Skip to Content

Vibe Coding for Business: Faster Delivery Needs More Discipline

What the research says, what to ask a provider, the engagement models that work, and what to expect for time and cost
November 15, 2026 by
Vibe Coding for Business: Faster Delivery Needs More Discipline
Hassan Tabbal

Vibe coding means describing what you want in plain language and letting AI write the code. In less than a year it went from a post on X to Collins Dictionary's word of the year, and most software teams now use AI every day. At NavyBits we've embraced it fully: we built our own AI Development Factory, which works around the clock on our customers' code and solutions.

Producing a lot of code quickly is no longer what holds a project back. Owning that code is the real work, because every line has to be understood, tested, secured and kept working for years, while app stores, libraries and security threats keep changing around it. This article covers what the research says about vibe coding, why faster delivery needs more discipline, what to look for in a provider, the engagement models that work, and what you can expect for time and cost.

What vibe coding means in 2026

Andrej Karpathy, a co-founder of OpenAI, coined the term in February 2025 for a way of building where you "fully give in to the vibes" and stop looking at the code. He meant it for weekend projects. By November 2025, Collins named "vibe coding" its word of the year, defined as using AI prompted by natural language to write computer code. In early 2026 Karpathy himself began calling the serious version "agentic engineering", where people direct AI agents that write, test and fix code, and engineers stay responsible for the result.

When businesses ask us about vibe coding, they mean software built mostly by AI from requirements written in plain language, with people deciding what to build and checking what comes out. That's the meaning we use here.

Where the industry stands

  • Almost every developer uses AI. In the Stack Overflow Developer Survey 2025, 84% of developers use or plan to use AI tools. Google's DORA research found that 90% of technology professionals use AI at work, for about two hours a day, and Google says 75% of its own new code is now generated by AI and approved by its engineers.
  • Few of them trust it. In the same survey, more developers distrust the accuracy of AI (46%) than trust it (33%). Their most common complaint, from 66%, is answers that are "almost right, but not quite".
  • Professionals rarely vibe code in the pure sense. 72% of developers said vibe coding isn't part of their work. They use AI heavily, and they still read and test what it writes.

Is it really faster?

It depends on where you measure. In a 2025 randomized study by METR, experienced open-source developers worked on real tasks in code they knew well. With AI they took 19% longer, although they believed they had been 20% faster. METR's 2026 follow-up suggests today's tools do save time, but it couldn't say how much.

Bain's 2025 technology report found that teams using AI assistants gain 10% to 15% in productivity, and 25% to 30% when they redesign the whole process around AI. Writing and testing code is only 25% to 35% of the time from idea to launch, so faster coding speeds up only part of a project.

Google's DORA team describes AI as an amplifier. In a team with clear requirements, good tests and small, frequent releases, it speeds everything up. In a team without them, it produces more code and more instability, and faster.

More code means more responsibility

When a person writes code by hand, the time it takes limits how much can go wrong in a week. AI removes that limit. Far more code means far more places for a mistake to hide, and someone still has to find them.

  • Quality. GitClear's analysis of 211 million changed lines found that blocks of copy-pasted code grew about eight times in 2024, and for the first time copied code outnumbered code that was reorganized and reused. Duplicated code is cheap to write and expensive to change later.
  • Security. Veracode tested code from more than 100 AI models in 2025, and 45% of the samples contained a security flaw from the OWASP Top 10. Newer and larger models were no safer. We explain the usual ones in Security Risks of AI-Generated Code and OWASP Top 10 2025 Explained.
  • Packages that don't exist. A study presented at USENIX Security 2025 found AI models recommending software packages that don't exist, at least 5.2% of the time for commercial models and 21.7% for open-source ones. Attackers now publish packages under those invented names, so the AI's suggestion installs their code. This is called slopsquatting.
  • Agents with too much access. In July 2025, an AI coding agent on Replit deleted a company's production database during a code freeze. In May 2025, a vulnerability (CVE-2025-48757) showed that apps generated with Lovable could expose their database when access rules were missing. Our guide to Supabase row-level security shows how to close that gap.

