✦AI in Aurora Hive Sales

Wherever something new is to be created from text — and nowhere else:

• Cover letters and text suggestions,
• Assessing profiles against a requirements profile,
• Summaries in research,
• Enriching contact details in the opportunity search,
• Translation suggestions (in Aurora Hive Sites).

Column detection during import and industry detection in research explicitly work without AI. Both would be simpler with AI and worse: more expensive, slower and with data leaving the company.

No. Every AI suggestion remains a suggestion and only takes effect once you approve it. There is no process at the end of which an email goes out without a human having seen it.

This is not about caution for its own sake, but about the nature of these models: they produce what sounds plausible — even when it is wrong. A suggestion that nobody reads is a random variable in your customer contact.

When assessing applicants, there is also a legal side: a selection made by the machine alone would be an automated decision within the meaning of the GDPR, and such use also falls under the high-risk classification of the European AI Act.

According to actual usage, not as a flat rate per click. What a run costs depends on the amount of text: a short cover letter is a fraction of a cent, a complete CV with assessment considerably more.

The most expensive single item is not the text output but the web search in research: one call there costs many times an ordinary AI request.

The overview shows usage for the current month as it accrues; the counter starts at zero every month.

Three safeguards that work together:

1. The preview before the click states the upper limit for chargeable runs.
2. A monthly budget per instance. Once it is reached, chargeable runs are refused instead of carried out — all other work carries on unaffected.
3. The licence is always the outer limit.

The right order is: budget first, then the first big run. A budget set after the expensive run only helps next time.

Useful as a rough draft, unreliable as an end result. The rule of thumb from practice: an AI text saves you the blank page, not the reading.

What to look out for in particular:

• Invented details. Figures, names, references, dates — if they are in the suggestion and you did not supply them yourself, they are guesses.
• An overly friendly tone. Suggestions tend to gush; in German business correspondence, that quickly looks unprofessional.
• Repetition across several recipients. If you have a hundred cover letters generated, you get the same sentence structure a hundred times.

Anything that uses AI is tied to modules, and modules depend on the licence. If you do not want to use AI in your company, the clean way is not to activate the modules concerned — then they do not even appear in the menu.

A budget of zero has a similar effect in practice: chargeable runs are then refused. The difference is that the buttons remain visible and lead nowhere — which confuses more than it helps.

That is not decided by each individual spot in the program but by a central configuration: calls go through one shared access point, and there it is set for each purpose which model is responsible.

The advantage for you: a change — because a model is retired or a cheaper one is enough — happens in one place and takes effect everywhere. The advantage in operation: there is no forgotten corner still using an old model.

By putting more in. The quality depends almost entirely on what you supply:

• For profile optimisation: the complete job description instead of the title. The more detail, the better the optimisation.
• For candidate matching: a real requirements profile, not three keywords.
• For texts: say for whom and against what — “short first approach to a trades business, no sales tone”.

What does not help: asking for it to be worded “professionally”. That produces empty phrases.

Because a language model actually generates the text rather than looking it up. For a complete CV, Aurora Hive Sales says so explicitly: this takes up to 90 seconds, please wait.

Do not click again during this time. In the worst case, a second click starts a second run, and that costs again.

Three common causes:

1. The budget is used up. Chargeable runs are then refused — everything else keeps working.
2. The licence is paused. During the grace period, sending and AI are suspended.
3. The input is too short. Some functions require a minimum — for candidate matching and profile optimisation, at least 20 characters.

On the overview, you see points 1 and 2 in one line.

Check two questions before you do:

• Does it need to go there? Column detection during import does not need AI, which is why there is none there. For rewriting a text, there is no other way.
• Is the processing covered? The AI service belongs in your legal texts as a sub-processor, and the person concerned should know about it.

What you should avoid in any case: copying access credentials, bank details or health information into a free-text field that goes to an AI service.

You have to read it in any case. Changing it is almost always worthwhile in three places:

• The first sentence. AI texts like to begin with an introduction that says nothing. Delete it.
• Superlatives. “Leading”, “innovative”, “tailor-made” — words everyone writes and nobody believes.
• The ending. Instead of “I look forward to hearing from you”, a specific, small question.

After that, the text sounds like you — and that is the difference that gets replies.

Often, yes — by the same features listed above: interchangeable introductions, lists of three, superlatives, an even sentence rhythm across a hundred emails.

Legally, you do not have to label a business text generated with AI; the labelling obligations concern interaction with an AI, not every text it has helped with.

In practice, it is still a question of impact: anyone who visibly writes automatically gets automated attention.