In Short
- AI coding agents are genuinely useful: I have been using tools like Claude, GitHub Copilot, and ChatGPT in my development workflow since late 2024. They have materially reduced the time it takes to build custom features — this is not hype, it is my daily experience.
- Custom applications are now affordable for small businesses: Work that would have cost $15,000–$50,000 two years ago can now realistically be delivered for $1,500–$5,000. That changes the equation for a lot of small businesses who previously could not justify the spend.
- AI agents are not magic: They make mistakes, they need experienced oversight, and they cannot replace an understanding of architecture, security, or user experience. They are powerful tools, not autonomous replacements for skilled developers.
- Browser-control agents are overhyped: Tools like ChatGPT's Operator are interesting for specific repetitive tasks but nowhere near ready for general business automation. Ignore the breathless commentary.
- The opportunity is real, the timeline is now: If you have been putting off a custom application because of cost, the economics have shifted dramatically in your favour.
The original version of this article was written by ChatGPT. It said so at the bottom, quite proudly. It also claimed AI agents could deliver "300–500% efficiency improvements" and "90% cost reductions" across your business. It read like a press release from a company trying to sell you something.
I am rewriting it because the core topic is legitimate — AI coding agents genuinely have changed how I build things for clients — but the original was exactly the kind of promotional noise I have written about elsewhere. Unsubstantiated claims, breathless excitement, zero nuance. If I am going to talk about AI, I should do it honestly.
What I Actually Use and How
Since late 2024, I have been using AI coding agents as a core part of my development workflow. Primarily Anthropic's Claude and GitHub Copilot, with ChatGPT for specific tasks. These are not novelty tools I dabble with occasionally. They are integrated into how I write code every day.
The way it works in practice: I describe what I need built — a plugin feature, a data migration script, a custom API integration — and the AI generates a working first draft. I review it, test it, identify what it got wrong, and iterate. Sometimes the first output is 90% right. Sometimes it is fundamentally wrong in ways that would break things badly if I deployed it unchecked.
The critical point that the hype cycle consistently misses: these tools amplify existing expertise. They do not replace it. I can use an AI agent effectively because I have been building web applications for over 30 years. I know what good architecture looks like. I can spot when the AI has made a security mistake, chosen the wrong data structure, or produced something that will fall over at scale. Someone without that background would not catch those problems — and the AI certainly will not flag its own errors.
The Real Cost Impact
Here is where the original article was directionally correct, even if the specific numbers were made up. Custom application development has become dramatically cheaper.
Two years ago, if a client came to me wanting to replace a complex spreadsheet-based workflow with a proper web application — something with user authentication, data validation, reporting, and integrations — I would be quoting $15,000 to $50,000 depending on complexity. That is not because I was overcharging. That is because building reliable custom software takes a lot of careful hours.
Today, I can deliver equivalent functionality for $1,500 to $5,000. The AI handles the repetitive structural work — boilerplate code, standard CRUD operations, form validation, basic UI components — in minutes instead of hours. That frees me to focus on the parts that actually require human judgment: architecture decisions, security, user experience, edge cases, and integration with the client's specific business logic.
This is not a marginal improvement. It is a fundamental shift in what is economically viable for a small business. Features and custom applications that were previously only accessible to companies with five- or six-figure technology budgets are now within reach of a sole trader or a team of five.
What AI Agents Get Wrong
I want to be specific about the failure modes, because the marketing around these tools is relentlessly positive and that does not serve anyone.
They hallucinate confidently. AI agents will generate code that references functions, APIs, or library methods that do not exist. They will present this code with complete confidence. If you do not know enough to recognise the hallucination, you will deploy broken software. This is not an edge case — it happens regularly.
They have no concept of your system as a whole. An AI agent works on the specific task you give it. It does not understand your overall architecture, your deployment pipeline, your security model, or how this piece fits with everything else. It will happily produce a solution that works in isolation but creates conflicts or vulnerabilities when integrated into the broader system.
They optimise for the immediate request, not for maintainability. Left unchecked, AI-generated code tends toward verbose, tightly-coupled solutions that work now but become difficult to modify later. Good software architecture requires thinking about the future — about what changes, what scales, what another developer will need to understand six months from now. AI agents do not think about any of that unless you explicitly force them to.
They struggle with nuanced requirements. "Build a booking system" will get you a generic booking system. The specific business rules that make your client's workflow actually work — the Tuesday exception, the capacity limits that vary by season, the integration with their existing accounting package — require human understanding and iterative refinement that AI cannot shortcut.
Browser-Control Agents: Interesting, Not Revolutionary
The original article was particularly breathless about ChatGPT's browser-control capabilities, calling them "revolutionary" and implying they could automate most business operations. Having used OpenAI's Operator and similar tools, I can report that the reality is more modest.
Browser-control agents can navigate websites, fill forms, click buttons, and perform multi-step web tasks. They are useful for specific, repetitive workflows — things like checking competitor pricing across multiple sites, or filling out the same form in multiple systems. For these narrow tasks, they save genuine time.
But they are slow, fragile, and easily confused by unexpected interface changes. They cannot handle CAPTCHAs reliably. They struggle with anything that requires judgment or context that is not visible on screen. They need constant babysitting for anything beyond the most routine tasks. Claiming they deliver "300–500% efficiency improvements" is the kind of stat that sounds impressive in a pitch deck and evaporates on contact with reality.
Will they improve? Absolutely. Are they worth experimenting with now? For specific use cases, yes. Should you restructure your business operations around them today? No.
What This Means If You Run a Small Business
The practical takeaway is straightforward, and it is genuinely good news.
If you have been running your business on spreadsheets and manual processes because custom software was too expensive, that constraint has loosened significantly. A purpose-built application that handles your specific workflow — client intake, job scheduling, inventory tracking, quoting, reporting — is now realistic at a price point that makes commercial sense for a small business.
The key is working with someone who understands both the technology and your business context. AI agents are tools that make experienced developers faster. They do not turn inexperienced people into developers. The cheapest quote you get will likely come from someone who plans to paste your brief into ChatGPT and hand you whatever comes out. That approach leads to software that works during the demo and breaks in production.
I would also caution against the consulting firms and agencies that are now marketing "AI transformation" packages. Much of what is being sold under this banner is the same repackaged common sense that the SEO industry has been peddling for years, wrapped in newer jargon. If someone promises to "AI-enable your business" without asking detailed questions about your actual workflows, they are selling a label, not a solution.
Where I Think This Is Heading
I do not have a crystal ball, but I have been watching this space closely for over a year and I will share what I observe.
AI coding agents are improving rapidly. Anthropic's benchmarks show significant jumps in code generation quality between model versions. The tools I am using today are materially better than what was available six months ago. I expect this trajectory to continue, which means development costs will continue falling.
The bigger shift is not technological — it is economic. Custom software has historically been a luxury item for small businesses. It is becoming a commodity. That changes competitive dynamics in ways that are hard to predict but broadly positive for smaller operators. The plumber who could never justify a $30,000 job management system can now get one built for a few thousand dollars. The consultant who has been managing clients in a spreadsheet can afford a proper CRM tailored to their workflow.
Google's own helpful content guidelines emphasise first-hand experience. So I will end with mine: AI coding agents have meaningfully changed my practice for the better. They let me deliver more value to clients at lower cost. But they require the same thing every powerful tool requires — skill, judgment, and honest assessment of what they can and cannot do.
If you have a business process that is costing you time and money because it is manual, repetitive, or held together with spreadsheets and workarounds, get in touch. The economics of building something better have changed, and it is worth a conversation to find out what is now possible.
