Making AI useful
Everyone says AI. What actually works?
AI rarely comes down to one big decision. It is a series of smaller choices, and many are not worth pursuing. The real value is usually in a few practical places: repetitive work, information that is already written down, and tasks where a good answer now is more useful than a perfect one next week. Those are the ones worth finding; the rest can be left alone.
What it looks like
The problem is rarely the technology.
The pilot that never landed
The demo looked promising several months ago. Nothing has reached production since, and nobody quite wants to say it has stalled.
Licences nobody is using
A seat for everyone, a burst of enthusiasm, and usage that quietly falls away by the second month.
Answers you cannot check
It produces something plausible, but nobody can see where it came from — so nobody will put their name on it.
Everything or nothing
The only options on the table are replacing an entire platform or doing nothing at all. The useful answer is often somewhere in between.
How we work
Not every problem needs AI.
It starts from what the organisation needs to do better, and works back to the simplest tool that will do it. Sometimes that is AI; just as often it is a rule, an integration, or a form that no longer has to be filled in twice. The measure is what the business can do afterwards, not how much of it is AI.
-
Discovery & assessment
Working through where AI would genuinely help and, just as importantly, where it is not worth the effort. You get an honest assessment rather than a sales pitch — a clear path forward, and the ideas I would talk you out of.
-
Data readiness
AI is only as reliable as the material behind it. A review of what you already hold, the gaps found, and the right foundations in place — whether that means tidying spreadsheets, improving source material, or connecting systems that have never spoken.
-
Build & integrate
Built into the tools your team already opens each morning, rather than another system for them to manage. Off-the-shelf where that is the sensible answer, custom where it is not, sized to your scale and budget.
-
Train & improve
Training for your team, a measure of what changed, and improvement as people use it. The point is not a launch — it is something still being used, and still earning its keep, six months later.
What you end up with
Fewer experiments. More things that work.
- A short list of where AI can genuinely help, and a shorter list of where it is not the right answer
- Answers your team can trace back to a source and stand behind
- Something running in production, rather than another pilot
- A clear view of whether the next idea is worth pursuing before you commit to it
Client story
The forms are automated. The data stays put.
Talking Trouble works with sensitive information, so any automation had to meet the same privacy and compliance obligations as the processes it replaced. The team was spending real time on routine forms, and sending that material to a third-party AI provider was not an option. The form automation built for them uses a custom-trained model, deployed in an environment the organisation controls. It does the one task it was trained for, keeps every piece of data inside the organisation, and costs very little to run. The team has the time back, and how sensitive information is handled has not changed.
Talking Trouble Aotearoa New Zealand
Further reading
A seat for everyone, or a system that does the work
There are two ways to pay for AI. One gives you a predictable invoice; the other gives you a number you can defend. Most organisations have only tried the first.