Where LLMs Actually Pay Off in the Enterprise (and Where They Don't)

Generative AI is genuinely useful in specific, bounded situations — and genuinely disappointing when applied everywhere else. After a run of AI discovery engagements, a pattern in what works has become clear.

Where it pays off

Retrieval over your own content. Answering questions grounded in your internal documentation, policies or knowledge base is the single highest-success-rate use case we see. The model isn’t inventing an answer — it’s finding and summarising your own, already-correct source material.

First-draft generation with a human editor. Drafting support responses, marketing copy variants, or code scaffolding — where a person reviews and edits before anything ships — reliably saves time without introducing unchecked risk.

Structured extraction from unstructured input. Pulling fields out of invoices, contracts or free-text forms into a structured record is a narrow, well-defined task LLMs handle well, particularly compared to the regex-and-rules approaches it usually replaces.

Internal tooling and copilots. Lower stakes than customer-facing features — if the model gets something wrong, an employee catches it before it reaches a customer.

Where it doesn’t — yet

Fully autonomous customer-facing decisions. Letting a model make an unreviewed decision that affects a customer (a refund, an eligibility call, a legal interpretation) is where we see the most expensive failures. The cost of being wrong is asymmetric, and the model doesn’t know when it’s wrong.

Precise numerical or compliance-critical output with no verification step. Generative models are language models, not calculators or compliance engines. Without a verification layer, using them for exact figures or regulatory determinations is a liability, not a feature.

Anything where “mostly right” isn’t good enough. If the acceptable error rate for a task is close to zero, that’s usually a sign you need a deterministic system, an existing ML technique with real accuracy guarantees, or a human in the loop — not a generative model on its own.

The actual question to ask

Before building anything, we ask: what happens the first time this is wrong? If the answer is “an employee reviews it before it matters,” the use case is usually worth pursuing. If the answer is “the customer finds out,” it needs a lot more design work before it ships — or it isn’t ready yet. Get in touch if you want a structured assessment of where this applies to your business specifically.