Right now we either seem to be being told that AI is the answer to everything, or that it is completely over hyped; even deeply problematic, considering things like AI slop, AI projects that have failed to deliver, AI powered fraud, deep fakes and all the rest.
There are also multiple products popping up with flashy marketing telling people they can ‘vibe code’ a software startup, just using prompts. And yet you’ll notice that developers still have jobs.
Where is reality in this emerging and fast-changing world? As a software development agency, we believe we can offer a balanced perspective.
For a start, there are definitely tasks that Machine Learning (ML) or Large Language Models (LLM’s) can do very well, but they certainly aren’t yet the solution to every problem.
In fact, one slightly confusing aspect is that these new tools are often somewhat the opposite of what we’d gotten used to computers being good at. Forever in a day computers excelled at doing very specific tasks super accurately, over and over again. Now, AI’s have expanded what computers can do into much fuzzier and more human-like realms, but the “super accurate” part doesn’t always apply.
This means that, for any given requirement, traditional vs AI approaches still need weighing up.
What we see working
- Staff using LLM’s to find and collate information, analyse large documents, produce draft content, convert data into report formats and much more.
- Chatbots in websites, software and apps that are context and subject aware and therefore able to provide meaningful help to users.
- Coding assistants that speed up software development.
- AI scheduling and note taking tools.
- Vibe coded (prompt driven) prototypes of digital products.
- Industry-specific applications where there’s high-quality data. E.g., matching rules in a workflow, extracting fields from forms, or tailoring recommendations.
What we see failing
- Hype-driven claims like “vibe code” that promise you can ship a reliable product from prompts alone.
- Overreliance on off-the-shelf AI models for specialised problems leading to hallucinations, inconsistent output and privacy leaks.
- AI automation projects that underestimate the need to properly prepare and clean data.
So how should Kiwi mid-market companies approach AI?
- Start with the problem, not the tech. What is the pain you’re solving? If AI isn’t the most direct, maintainable or cost-effective route, don’t use it.
- Measure before you build. Define clear KPIs for any experiment; time saved, conversion lift, error reduction, cost per transaction. If you can’t measure it, you can’t prove its value.
- Check your data. Is it clean, representative and accessible? Will you own and control it? If not, fix data plumbing first; models are only as good as the data they’re fed.
- Build small, then iterate. Proof-of-concept → controlled pilot → production release. At each step check reliability, latency, cost and user feedback.
- Plan for ops and maintenance. Models drift. Prompts need tuning. Logging, monitoring, retraining and a rollback plan are non-negotiable.
- Keep humans in the loop. For many business processes, a human + model is far better and safer than fully automated decisions.
- Think about risk and compliance early. Privacy, IP, explainability and the risk of deepfakes or biased outputs should inform design choices, vendor selection and contract terms.
Vendor tooling and hosted models
There are fantastic off-the-shelf APIs and hosted LLMs that can dramatically lower the barrier to entry. They’re great for prototyping and for features that don’t require sensitive data. However if your use case touches confidential customer information or material IP, consider private deployment options and keep an eye on contractual protections (who owns the outputs, how is data stored, can the vendor use your data to train other models).
The developer reality (and why agencies still matter)
No matter how glossy the marketing is, building reliable software remains engineering work. Integrations, secure APIs, testing, monitoring, CI/CD; all of these fundamentals still need to be handled properly. Bringing AI into a product adds complexity: new infrastructure, new testing regimes, and often new expertise. That’s where a pragmatic development partner can help; not to sell you a dream, but to design a viable path from idea to repeated value.
A simple playbook we use
- Problem discovery and value hypothesis.
- Data audit and feasibility check.
- Lightweight prototype with measurable KPIs.
- Pilot with a selected user group and monitoring.
- Production rollout plus ongoing model ops and user support.
Final word: be curious, stay thoughtful
AI will reshape how we build and use software; for the better in lots of cases. But it’s not a shortcut to product strategy, nor does it absolve the need for good engineering and design. If you obsess over the problem, own your data, and plan for operations, you’re far more likely to turn AI from hype into genuine advantage.
If you’d like to talk through where AI could actually move the needle for your business give us a shout. We’re Putti: curious, pragmatic and based in Auckland. We’ll help you separate the slop from the stuff that actually works.