Vibe coding alone compared with disciplined AI delivery: who decides what to build, who reads the code, tests, security and what happens after launch

The ground keeps moving after launch

A system that works today won't keep working on its own. Platforms, rules and threats change every year, and every change means work on code that someone has to understand.

  • Google Play. From 31 August 2026, new apps and updates must target Android 16 (API level 36), and from 1 February 2027 updates must support 16 KB memory pages. Since September 2026, apps on certified Android devices in Brazil, Indonesia, Singapore and Thailand must come from verified developers, and the rest of the world follows in 2027.
  • Apple. Since 28 April 2026, apps must be built with Xcode 26 and the iOS 26 SDK. Apple added 13+, 16+ and 18+ age ratings, and its review guidelines require apps to tell users when personal data is shared with third-party AI and to ask their permission first.
  • Security. A record 48,185 vulnerabilities (CVEs) were published in 2025. Black Duck's 2026 audit of 947 commercial codebases found an average of 581 known vulnerabilities in each, more than double the year before, and 92% used components at least four years out of date.
  • Regulation. Since September 2026, the EU Cyber Resilience Act has required makers of software products sold in the EU to report actively exploited vulnerabilities within 24 hours, and its main obligations apply from December 2027. The law doesn't care whether a person or an AI wrote the code.

Long before AI, maintenance already made up 40% to 80% of what software costs over its life, about 60% on average, according to Robert Glass's well-known IEEE Software column. AI makes the first version cheaper. The years after it cost what they always did, and more code means more to maintain.

Think before you prompt

The most expensive mistake in an AI project happens before any code exists, when the wrong thing gets built quickly. AI follows instructions very well, including vague or wrong ones, and it rarely asks the questions a good analyst would. Who uses this screen? What happens when a payment fails halfway? Which rules will your accountant or regulator expect?

That's why serious teams now write the specification first. GitHub's open-source Spec Kit makes the written spec the source of truth and breaks the work into small changes a person can review. Anthropic's guidance for its coding agent says much the same: write the spec, give the agent tests to pass, and have the change reviewed by someone who didn't write it. Time spent understanding your business costs far less than rebuilding the wrong system.

How we use AI at NavyBits

Our AI Development Factory works around the clock, so customers see working software sooner and changes come back faster. Speed only helps when the code that comes out is code we'd be willing to maintain ourselves, and that shapes how we work.

How a disciplined AI project runs: assess, write the spec, build in small reviewed changes, test and check security, then release and maintain

  • A senior engineer leads every project. They own the specification, the decisions AI can't make and the final review. We don't hand projects to juniors, with or without AI.
  • Understanding comes first. Every engagement starts with an assessment of the business problem and of the existing system, if there is one. It ends in a written scope that the factory builds against.
  • Small changes, each one reviewed. AI output arrives as small changes that an engineer reads and tests before they're accepted.
  • Security is checked before release. Code goes through quality and security reviews before it goes live, the same kind of review we offer as our AI code assessment.
  • Your ideas stay protected. We sign an NDA before receiving any code or idea.
  • We plan for the second year. Store requirements, library upgrades and security patches are planned from the start, so the system keeps working after launch.

What to look for in a vibe coding provider

Anyone can show an impressive demo built in a weekend. These questions tell you whether a provider can deliver software you'll rely on for years.

  1. Who reads the code the AI writes? Ask for the name and experience of the engineer who reviews and approves each change. "The AI checks itself" doesn't count.
  2. How will you make sure it does what I need? Look for a written specification, tests that run on every change, and a working demo before you pay for the next phase.
  3. How do you check security? Ask which checks run on the code and its components, whether passwords and keys are kept out of the code, and whether they can list every component inside your software (a software bill of materials). The US cybersecurity agency's Secure by Demand guide has more questions worth asking.
  4. Who owns the code? The contract should give you the code, the repository and the right to take them elsewhere. In some countries, material generated only by AI may not be protected by copyright, as the US Copyright Office said in January 2025, so ownership has to be written down.
  5. Where does my data go? Ask which AI services see your code and data, under which terms, and whether the NDA covers it.
  6. What happens after launch? Ask who handles store requirements, upgrades and security patches, how quickly they respond, and what that costs per month or per year.
  7. Can I see something you maintain? A system that has run for years, with references you can call, tells you more than any demo.
  8. Will I get documentation? You should receive documentation and a handover plan, so that your own team or another provider can take over.

