Generic AI output: why it fails in client work
5 minutes read
A generic draft is fluent and well organised, and you could send it to any client in the sector unchanged, which is the problem. Where generic output comes from and how to get rid of it.
Generic AI output fails in client work because it could have been written for any client. A consultancy is paid for the part of the answer that applies to this client and no other, and a draft produced from a general model and a short prompt contains none of it. The draft may be accurate, but it answers a broader question than the one the client asked.
The model is rarely at fault. It can only write from what it was given, and in most cases that was a paragraph of instructions and nothing else.
How to recognise a generic draft
There is a simple test for a generic draft: swap the client's name for a competitor's and see whether anything else has to change. If the draft still reads correctly, it was never about the client.
There are other signs too. The recommendations are ones any firm in the sector would hear, and the evidence is general industry knowledge with no document behind it. Nothing in the draft would surprise the client, because none of it came from the client or from work only this firm has done.
Reviewers usually sense this before they can name it. The draft is hard to fault line by line and still not worth sending.
Where generic output comes from
Generic output comes from four gaps in what the model was given.
| What was missing | What the draft does instead |
|---|---|
| The client's own material | Describes a typical company in the sector |
| The firm's method and view | Uses the most common framework for the topic |
| Sources chosen for the brief | Draws on whatever general knowledge the model holds |
| The specific question | Answers the broad version of the topic |
With each gap, the draft moves closer to the average of everything the model has seen on the subject. Leave all four open and you get that average, competently written.
Why better prompts do not fix it
A longer prompt changes the instructions but adds none of the missing material. Telling a model to "be specific to the client" does not give it the client's documents. It only pushes the model to sound specific, which is worse, because invented detail is harder to spot than obvious generality.
Pasting material into the prompt helps with one task and no further. The next person starts from nothing, the pasted text is a fragment chosen in a hurry, and nobody reading the result can tell which document a claim came from. The firm ends up with a specific-sounding draft and no way to check it.
How to get rid of it
Generic output goes away when the missing inputs are present before the draft is written.
Start with the client's material: what they sent, what the firm produced for them before, and what was agreed in the brief. Add the firm's own method for this kind of work, so the structure of the answer is the firm's and not the internet's. Limit external sources to ones the firm would cite. Then ask the narrow question that is actually in the brief.
A draft built this way differs between clients because its inputs differ. It also has something a generic draft cannot have: a trail from each claim back to a document. How to choose approved sources for AI covers the source side of this in detail.
A specific draft still needs review
A draft built from the right material is a better starting point, and it still needs an expert to read it. Specific claims can be wrong in specific ways: a figure taken from an outdated file, or an inference that joins two documents which should not be joined. The difference is that you can check these, because each one links back to a document.
The time saving shows up in review. A generic draft has to be rewritten before it can go out, while a sourced one only has to be checked. The guide to reliable AI client work sets out the whole approach.
How this looks in Hebno
Hebno writes from the firm's library and the client files a person can already open, alongside public sources the firm has chosen. Each source is labelled internal, client, or public, and each finding is marked supported, inferred, or uncertain, so a reviewer can see what each claim is based on. Citations stay attached while the draft is edited. The work feature page shows a draft with its sources beside it.
