05/06/2026
AI agents need more than good prompts
In my own work building AI workflows and support systems, one lesson keeps coming back:
An AI agent is not just a smarter chatbot.
Once an agent can browse websites, read files, call tools, update systems, send messages, or act on someone’s behalf, the risk changes.
The issue is no longer only:
Can the AI give a useful answer?
It becomes:
What can this agent access?
What can it change?
Which instructions is it allowed to trust?
What happens if a webpage, file, email, or document contains hostile instructions?
What must still come back to a human before action is taken?
This is where prompt injection becomes a real concern.
A prompt injection attack can hide instructions inside content the agent is reading.
That hidden instruction may try to override the agent’s real job, bypass safeguards, suppress logging, expose information, or take actions the owner never authorised.
This is why I am cautious about rushing organisations into agents before the basics are clear.
Good prompts are not enough.
You need clear architecture.
Clear permissions.
Trusted instruction sources.
Human approval for higher-risk actions.
Logging that cannot be quietly bypassed.
Safe handling of sensitive information.
And a clear decision about what the agent must never be allowed to do.
For Māori, Pacific, community, and values-led organisations, this matters even more because the risks are not only technical.
They can affect trust, relationships, privacy, cultural integrity, service delivery, and accountability.
AI agents can be useful.
But they should not be given broad access, vague authority, or unchecked ability to act.
Start with the mahi.
Then decide what the agent is allowed to do, what it must never do, what needs human judgement, and what safeguards need to be in place before it touches real work.
JamesPratt.com
Start with the mahi, not the tool
