✦AI in everyday use

In clearly defined places — and always as a suggestion, never as a decision:

• Text suggestions for emails, templates, descriptions.
• Fit assessment between a profile and a vacancy, with reasons.
• Column detection when importing files, if the headings give nothing away.
• Evaluating search results in the opportunity search.
• Summaries of long texts.

What runs without AI: contacts, campaigns, sending, inbox, template management, reports, contracts, deliverability.

This matters for putting it in perspective: AI is an ingredient, not a foundation. If you switch it off, Aurora Hive Sales keeps working — just with more manual work in some places.

With good input — the quality of the result depends almost entirely on it.

What helps:

• Be specific. “First approach to trades businesses in Saxony about recruitment services” achieves more than “sales email”.
• Give examples. If you have a text that worked well, include it as a model.
• Say what should not be in it. No superlatives, no exclamation marks, no “competent partner”.
• Set the length. “120 words at most.”

What does not help: a very long input with everything that comes to mind. An AI that reads thirty pages answers less precisely than one that gets one page, and it costs more.

And expect two or three attempts. The first suggestion is rarely the one you use.

In four places, and all four are predictable:

1. Facts. It invents years, certificates, references and legal citations — convincingly and without it being noticed.
2. Judging people. Whether someone fits into the team, whether a customer is ready right now, whether a CV tells a story.
3. Your business. It does not know your prices, your capacity or your particular features — unless you tell it.
4. Topicality. It does not reliably know what has changed recently.

What follows from this: check every factual statement. Figures, names, deadlines, sections of law — anything someone could look up.

The rule that has proven itself: AI may word things, not claim things.

Four things, in this order:

1. Are the facts right? Every figure, every name, every date, every legal statement.
2. Does it contain anything I cannot honour? AI words things generously — a promise in an email can be binding.
3. Does it sound like me? If not, rewrite it. A text that sounds like nobody has the effect of nobody.
4. Is there anything generic that can go? Usually a third.

What you will notice: the usable part of an AI suggestion is rarely the whole of it. Take the structure and one or two sentences, and replace the rest with specifics from your business.

And with anything that goes to customers: read it aloud once. What you would not say yourself, you should not write.

As a pre-sort, and the reasoning is more important than the number.

What works well:

• Picking the ten out of fifty profiles that you look at first.
• Reading the reasoning — it often points to something you would have missed.
• Checking contradictions: if the score is high and your gut feeling says no, look more closely. Usually your gut feeling is right.

What you should not do:

• Sort by the number and ignore the rest. The best candidate is often the one with the unusual CV.
• Reject applicants on the basis of the score. That is the most delicate point legally: automatically screening out applications falls under the high-risk rules.

A human sees every application. That is the rule that should not be tampered with.

With three steps, and the first is the most important:

1. Set a budget. An upper limit per month, for every chargeable service. Without a limit, usage keeps running, even if someone accidentally starts twenty runs.
2. Shorten inputs. A twelve-page PDF costs more than a paragraph — and delivers a less precise result.
3. Look at usage, once a month. An outlier always has a cause.

What you should not do: avoid using it at all to save money. Anyone who switches off the AI to save ten euros and then works two hours by hand has miscalculated.

Think in ratios: what did the usage achieve? Ten euros for a job is cheap.

Try again — it usually works the second time. Language models do not always respond the same way.

If it keeps happening, the cause is almost always the input:

• Very long texts. Cut them down to the essentials.
• Garbled characters from a PDF that could not be read properly.
• Unusual special characters.

Good to know: a failed attempt costs too. The request was made. With repeated failures, it is worth looking at the usage.

And if it happens reproducibly in a particular place: report it — stating which function and which input. That can be traced; “the AI isn’t working” cannot.

Only what is passed on for the task in question — not your records.

Specifically:

• For a text suggestion: your input and the context in which you request it.
• For a fit assessment: the profile and the job description.
• For column detection: the headings and a few sample values.

What is not passed on: your entire contact records, your history, your notes.

What you should still bear in mind: whatever you write into an input yourself goes with it. Anyone who pastes in a case with names and details transfers it.

One more rule that should apply in the team: no customer data in a private AI account. That is the most common unchecked data leak in small businesses.

Yes. The core of Aurora Hive Sales works fully without it.

Without AI, the following keep working: contacts, companies, opportunities, campaigns, templates, sending, inbox, follow-ups, contracts, deliverability, reports.

The following are lost: text suggestions, fit assessment, AI-supported column detection, evaluation of search results.

If you are considering this, check the reason:

• Data protection concerns? Usually solvable through a contract and data minimisation.
• Costs? Then budgets help more than switching off.
• A decision of principle? Legitimate — then do it consistently and document it.

And in any case: a written rule in the team. Otherwise someone uses it anyway, with customer data, in a place nobody has checked.

Five sentences on half a page:

1. Where AI may help and where it may not.
2. What may be put into it — and explicitly what may not: customer names, health information, confidential matters from client mandates.
3. That every suggestion is checked before it goes out.
4. That nobody is criticised for discarding a suggestion.
5. No private AI accounts for work content.

Point 5 is the most important and the one most often missing. If you have no rule, you do not get abstinence — you get uncontrolled use.

And if there is a works council: introducing AI tools is regularly subject to co-determination. Talk early, not after the introduction.

Increasingly, yes — and that is a practical problem, not a legal one.

How AI texts can be recognised:

• Perfectly built sentences without rough edges.
• Lists of three, everywhere.
• Phrases such as “in today’s fast-paced business world”.
• Lots of affirmation, little substance.
• No point of view.

What follows from this: an unchanged AI text in a first approach comes across like a mass email — exactly the opposite of what you want to achieve.

What helps: put in specifics. Figures, places, names, an example from your business, a sentence only you would write. That is the part no AI can supply, and the part that works.

Where you improve something, not where you replace something:

Worth it:

• When you are sitting in front of an empty field and cannot get started.
• When you want to shorten a text of your own that is too long.
• When you need ten out of fifty profiles to look at.
• When a file has fifty columns and you do not want to map them yourself.

Not worth it:

• For short texts. By the time you have worded the input, you could have written the email.
• For anything where your personal touch matters — an apology, a condolence, a personal offer.
• For decisions.

The rule of thumb: AI is good at getting started and at shortening. In the middle — where the substance matters — you are better.