In cold outreach you can automate the research, not the cold approach. Anyone who lets software blast a scraped list is not automating a sales process, they are automating a breach of the law: section 7 of the German Act Against Unfair Competition (Gesetz gegen den unlauteren Wettbewerb, UWG) treats advertising by electronic mail without the addressee's prior express consent as an unreasonable nuisance, towards consumers and towards companies alike. For advertising calls, at least presumed consent is enough towards other market participants, and that is a narrow exception, not an open door. An AI voice bot falls under the automatic calling machine and therefore always needs prior express consent. Everything that happens before contact, on the other hand, can be fully automated: finding candidates in public sources, checking them against fixed criteria, spotting a verifiable trigger and preparing a personal first contact. You can also automate the routes that may end in provable consent; once it exists, sending may run automatically too, it is simply no longer a cold approach. That groundwork is not rule free either: as soon as sole traders, named contacts or people are identifiable in the list, collecting and rating them is already processing of personal data under the GDPR. So the line does not run between human and machine, it runs between preparation and sending, and both sides of it have rules of their own.
I have wired that separation into my setup rather than promising it to myself. My AI employee Olaf researches and rates, he sends nothing and calls nobody, and that stands as a rule in his working instructions. Which tasks can sensibly be handed over is sorted out in Which Tasks You Can Hand Off to AI; the overview of the five routes to new clients is in Finding New Clients as a Freelancer. This post takes one of them and goes deep.
What section 7 UWG says word for word
The decisive passage is subsection 2. It states that an unreasonable nuisance is always to be assumed, and the original German wording reads:
"1. bei Werbung mit einem Telefonanruf gegenüber einem Verbraucher ohne dessen vorherige ausdrückliche Einwilligung oder gegenüber einem sonstigen Marktteilnehmer ohne dessen zumindest mutmaßliche Einwilligung,
- bei Werbung unter Verwendung einer automatischen Anrufmaschine, eines Faxgerätes oder elektronischer Post, ohne dass eine vorherige ausdrückliche Einwilligung des Adressaten vorliegt"
In plain terms: advertising by telephone call to a consumer without their prior express consent, or to any other market participant without at least their presumed consent, counts as an unreasonable nuisance, and so does advertising using an automatic calling machine, a fax machine or electronic mail without the addressee's prior express consent.
Three things in there matter for automation. First, for email the text draws no line between consumers and companies, and consent has to exist beforehand, not be collected afterwards. Second, presumed consent applies only to the telephone call in business dealings. It means a concrete interest in that particular call, recognisable from the circumstances, not the assumption that a business will be pleased about more revenue. Third, the automatic calling machine sits in the same item as email. An AI voice bot therefore does not get the business relief, however natural it sounds.
Subsection 2 item 3 hits exactly the tools mass sending usually relies on: an advertising message is also an unreasonable nuisance where it conceals or hides the identity of the sender on whose behalf it is transmitted, or where there is no valid address to which the recipient can send a request to stop such messages. Rotating sending domains and unfamiliar sender names is therefore a second forbidden track, quite apart from the consent question. I am not a lawyer and this is not legal advice; for an individual case, read the statute or ask one.
The existing-customer exception and its four conditions
Section 7 subsection 3 UWG names one case in which email advertising is permitted without prior express consent. The catch: all four conditions have to be met at the same time. Drop one, and subsection 2 applies again.
Two spots get read over regularly. The first condition speaks of a sale, not of a conversation: an address from an enquiry that never closed does not carry the exception. And the fourth requires the notice twice, at collection and on every single use.
For automation that is good news, because it maps cleanly onto data. If your system carries with every address where it came from, whether a sale sits behind it, whether an objection exists and when the right to object was last pointed out, software checks those four points before an email is composed. In my experience it forgets less than a spreadsheet kept by hand. Getting those four fields cleanly into your own system is real work; if you would rather not do it alone, bring the question into the community, where people have built the same list before.
What can genuinely be automated
The part of cold outreach that eats time is not the sending. It is the question of who to approach, and why that business in particular. That is diligent work along fixed rules, which is exactly what machines are good at.
- Find candidates. Pull together businesses of one kind in one area from publicly accessible directories, with address, website and public rating. That is pure reading: no message reaches the business, nobody is called or written to. Trace free it is not, since page requests and API calls can show up in server and provider logs.
- Qualify. Check every candidate against the same criteria instead of picking by gut feeling: encrypted connection, usability on a phone, age of the site builder in use, a reachable way to get in touch.
- Spot the trigger. A trigger is something you can see from the outside and put into words: a site without HTTPS, a site that is unusable on a phone, excellent ratings but no website. Without a trigger a first contact is an interruption.
- Sort out. The most valuable automation step is the one that throws candidates away. No reliable contact route from the legal notice or contact page, no lead. No website, no lead, because the comparison material is missing. Objection on record, no lead.
- Prepare the first contact. A short dossier per business: what stands out, what backs it up, what it means for that business. From that, a human writes a true opening sentence in two minutes.
- Document. When it was checked, from which source, with what result, and when the entry gets deleted. You need that for yourself and for the answer when someone asks where you got their data.
What is deliberately not on the list: sending, calling, following up. The step from item six to contact belongs to a human.
