I spent three months building an ERP system for my wife’s business. It was almost ready to launch. Then I discovered agentic AI — and ended up rethinking much of what I’d built.
There’s something particularly frustrating about spending three months developing a system, getting it almost ready to go live, and then discovering a completely different way of doing things.
Especially when you’re the technology consultant who designed it.
But that’s exactly what happened to me.
And it made me think about how many New Zealand businesses might be making expensive technology decisions right now, without realising how quickly their options are changing.
It started with a fairly ambitious project
My wife runs an Auckland-based event styling business.
Like many growing businesses, there are a lot of moving parts behind the scenes.
Customer enquiries, quotes, bookings, inventory, staff scheduling, vans, storage, deliveries, logistics, emails, invoicing — and somehow, all of it needs to work together.
So I decided to build an ERP-style system to manage the entire customer experience, from the first enquiry through to the completed event.
And I didn’t exactly keep it simple.
We had a PostgreSQL database, a custom web dashboard, Microsoft 365 authentication, CRM functionality, inventory management and a fairly extensive set of n8n automations connecting everything.
We even trained a local AI model to help manage customer emails.
The idea was that when a customer emailed us, the system would gradually gather all the information needed to manage their booking.
We had roughly 20 questions that needed answering, but instead of bombarding customers with all 20, the AI would ask two or three at a time, naturally throughout the email conversation.
Behind the scenes, the automations would update the CRM, manage inventory, coordinate logistics and keep everything connected.
It was a pretty impressive setup.
At least, I thought so.
Three months later, we were almost ready
Most of the system had been developed.
The database was there. The dashboards looked great. The workflows were built.
I was about to enable the email connector and start letting the system handle real customer enquiries.
There was just one small problem.
The more complex the system became, the more effort it took to maintain.
We had automations triggering other automations, database updates happening in the background, and scripts handling different parts of the process.
When something went wrong, tracking down the cause wasn’t always straightforward.
Sometimes fixing one workflow meant rethinking another.
It worked, but there was a lot of complexity behind what should have been relatively simple business processes.
And then I came across OpenClaw.
Enter agentic AI
I decided to give OpenClaw a proper test.
I installed it, connected it to some free models through OpenRouter and started experimenting.
Initially, I was just curious.
But the more I tested it, the more I realised that agents could approach the same problems very differently.
Instead of building complex chains of automations to tell the system exactly what to do at every step, I could give individual agents specific responsibilities, controlled access to information and the tools they needed to complete their tasks.
For example, one agent could read information from a particular database table, another could process that information, and a separate agent with the appropriate permissions could update the relevant records.
Skills, schedules and defined responsibilities made the whole setup easier to organise.
The biggest difference wasn’t necessarily speed.
It was how much easier the system became to manage, understand and adapt.
Instead of constantly maintaining complicated automation sequences, I could focus on the business outcome and give agents clearly defined tasks.
That was the moment I started questioning the architecture I’d spent three months building.
So, naturally, I started rebuilding it
Eventually, we moved from experimenting with OpenClaw to using Hermes, which I felt was a better fit for the security and operational controls I wanted.
The important change wasn’t simply swapping one AI platform for another.
It was redesigning the system around smaller, clearly defined responsibilities.
We kept the database and much of the underlying functionality, but removed a substantial amount of the automation work.
The front end also needed redesigning to suit the new structure.
We separated responsibilities, limited what individual agents could access and created clearer boundaries around reading information and making changes.
That distinction matters.
An AI agent capable of taking actions isn’t automatically safe or reliable. You still need proper access controls, validation, monitoring and human oversight, particularly when customer information and bookings are involved.
But with those controls designed into the system, I found the agent-based architecture considerably easier to work with.
We weren’t necessarily making every individual task faster.
We were making the overall system less complicated to manage and easier to change.
And that’s a pretty significant improvement.
Now imagine this was a $50,000 business project
Fortunately, this was my own development project.
I lost a fair bit of time rebuilding things, but I also gained something valuable from the experience.
Now imagine a business owner who has just received a $50,000 proposal to replace their CRM, upgrade their infrastructure or implement a new business management platform.
Their existing technology provider has recommended a solution.
The quote looks professional. The project has a timeline. The vendor seems confident.
And why wouldn’t the business owner trust them?
Most New Zealand businesses have a good relationship with their technology providers. They rely on those providers to recommend the right solutions because technology isn’t their area of expertise.
There’s nothing inherently wrong with that.
But here’s the question:
Is the recommended solution still the best option for the business, or simply the solution the provider knows how to deliver?
Those aren’t always the same thing.
And with AI capabilities evolving so quickly, that question is becoming increasingly important.
Should businesses put technology projects on hold?
Not necessarily.
I certainly wouldn’t suggest every business should stop investing in technology and wait for AI to settle down.
We could be waiting a very long time.
And sometimes the traditional solution is still absolutely the right one.
A well-designed automation can be more predictable, cheaper and easier to audit than an AI agent. Not every process needs AI, and not every new AI capability belongs in a production environment.
But I do think businesses should be asking better questions before committing to major technology investments.
Before buying new technology, it is worth asking what problem you are actually solving.
For example:
- Are we solving the actual business problem, or just replacing the existing technology?
- Have we considered what AI or automation could realistically do for this process?
- Are we being locked into a solution that will be difficult or expensive to change?
- Can the proposed system integrate with new capabilities as they become available?
- Have we explored alternatives, or are we relying entirely on one supplier’s recommendation?
- What happens if our requirements change significantly in the next 12 months?
The goal isn’t to predict what AI will look like in two years.
It’s to make sure today’s investment doesn’t unnecessarily limit tomorrow’s options.
Sometimes a second opinion is worth more than the technology itself
If you’re about to approve a significant technology project, my advice is simple.
Get another perspective.
Talk to someone independent. Speak with business leaders who’ve implemented similar systems. Ask questions about the architecture, costs, risks and alternatives.
You might discover that the original recommendation is excellent.
Great. Now you can proceed with more confidence.
But you might also discover a simpler approach, an overlooked risk or an opportunity to get more value out of technology you already own.
And that conversation could save you a substantial amount of money, time and frustration.
In my case, I was fortunate enough to be both the person recommending the technology and the person questioning it.
I still managed to spend three months building something before deciding to rethink it.
At least I didn’t have to explain the change of scope to a client.
Well, apart from my wife.
And somehow, I suspect she’s the tougher client.
The takeaway
AI isn’t just introducing new tools.
It’s changing the options businesses have when deciding how to solve problems.
That doesn’t mean everything needs replacing, or that the newest technology is always better.
It means the decisions we made even six months ago might be worth revisiting before we commit more money to them.
Before you approve your next technology investment, make sure you’re solving tomorrow’s problem — not just buying a newer version of yesterday’s solution.
Because sometimes the most valuable technology advice isn’t what to buy.
It’s knowing when to question what you’re about to buy.