Custom AI agent or one off the shelf: how do you choose?
Most teams that ask for a custom agent do not need one, and the ones that do usually know exactly why. Here is the line between them.
Domande e risposte
When is a catalogue agent enough?
When the task is one most businesses do the same way: screening CVs against stated criteria, triaging a support queue, drafting a research brief from supplied sources. The work is standard even when your business is not, and a ready agent starts producing the same afternoon.
When do I genuinely need a custom one?
When the judgement in the middle is yours specifically — your underwriting rules, your product's failure modes, the six exceptions everyone in your team knows and nobody has written down. That knowledge is the thing being built. If you cannot describe it, the agent cannot use it.
What do I have to supply for a custom brief?
What the agent takes in, what it must hand back, and the rules that decide between them — including the exceptions. Examples are worth more than descriptions: five real inputs with the output you would have wanted is a better brief than two pages of prose.
Can I start with a catalogue agent and move later?
That is usually the right order. Run the standard one for a fortnight and the gaps announce themselves in specifics — this field, that exception, this format. A brief written after two weeks of real output is a far better brief than one written from imagination.
How long does a custom agent take to build?
Days rather than months, because the definition is the deliverable rather than a codebase. The scoping is the slow part, and it is slow in proportion to how clearly you can state the rules.
What if the custom agent turns out to be wrong?
Agents are versioned, so a change is a new version rather than a rewrite, and you can see what altered between them. Expect the first version to be about eighty per cent right — that is normal, and the last twenty comes from real inputs rather than from more discussion up front.