Should You Use ChatGPT or Build a Custom AI Agent?

A fair number of the enquiries we receive should not become projects. The honest answer is often "a subscription and an afternoon of setup will do this".

Here is how to tell which situation you are in.

Use the off-the-shelf tool when

A person is driving it. ChatGPT, Claude and similar are excellent when a human is in the loop — drafting, summarising, researching, rewriting. If your use case is "help me write things faster", you do not need a build.

The volume is low. Something done a few times a week rarely justifies engineering.

It does not need your systems. If the work does not require reading your calendar, writing to your CRM, or pulling from your price list, a general tool is fine.

The process is still changing. If you are not yet sure what the workflow should be, do it manually with an assistant until it stabilises. Automating an unsettled process is how you end up rebuilding twice.

Build custom when

It has to run without a person. The moment the work needs to happen at 2am, or on every enquiry automatically, a chat interface stops being the answer.

It needs to touch your systems. Booking into your calendar, creating a job, pulling from your inventory. This is the most common genuine reason.

It has to be consistent. General assistants are variable by design. If you need every quote to apply the same margin rules, you need something deterministic where it matters.

Data cannot leave. Health, legal or financial data often has constraints a general subscription cannot satisfy.

The volume is high. A few hundred instances a month is where per-seat subscriptions and manual handling both start to strain.

The middle option people forget

Between "use ChatGPT" and "commission a build" sits configuring existing tools properly. Most businesses have a CRM, an accounting package and a scheduling tool that already talk to each other more than they realise.

Before commissioning anything, the cheapest question is: what does our existing stack already do that we are not using?

We have talked several enquiries out of a build with that question. It costs us the project and saves the relationship.

What a custom build actually costs

Against a subscription at $30 to $100 per seat per month, the arithmetic only favours building at genuine scale — or where a build retires several subscriptions at once. One build we delivered replaced three tools and saved $130,000 a year in licence fees, which changes the calculation entirely.

A test that saves money

Before commissioning anything, do the process manually with an off-the-shelf assistant for two weeks. Log every instance: how long, where it broke, what needed judgement.

You will end up in one of three places:

1. Manual with the assistant is fine — stop, you are done
2. It works but the volume hurts — automate the specific bottleneck, not the whole thing
3. It keeps needing information the assistant cannot reach — that is your build, and now you can specify it precisely

That fortnight is the cheapest specification document you will ever produce.