This stage has a ceiling, and you can recognise it. The tool starts to forget. Bolt's own documentation warns that once a project is large, a long prompt adds to a context window that is already close to full. Every prompt fixes one screen and breaks another, and you spend Thursday getting the app back to where it was on Monday. In Stack Overflow's 2025 survey, 66% of developers counted «AI solutions that are almost right, but not quite» among their frustrations with AI tools, the most cited of all. Hitting that ceiling calls for cleaning up the project, splitting it, or bringing in one experienced pair of hands for a week. By itself it does not call for a new provider.

When the app outgrows the tool, six signals show it

The tool vendors draw the line themselves. Lovable's security documentation says you are responsible for ensuring that your app meets the security requirements appropriate for its use case, and for apps that handle sensitive data or critical functions it recommends «an additional professional security review». Supabase, the backend under most Lovable apps, publishes a shared responsibility model that leaves the access rules on tables with sensitive data to you. The open question is when «you» has to be someone who has done this before. These are the six signals I look for.

  1. Paying users. Money brings refunds, chargebacks and a payment provider that retries webhooks in whatever order the network chooses. It also brings the message at eleven at night from a customer who paid and cannot log in. Somebody has to own that message, and a subscription to a builder includes nobody on call.
  2. Personal data and real authentication. The day a second company's records land in the same tables as the first one's, who can read what becomes a legal question. Inside the builder, the answer is a set of database policies nobody has tested against a hostile request.
  3. Integrations with systems you do not control. A CRM, an ERP, a customer's single sign-on, a payment gateway. Each one changes and fails on its own schedule, and each needs a test environment that cannot touch production data. Replit only separated development and production databases on 21 July 2025; until then, an app built there used one database for both. Integrations are where a single environment stops being acceptable.
  4. Compliance that someone will check. A security questionnaire from a larger customer, a data processing agreement under Article 28 of the GDPR, a HIPAA business associate agreement, the security certification a public administration demands of its suppliers. The questions are plain: where the data lives, who has access, how keys rotate, how you would restore yesterday's backup. They need an owner who can answer them in writing.
  5. A raise. Technical due diligence asks who owns the repository, the cloud accounts and the intellectual property, whether anyone besides the founder can maintain the product, and what it would cost to run at ten times the users. «It lives in my Lovable workspace» is an answer, though rarely the one the investor hoped for.
  6. A team. The second engineer needs branches, code review and a pipeline that runs the tests before anything ships. Lovable's GitHub integration edits and syncs one branch at a time. That is the right design for one person and a prompt box; for three people working in parallel it becomes the queue everyone waits in.

The rule I give founders is short. One signal is a reason to get a written verdict on what you have. Two signals, above all money plus personal data, are a reason to move. With none, keep building.

A provider is one of four ways forward, and not always the best

The other three are sometimes the better call.

Keep iterating, with guardrails. With one mild signal, stay in the tool and close the obvious gaps. Connect the code to a GitHub organisation you own, run Lovable's deep scan before each release (it does not run on its own), and switch on the workspace setting that blocks publishing while critical findings are open. That is an afternoon of work.

Hire a freelancer for a scoped job. When the problem lives in one module, the payment webhook or the email sender, a good freelancer fixes it faster and for less than any company. I concede that without hedging: for one fix by Friday, the freelancer is the better call. The risk arrives later, when the person who understands the payments module is on another project and the accounts are in their name.

Hire a provider to harden what exists. Once two signals have appeared, this is the most common right answer. In most of the apps we read, the screens and most of the data model survive, and the work goes into what surrounds them: authentication, access rules, payments, separate environments, tests on the paths that move money or personal data, monitoring, and a backup someone has restored at least once. Veracode tested code from more than 100 language models and found that AI-generated code introduced risky security flaws in 45% of tests. Hardening therefore starts by reading the code, line by line where money moves.

Hire a provider to rebuild it as a custom app. A rebuild is right in fewer cases than founders fear: when the data model is wrong for the business, when the product needs what the platform cannot give it (native mobile, heavy background jobs, a backend in your own cloud for a regulated customer), or when no module can be tested without being rewritten. Then the prototype becomes the specification, and custom app development starts from a working example instead of a slide deck. When the core has to be rebuilt, we do it as an MVP in 6 weeks at a fixed price.

Harden or rebuild is rarely one decision for the whole app. You make it module by module, and the module that most often gets rebuilt is authentication.

The cost of moving is mostly decided in the first two weeks

Founders ask me for a number, and the number depends on choices most of them have not made yet. The cost has three parts.

The verdict. Someone has to read the code, the database and the cloud console and write down what to keep. Our Diagnostic & Blueprint takes two weeks with read-only access and ends in a verdict per module and a plan with a closed price for the next phase. If we build that phase, what you paid for the diagnosis comes off the MVP; if you build it elsewhere, every file stays yours.

The work. Hardening takes days per module once the verdict has named the problem; a rebuild takes weeks. Both are sized from the verdict, which is why I distrust any quote given before somebody has read the code. I set out the arithmetic of hours, rates and team size in how much an app costs to build, and it applies here unchanged.

The running cost. Most budgets miss this line. Hosting moves to accounts you pay for, someone watches the error alerts, dependencies get updated, and the model calls get a spending cap. After the build, that line never goes away. Leaving it out makes the tool look cheaper than it is and the provider dearer than it is.

How I would start the move this month if the app were mine

If the app were mine and I had just seen the second signal, in this order:

  1. Count the signals in writing. Paying users, personal data, integrations, compliance, a raise, a team. The count sets how urgent everything below is.
  2. Put every account in your own name. GitHub organisation, Supabase organisation, cloud account, Stripe, domain, email sender, app store listings. A provider should be invited to each of them and removable in a minute. If a freelancer or an agency holds any of these today, fix that first.
  3. Export the code to a repository you own and check that it builds outside the tool. A copy that only runs inside the builder is yours on paper and nowhere else.
  4. Pay for a discovery before you pay for a build. Two weeks, read-only, a written verdict per module with the measurements next to it. A provider who wants to start building before reading has already decided to rebuild.
  5. Sign a phased contract. Each phase with a closed scope, a closed price and acceptance criteria you can test, and an exit at the end of every phase with a handover of code, accounts and documentation. That clause protects you from us as much as from anyone else.
  6. Keep the tool for what it does well. Sketching the next feature in Lovable or v0 and handing it to the team as a specification is a sound workflow. Leave the context files (CLAUDE.md, AGENTS.md) and the tests in the repository, so Cursor or Claude Code keep building inside the plan.

The tool found your customers; the move is how you keep them

Lovable, Bolt, Replit, Cursor and v0 did the hardest part in a weekend: they proved that someone wants what you built. Leave them too early and you pay professionals to polish a guess. Leave them too late and your customers run the security review for you. The six signals tell you which side of that line you are on.

If you are already past it, the first step of how we take an AI-built app to production is the two-week Diagnostic & Blueprint: a written verdict on your app that you can carry out with us, with your own hire or with the tool you started with.

Work with the authors

The CTOs who write this also build

Fractional CTO engagements, AI engineering squads and an AI development platform we install on your codebase in 24 hours. If this piece matched how you think, the conversation is short.