Research has rules too
Publicly findable does not mean free of data protection duties. The GDPR protects natural persons, not the bare company data of a limited company. In a lead list a personal reference appears fast, though: a sole trader, a named contact, an email address with a first name in it. From that point on, four questions belong in the process.
- A legal basis under Art. 6. In sales that is usually legitimate interest, which requires a balancing you write down rather than assume.
- Information under Art. 14. Data not collected from the person themselves triggers a duty to inform. The first message to them carries it at the latest.
- Data minimisation and a retention period. Only the fields the rating genuinely needs, plus a date on which the entry disappears again.
- Objection under Art. 21. Direct marketing can be objected to at any time without reasons. The objection has to block every later run.
A tool carries that part well, because it is fields and deadlines. What it cannot give you is a blanket assurance that a particular piece of research is lawful; that depends on the data you actually store.
Routes to lawful first contacts and provable consent
Instead of working around the consent requirement, automate the routes that come before it. They lead to three different outcomes, and only one of them is consent.
This is where consent you can prove comes into being:
- A newsletter with documented double opt-in. Consent is given at sign-up and made provable by the confirmation email. What you may automate afterwards is sending within the purpose consented to, until the consent is withdrawn. "Freely" it is not.
- Forms with a clear statement of purpose. That statement is the basis for later saying what the consent covered. Without it you have a tick box with no content.
This is where a statutory exception applies, not consent:
- Existing customers under subsection 3. The route described above, when all four conditions hold at the same time and are checked before every send.
This is where only a contact arises, and the consent question stays open:
- Content that answers a real question. Whoever arrives through search has taken the first step themselves. A visit is not permission for a marketing email, though; the form on that page is what creates it. The content itself can largely be handed to AI, as described in Automating Processes With AI: 8 Examples.
- Referral and network. Here a third party takes the first step, and the first contact is personal rather than cold. That still does not grant marketing consent; you ask for that expressly when the moment comes.
- The human call in business dealings, where a concrete interest in that particular call is recognisable from the circumstances. The groundwork for it, namely the evidence of that interest, is the automatable part. The call itself is not.
One observation from my own work, not a measurement: the legal narrowness forces you into the behaviour that is more careful anyway. A hundred identical emails name no trigger; ten documented triggers produce ten different opening sentences.
How Olaf maps that line
Olaf is my AI employee for this groundwork, and expressly not an outreach tool. I give him a town and an industry, he gives me back a rated list of local businesses with an outdated website: every entry with a reason, evidence and an A, B or C grade, plus a brief for the website build. His toolchain has three parts: local research via the Google Places API, a technical quick check per website, and screenshots of the site, on a phone as well where needed, because he only judges a design with the picture in front of him, never from a measurement alone.
The most important sentence in his instructions: he contacts nobody. No emails, no calls, no message that arrives at the business. On top of that comes the rule that separates him from a bought list: without a reliable contact address from the legal notice or contact page there is no lead, only a C grade. Addresses are never guessed.
How hard that filters shows in the numbers from my lead runs in July 2026: 198 candidates checked across eight assignments. In the seven local assignments with 20 candidates each, that typically yielded 3 to 5 real leads; the eighth was a differently built community list. In the first cross-check by my website colleague 4 out of 10 briefs were faulty, and since then every lead is checked a second time. The frame for rebuilding it is in AI Employees for Consultants and Agencies. Finding the criteria for your own industry is the part that costs thinking, and that goes faster in a group than alone.
Frequently asked questions
May I email companies without consent?
No. Section 7 subsection 2 item 2 UWG names advertising by electronic mail without the addressee's prior express consent as an unreasonable nuisance and draws no distinction between consumers and companies. The only relief named in the statute is the existing-customer exception in subsection 3 with its four conditions. This is not legal advice; in case of doubt the individual case needs checking.
Is an AI voice bot allowed for cold calling?
It stands on worse ground than a human on the phone. The automatic calling machine sits in section 7 subsection 2 item 2 UWG alongside fax and email and requires prior express consent. The relief of presumed consent applies only to a telephone call towards another market participant, so not to the bot.
What does presumed consent mean for a B2B call?
A concrete interest in that particular call, recognisable from the circumstances, not the general assumption that a business welcomes every offer. The test attaches to the specific case: does the offer visibly fit the recipient's ongoing business? Because that has to be set out if it is ever disputed, the trigger belongs on record before anyone dials.
Are bought lead lists a shortcut?
They do not solve the problem, they move it. A bought address brings no consent you could demonstrate to the recipient, and it brings no trigger to carry your opening sentence. On top of that sits the data protection question of where the data came from and whether the person concerned knows about it; what separates bought databases from interfaces and scrapers is in the tool comparison.
What is the difference between automated research and automated outreach?
Research reads publicly available information without a message arriving at the business; a request can still sit in a server log. Outreach sends something to a person and is therefore tied to consent. That line is where the machine stops in my setup. The GDPR applies on both sides of it as soon as the list makes people identifiable.
How to take this further
Take your acquisition route and draw the line through it once: which step only reads, which one sends? Everything before the line you may automate, and that is where most of the effort sits. For everything after it you need consent, a sound existing-customer case, or a documented interest for a call in business dealings. If you are unsure whether a case holds, that is the moment for a lawyer, not for an experiment.
How I handed the research before the line to an AI employee is on the page about Olaf, my lead scout. The template for it, with toolchain, rating matrix and exclusion rule, sits in my community Claude Practitioners.
