Generic AI vs custom AI: when off-the-shelf stops being enough

Most companies should start with the off-the-shelf tool. A ChatGPT team plan, a Copilot license, the AI feature inside the SaaS you already pay for. For a lot of work that is the right answer, and building custom would be wasted money.

The real question is not whether to use AI. It is where the generic tool stops fitting, and whether that gap is worth building for. Here is how we draw the line.

Where generic quietly breaks

Off-the-shelf AI caps out the moment the work depends on things the model has never seen:

• Your data. Pricing rules, inventory, contracts, patient records, the internal wiki. A public model knows none of it, and pasting it in by hand does not scale.
• Your logic. The exceptions and approval steps, the ‘we never do X for customer type Y’. Generic tools flatten that into a plausible average, wrong in exactly the cases that matter.
• Your systems. The output has to land in your CRM, ERP, or database, not in a chat window someone copies from.
• Your standards. Regulated data, access control, audit trails, and a record of why the AI decided what it did.

What custom AI actually means

Custom does not mean training your own model. It means wiring the right model to your data, your logic, and your systems, behind evals and guardrails, so it does real work in production and holds up once real users arrive. Agents that run tasks, retrieval over your own knowledge, automations that move work through the tools you already use.

How to decide without overbuilding

Start with the subscription. Log where it fails: the answers it gets wrong, the steps people still do by hand, the output they copy-paste. That log is your spec. Build for those failures, not for the hype, and you spend only on the 20% that pays back.