Building Software Faster: How We Work with Claude Code and AI Coding Agents
Digital Airways·
Two years ago, the question «Should we have our own software built?» was quickly answered for many Swiss SMEs: too expensive, too slow, too risky. That calculation has shifted. Not because developers have suddenly become superfluous, but because the way software is created has changed noticeably.
The reason is called the AI coding agent. In this article we describe openly how we at Digital Airways work with such tools, where they genuinely save time, where they do not – and what that means for you as a client in concrete terms.
What distinguishes an AI coding agent from code completion
The first generation of AI aids in programming suggested individual lines while the developer typed. Useful, but essentially a better autocomplete.
An AI coding agent works differently. It receives a task in natural language – for example «Add a check to the order form that the delivery date is not in the past» – and then completes several steps independently: it searches the project for the relevant files, understands the existing structure, writes the change, runs the tests and corrects itself when something fails.
Well-known examples are Claude Code (Anthropic), ChatGPT Codex (OpenAI) and agentic features inside development environments. Viewed in a vendor-neutral way, they solve the same basic problem: they shift the developer’s work from «writing lines» towards «describing, checking and taking responsibility for a result».
Which assistants suit which task is covered in more depth in our article Claude, ChatGPT & Copilot Compared.
Where the time is actually saved
The biggest misconception in the public debate is the assumption that AI makes programming «ten times faster» across the board. The realistic picture looks different: the time gain is distributed very unevenly.
| Activity | Effect of AI agents |
|---|---|
| Scaffolding, forms, standard views | Very high – hours instead of days |
| Understanding and documenting existing code | Very high – especially with legacy systems |
| Writing tests and narrowing down bugs | High |
| Defining the data model and architecture | Low – remains headwork |
| Clarifying business requirements | None – this emerges in conversation with you |
| Security, permissions, data protection | Low – requires human review |
The pattern is revealing. Everything that involves a lot of routine work and few decisions becomes dramatically faster. Everything that demands judgement, contextual knowledge and responsibility stays largely the same.
That is precisely why the developer’s role does not disappear – it shifts. Those who develop well today spend more time on precise task formulation, code review and quality assurance, and less on typing out familiar patterns.
How we work in practice
A typical development step runs through four phases for us:
- Sharpen the task. Before the agent starts, the requirement is written down: what should come out of it, which rules apply, what must not change? This care is not bureaucracy – an unclear task leads an AI agent to a wrong result just as reliably as it does a human.
- The agent implements. It works within a defined area of the project, has access to the existing code and the tests, and proposes a complete change.
- A human reviews. Every line that goes into a client project is read and understood by one of our developers. That is non-negotiable. An agent can deliver a solution that works and is nevertheless wrong – for instance because it does not know a special case in your business.
- Test and roll out automatically. Tests and deployment run through established processes as before. The agent does not change how carefully we release.
The effect shows up above all in the early phase of a project. A working prototype that you can handle and comment on is often created in days rather than weeks today. And an early prototype is the best protection against the most expensive kind of mistake: software that does exactly what the specification said, but not what the business needs.
What this means for you as a client
For Swiss SMEs, four practical consequences follow.
Custom development becomes realistic for smaller projects. Undertakings that previously failed the economic viability test because of the effort involved now add up. This also shifts the trade-off between standard software and a solution that fits your process exactly.
Interim results arrive earlier. You see something workable sooner – and can intervene sooner. Use that actively: feedback in week two is many times cheaper than feedback in month four.
Quality still depends on the team, not the tool. AI agents are available to everyone. The difference between a solid and a fragile solution arises in the architecture, in testing and in review – exactly where the agent contributes least. So do not ask a provider whether they use AI, but how they ensure quality.
Operating it afterwards remains your concern. Software built quickly needs just as much maintenance, updating and understanding as software built slowly. Clarify from the outset who owns the code, where it lives and who can develop it further in two years’ time.
Data protection: what happens to your code?
A legitimate objection runs: if an AI agent sees our code – where does it go? Since the revised Data Protection Act (revDSG), the question is rightly present in Switzerland.
Three points help with the assessment:
- Business plans differ clearly from free offerings. With the paid business and enterprise variants of the major providers, inputs are generally not used to train the models. That is the framework in which professional development should take place.
- Not all code is equally sensitive. An order form is something different from a data export containing personal data. Test data instead of live data is good practice in development anyway – with AI agents this rule becomes more important, not new.
- Responsibility stays with the service provider. Whoever builds your software must be able to tell you which tools are involved and under what conditions. Ask about it.
What AI agents (still) cannot do
Honesty demands the other side too. In our daily work we regularly hit the same limits:
- They do not know your business. Why an invoice requires two approvals at your company and who decides in exceptional cases is written in no repository.
- They often make architecture decisions that are plausible but short-sighted – viable for today, impractical in two years.
- They are confident even when they are wrong. Without review, that is a real risk, not a theoretical one.
- They are no substitute for a conversation. The most valuable hour in a project is usually the one in which someone watches on site how the work is actually done.
For smaller automations without in-house development, no-code approaches such as the Microsoft Power Platform remain a sensible route – AI agents and no-code are not mutually exclusive, they cover different needs.
Conclusion
AI coding agents such as Claude Code or ChatGPT Codex have genuinely changed software development – but differently from what the headlines suggest. They accelerate the routine work considerably and leave the demanding parts largely untouched. For Swiss SMEs this means: a tailor-made solution is achievable in less time and on a smaller budget than two years ago – provided someone continues to take requirements, review and operations seriously.
Are you wondering whether your own application is worthwhile for one of your processes? In a free initial consultation we look at your workflow and tell you honestly whether the effort pays off – including when the answer is no.
