The rush is reversing
For the past two years, the instruction in many boardrooms was simple. Get AI tools and agents into the business, and get it in fast. Now some of those businesses are doing the opposite.
A recent Guardian report describes employers dealing with a steady stream of errors and poor-quality output from AI tools. The clean-up is eating into staff time, adding privacy and security risk, and driving up cost. Some organisations that once encouraged open experimentation are now restricting it, rehiring experienced people and investing more in training.
New Zealand businesses don’t have to look far for an example. Commonwealth Bank of Australia reversed the redundancies of 40 customer service staff after its AI voice bot couldn’t cope with call volumes.
Hitting the brakes treats the symptom
Pulling back is an understandable reaction, but it’s a blunt one. Restrictions reduce the number of things that can go wrong, but they don’t change why things went wrong in the first place.
Often the AI did exactly what it was asked to do, and the trouble sat underneath it. The process it was dropped into was already slow, inconsistent or held together by workarounds, and the data feeding it was incomplete or unreliable. Automating that doesn’t remove the flaws, it just runs them at higher volume.
Faster steps, longer queues
The first problem is that an agent placed on a broken process doesn’t fix it. It inherits every flaw in the steps around it, and then repeats those flaws faster than any person could.
The second problem is less obvious. A process is a chain of steps, and it moves only as fast as its slowest link. Speed up one step with an agent and the work arrives sooner at the next one, where it waits. A report compiled in minutes still sits for a week if approval takes a week. The bottleneck hasn’t gone anywhere. It’s just moved further down the line, with a longer queue in front of it.
So before you decide where an agent should go, it’s important to step back and look at the whole process it’ll be part of.
First, fix the data and the process
This is where Mero starts. When we find a slow, error-prone or broken process with a client, we look at the data and the process before we look at the technology.
Before anyone adds an agent, we ask three questions.
- Where does the data come from, and can it be trusted?
- Which steps add value, and which exist only to work around a problem further upstream?
- Where do things get held up, and why?
The answers decide what comes next. Sometimes the answer is an agent. Sometimes it’s a cleaner data pipeline or a simpler workflow, with no AI involved at all.
For a data platform consultancy, this is familiar ground. An agent is only as reliable as the data it works from. Fixing that first is what makes anything built on top of it worth trusting.
Then build it with controls in place
Once the process is sound, how the solution is built and deployed matters just as much. Many of the problems in the Guardian report trace back to tools introduced quickly, with little oversight of what they could access or what they produced.
Mero builds and deploys these solutions on AgentAlloy, our production platform for custom applications, including AI and agentic solutions. It’s designed around the controls that rushed rollouts tend to skip.
| What’s gone wrong in rushed AI projects | How AgentAlloy handles it |
| Errors reaching the business unchecked | Every solution passes security, code, data quality and cost review before it reaches a live environment |
| Privacy and security exposure | Identity management and access controls are in place before anything touches production. Agent-based firewall protects against PII or sensitive data leakage |
| No way to trace what went wrong | Inputs, decisions and outputs are logged and available for audit at any point |
| Agents added where they aren’t needed | Mero recommends the right approach for each problem. Not every application needs to be agentic |
These controls are already built into the platform, so they don’t have to be rebuilt for every project. Teams with an existing build typically reach production in weeks rather than months. And when something does go wrong, someone can see what happened and fix it at the source.
Why companies are rehiring experienced people
What stands out in the Guardian report is that companies are bringing back experienced workers.
Those people hold the knowledge the AI was missing. They know which rules have exceptions, which data can’t be trusted, and where the process breaks on a bad day. Without them, errors are harder to spot and slower to fix.
AgentAlloy is built on the assumption that those people stay. The client brings the business logic, meaning the rules, workflows and domain knowledge specific to their organisation. Their team stays in control of what the solution does and how it behaves. The aim is to give experienced people more capacity, not to design them out.
Start with what sits underneath
There’s nothing wrong with moving quickly on AI. The trouble starts when you move quickly without sorting out what’s underneath first and the rules around it.
If you have a process that is slow, error-prone or held together by workarounds, start there. Fix the data, fix the process, then decide where AI belongs and build it with the right controls from day one.
See how AgentAlloy works, or get in touch to talk about the processes slowing your business down and how we can help fix them.