A specific word

Generative AI produces new content — text, mostly, in a business system.

That distinguishes it from the other two kinds you meet.

Extraction reads a document and pulls out what is there. The invoice total either is 4,500 or it is not.

Prediction estimates an outcome from history. The forecast is 400 units, and in three months you find out how close that was.

Generation writes something that did not exist. There is no right answer to check against, only a judgement about whether it is good and true.

That difference is the whole subject of this article. Extraction and prediction can be verified. Generation has to be read.

Where it is used in an ERP

Six places, roughly in order of how safe they are.

Internal summaries. A long email thread, a customer’s history, a project’s status. Nobody outside sees it.

Internal notes. Turning a rough call note into something a colleague can act on.

Product descriptions. Written from attributes, at scale.

Customer email drafts. Replies and follow-ups, edited before sending.

Knowledge base articles. First drafts of support content.

Report commentary. A written explanation alongside the numbers.

The problem with fluency

Here is what makes generative AI different from every other kind of software you have used.

Bad software produces obviously bad output. A broken report shows an error, or numbers that are visibly wrong. You notice.

Generative AI produces well-written wrong output. A product description with a specification it invented reads exactly as well as one that is correct. There is no formatting cue, no error message, nothing to catch your eye.

Confidence is not correlated with accuracy. This is the single most important thing to understand about it.

The practical consequence: everything generated needs reading by someone who knows whether it is true. Not proofreading for grammar — checking for facts.

Where errors actually appear

Being specific is more useful than a general warning.

Invented specifications. Dimensions, materials, compatibility, certifications. It will produce plausible ones.

Wrong prices or terms. It has no reliable knowledge of your commercial terms unless given them.

Overstated claims. “Industry-leading”, “guaranteed”, “fastest”. These create expectations and occasionally legal exposure.

Tone mismatch. Fine in isolation, wrong for your company or that customer.

Confident answers to questions it cannot know. Lead times, availability, whether you can do something.

Where it genuinely earns its place

Three situations where it is clearly worth it.

Volume that is otherwise impossible. Four thousand product descriptions is weeks of writing. Generated and reviewed, it is days.

The blank page. For many people the hardest part is starting. A rough first draft to react to is faster than writing from nothing.

Repetitive variation. The same message adapted for twelve customer segments.

The pattern: it is most valuable where the volume is high, the content is routine, and a knowledgeable person is reviewing.

Where it is a poor fit

Content requiring specific expertise you have not given it. Technical documentation, regulatory text, anything where being subtly wrong matters.

Anything genuinely original. It produces competent, conventional output. If your differentiation is your voice, generated text will flatten it.

Content that must be legally precise. Terms, contracts, compliance statements.

Small volumes. If you need three descriptions, write them. The review takes as long as the writing.

Making it work

Five practices.

Give it examples. Showing it three descriptions you like produces far better output than describing what you want. This is the single highest-return technique.

Give it the facts. Feed it real attributes from your product records rather than letting it fill gaps. Most invented specifications come from unfilled gaps.

Review for truth, not style. The grammar will be fine. The question is whether the claims are correct.

Have the right person review. Somebody who knows the products, not just somebody who can read.

Keep a short list of things it must never claim. Certifications, guarantees, compatibility. Make it an explicit rule.

The quiet risk

Worth naming, because it develops slowly.

Review quality decays. In the first week everything is read carefully. By the third month, output has been consistently good and reading becomes skimming.

Then one description with an invented certification gets published across four thousand products.

Two safeguards:

Spot-check a sample deliberately, even after you trust it. A fixed percentage, reviewed properly.

Never remove review from anything published at scale. The volume that makes generation valuable is the same volume that makes one error expensive.

What this means for Odoo

Odoo includes generative AI in the screens where the content is needed — product forms, email composition, knowledge articles, record summaries.

The practical advantage is context. Because it is one database, a draft can use the actual product attributes, the actual customer history, the actual order details. Less is invented, because less is missing.

That reduces the error rate. It does not remove the need to read.

The short version

Generative AI is very good at volume and very good at first drafts.

Its distinguishing feature is that wrong output looks exactly as good as right output. There is no error message and no visual cue.

Give it examples and real facts. Review for truth rather than style. Never remove review from anything published at scale.

Used that way it saves weeks of routine writing. Used without review, it publishes confident errors faster than any tool you have owned.