AI agents can write an app before lunch, but they cannot decide what to build, check that it is safe, or answer to the Regulator when it leaks. That is still a person’s job.
- Trust the agent with scaffolding, repetitive changes, first drafts, tests and options
- Keep a human on architecture, the data model, security, money, live systems and what not to build
- Almost half of AI-written code fails basic security tests, and you carry the POPIA risk
- A proper build still takes weeks: discovery, planning, review, testing and hand-over
Type “build me a website for my plumbing business” into an AI tool today and you will have something that looks like a website within minutes. Ask an AI agent for a booking app and it will happily write thousands of lines of code before lunch. That is remarkable, and it has changed our industry for good.
It has also created a new problem. Anyone with a laptop and a subscription can now call themselves a developer. The results often look finished. Many are not. The gap between “it works on my screen” and “it is safe to run a business on” is where the real cost hides.
Here is a straight-talking look at what AI agents are brilliant at, what they should never decide alone, and why proper agentic development still takes weeks, not minutes.
What changed, and why it is good news
Agentic development means AI that does more than suggest a line of code. An agent can read a whole project, plan a change, write it across many files, run it, test it and fix its own mistakes, all from a plain-English instruction.
For business owners, that is good news. Work that used to take a developer days now takes hours. The repetitive building blocks every app needs (forms, tables, admin screens, the wiring between them) no longer eat the budget. That puts custom software within reach of businesses that could never have justified it five years ago.
Speed only moves the work. It does not remove the thinking.
Someone still has to decide what gets built, check that it is right, and take responsibility when it goes live.
What to trust the AI with
Hand the agent work that is repetitive, well understood and quick to check.
- Scaffolding Project setup, standard pages, forms and admin tables.
- Repetitive changes Renaming a field across 40 files, updating a library, rolling a design change out everywhere.
- First drafts of features A working starting point that a developer then reviews and shapes.
- Tests and documentation The checks and notes humans tend to skip.
- Exploring options Mocking up two or three approaches quickly, so the right one gets picked.
The common thread: the result is fast to check. If an experienced person can look at it and know within minutes whether it is right, the agent is a superb assistant.
What still needs an experienced human
Keep a person in charge of anything where a mistake is expensive, invisible or permanent.
- Architecture How the system fits together decides what every future change costs. Agents solve the task in front of them, not the problem you will have in year three.
- The data model How your customers, orders and invoices relate to each other. Get it wrong and every report, integration and upgrade fights you.
- Security and access An agent will happily build a login screen. It will not reliably make sure customer A can never see customer B’s records.
- Money and personal data Payments, refunds and POPIA need review by someone who has seen them go wrong.
- Live systems Real databases, deployments and backups. Agents should work on copies, never the real thing.
- What not to build The most valuable decision in a project is often cutting a feature. AI will build whatever you ask for. It will not tell you it is a bad idea.
- The user experience Knowing how your staff and customers actually work comes from talking to people, not from code.
Why “everyone is a developer now” is risky
AI-written code looks finished long before it is safe. Building software by prompting an AI and shipping whatever comes out even has a name: vibe coding. It was Collins Dictionary’s Word of the Year for 2025. The trouble is what the research keeps finding underneath.
- 45%of AI-written code failed basic security checksVeracode, 100+ models, 80 tasks
- 31.6%was exploitable exactly as writtenIOActive, almost 20,000 samples, 2026
- 75%+of developers believed AI code was more secure than their own56% admitted it often introduced problems
And it is not theoretical:
- The database that disappeared SaaStr founder Jason Lemkin spent nine days building an app with an AI agent. Despite repeated instructions not to change anything, it deleted the live database holding 1,206 executive records and over 1,196 company profiles, then claimed recovery was impossible. It was not, and the data was eventually recovered.
- Apps with the front door open A scan of 1,645 apps built with one popular AI app builder found 170 whose databases anyone could read, including personal and financial records.
- “I didn’t write a single line of code.” That was the founder of Moltbook, a platform that went on to expose 4.75 million records in January 2026, including 1.5 million access tokens and 35,000 email addresses.
In South Africa, a leak like that is not just embarrassing. The business that collects personal information is responsible for protecting it, and the Information Regulator can impose fines of up to R10 million. The person who built your app over a weekend of prompting is rarely the one answering to the Regulator. You are.
What gets overlooked
None of these show up in a demo. All of them show up eventually.
- Database rules, so one customer can never see another customer’s data
- Secret keys kept off the public side of the site
- Separate test and live versions, so experiments never touch real data
- Automatic backups that have actually been restored at least once
- Spam and bot protection on every form
- POPIA basics: a privacy policy, consent, and knowing where data is stored
- Email set up properly, so your enquiries do not land in spam
- SEO foundations: page titles, descriptions, a sitemap and structured data
- Accessibility, so everyone can use it
- Speed on a mid-range phone over mobile data
- Analytics, so you know what is actually working
- Who owns the code, the domain and the hosting accounts
- Who fixes it when it breaks, and how quickly
Why proper agentic development still takes time
AI made the typing fast. It did not make the deciding, checking and testing fast.
Ask an AI to “build my entire website in HTML” and you will get a page that looks the part in minutes. Here is the difference between that and a site built properly:
| One prompt | Built properly | |
|---|---|---|
| Updating content | Edit the code, or prompt again | Change it yourself in a simple admin |
| Contact forms | Often go nowhere | Delivered, spam-protected and tracked |
| Search engines | Basic or missing | Titles, schema, sitemap and local SEO |
| Security | Untested | Reviewed and tested |
| Hosting and backups | Up to you | Set up and monitored |
| Next year’s changes | Start again | Built to grow |
| Look and feel | Generic | Designed around your business |
Where the time actually goes:
- Discovery Understanding the business, the users and the goal before anything is built.
- Planning The data model, the structure and the order things get built in.
- Building in small steps Agents work in short, reviewed chunks, never one giant prompt.
- Review Every change read by a person who understands what it does.
- Testing On real devices, with real data, including the ways people will misuse it.
- Launch and hand-over Hosting, backups, monitoring and training.
An agent might write the code for a feature in an hour. Making sure that hour does not cost you a weekend of downtime six months later is the job. It is also why a custom web app still costs what it costs: the price was never the typing.
How to choose who builds it
Ask any developer or studio these questions before you sign.
- Who reviews the code before it goes live?
- Where is my data stored, and who can get to it?
- How are the test and live versions kept apart?
- How are backups done, and when was one last restored?
- Who owns the code and the accounts when the project ends?
- What happens when something breaks at 9pm on a Friday?
- Can I see something similar you built that is still running today?
A good answer is specific and calm. A vague one, or “the AI handles that”, is your answer.
The bottom line
AI agents are the best thing to happen to custom software in years. They make good developers faster and put custom builds within reach of more businesses. But they are a power tool, not a builder. The value is still in the person holding it: knowing what to build, what to check and what can go wrong.
Thinking about a custom app or website?
A few questions, and a person comes back to you, usually within one working day.
- Cloud Security Alliance, Vibe Coding Security Debt (April 2026): Veracode figures, the developer survey, the app builder scan, Moltbook and Collins Word of the Year.
- IOActive, The Security Gap in AI-Generated Code (May 2026).
- SaaStr, Jason Lemkin’s write-up of the deleted database.