Engagement models that work

AI is changing how software companies charge. Forrester reported in October 2026 that providers are moving away from billable hours toward pricing based on results. Gartner, on the other hand, found that only 13% to 19% of service contracts use outcome-based pricing today, and expects fewer than a quarter to use it by 2031. In practice, these models work well for software built with AI:

  • Assessment first. A short, paid phase maps the problem, the users and the risks, and ends in a written scope and estimate. It's the cheapest place to catch a misunderstanding.
  • A fixed price for each phase. Once the scope is written, each phase (a first release, then each group of features) gets a fixed price and a demo. AI helps here, because you see a working version sooner and can judge it before you commit to the next phase.
  • Monthly capacity while the scope is still moving. If you're still discovering what you need, a fixed monthly team capacity keeps you flexible. Forrester recommends this until results can be measured.
  • A maintenance plan after launch. A monthly or yearly plan covers store requirements, upgrades, security patches and small changes. This is where the long-term cost of more code is managed.
  • You build, we review and maintain. AI made this one common. You build with AI tools yourself, as in our vibe coding guide, and hire a partner to review what you've built and look after it, including systems we didn't write.

What to expect on time and cost

  • The first working version arrives much sooner. You can see and test your idea early, and change direction before much money is spent.
  • The whole project saves less than the coding does. Coding is about a quarter to a third of the time from idea to launch. Requirements, decisions, testing, security review, data migration and training take as long as before, and review takes longer, because there's more code to read.
  • Overall gains of 10% to 30% are realistic, according to the research above, and more when the whole process is built around AI. Be careful with anyone who promises a complete business system in a weekend.
  • Building costs less, but looking after the system doesn't. Budget for review and testing during the build, and for maintenance after it. Over the life of the software, maintenance is still the bigger share.
  • AI has a running cost too. AI services are paid per use. Ask whether they're included in your price or billed separately.

Where to start

If you have an idea, start with an assessment: a written scope tells you what to build first and what it will cost. Our Innovation Lab builds web, mobile and custom software with AI, with the price quoted for your scope after the assessment. If you've already built something with AI, our AI code assessment, from $1,000, checks it for security, best practices, future maintainability and production readiness. For a first view in a few minutes, try our free self-check for AI-built software.

FAQ

What is vibe coding?

Vibe coding is building software by describing it in plain language and letting AI write the code. Andrej Karpathy coined the term in February 2025, and Collins Dictionary made it its word of the year for 2025.

Is vibe coding safe for business software?

It can be, with discipline. Research has found security flaws in close to half of AI-generated code samples, so every change needs review by an experienced engineer, automated tests and security checks before it goes live.

How much faster is development with AI?

Independent research puts the overall gain at around 10% to 30%. The first version arrives much sooner, but requirements, testing, review and launch still take time.

Does vibe coding make software cheaper?

The build usually costs less. Maintenance, which is 40% to 80% of what software costs over its life, doesn't shrink, and more code means more to maintain. Plan for both.

Who owns code written by AI?

Your contract should say so. Copyright in material generated only by AI is uncertain in some countries, so make sure the agreement assigns the code and the repository to you.

Can NavyBits review an app I built with AI tools?

Yes. Our AI code assessment checks security, best practices, maintainability and production readiness. The report documents each problem with how to reproduce it and how to fix it.

Talk to an engineer before you build

Tell us what you want to build, or show us what you've already built with AI. In a free 30-minute call, a senior engineer will tell you what it takes to make it ready for real customers, and to keep it that way.

Vibe Coding for Business: Faster Delivery Needs More Discipline
Hassan Tabbal November 15, 2026
Share this post
WhatsApp usAI agent replies 24/7
Book a free 30-minute call