Entexia™Intelligent ecosystem
First touch · Evidence beside every sentence

A personalised email that rests on facts

Not a template with a name dropped in. The email is written around one measured finding about the recipient, every claim names its source, and a person always approves the sending.

No invented numbers Evidence panel beside the draft No automatic sending Consent and guards built in The recipient's language

Why cold emails do not work

A recipient needs three seconds to tell an email written for them from one written for a thousand addresses. The difference is not tone and not length, but whether the sender knows anything about them. A template with a company name dropped in knows nothing, so it falls back on phrases anyone could write about anyone. A language model with no guards makes the same mistake faster and more smoothly.

Template and mail merge
  • The text is the same for everyone, only the name and the company change.
  • The claims are general, so they cannot be verified.
  • The numbers are decorative, because they do not come from the recipient's data.
  • Sending runs automatically, with no control over the individual email.
Personalised email
  • The text is built around one finding about this recipient.
  • Every sentence names its source in the Evidence panel.
  • Amounts come from the service catalogue, not from the model.
  • A person reviews and approves every email before it leaves the house.

Time-saving calculator

Enter your own numbers. The calculation compares the time spent writing by hand with the time spent reviewing a prepared draft, and contains no assumption of ours about response rates.

Your numbers

How many first emails to new companies you send in one month.

Time to look at the website, search the registers and write the text.

Every draft is reviewed and approved by a person. Without approval the email is not sent.

Total employer cost of wages divided by hours worked. If you do not have the figure, leave the default.

Time saved per month
12.0 h
144.0 h per year
Writing all approaches by hand14.7 h
− Reviewing and approving drafts−2.7 h
= Difference12.0 h
Value of the time saved
420 €
5,040 € per year
Approaches in the same time
220
instead of 40

What this calculation is not

The calculation compares spent time only. It does not include the cost of running the analyses and generating the texts, because that depends on the chosen package and the scope of the analyses - prices are on the pricing page. It also does not predict recipient response: we do not measure reply rates and therefore do not promise them.

How one email comes about

Seven steps. Each of them can stop or correct the next one.

1

Context

A picture of the recipient is assembled from the measured analyses and from your own history with them. None of it is guesswork: every finding carries a label and a source.

2

Offer selection

The model proposes one item from your service catalogue and must state which findings justify it. The code reviews the proposal and may reject it.

3

Hook selection

One concrete finding is chosen as the hook the email is written around. When there is none, the email says so and does not invent another.

4

Writing

The text is produced in the recipient's language, at the length of a first approach, with one hook and one call to action.

5

Fact check

Every number, amount, date and claim about the recipient is compared with the structured findings. Whatever has no source falls out.

6

Style linter

Banned phrases, headings, bullet points, emoji, exclamation marks and dashes are rejected. The text is stripped of invisible characters and formatting residue.

7

Editor and Evidence

The draft opens in the editor, next to it the Evidence panel with a source for every sentence. Sending is an act of a person, not of the system.

The model proposes, the code verifies

Which service to offer

The model does not choose freely. It returns one item from your service catalogue and must state the labels of the findings that justify the choice, plus a short explanation. The code then reviews the proposal. A rejection is not an error and is not retried: the email simply stays general, and the reason for the rejection is visible in the interface.

The code rejects the proposal when
  • the proposal is not valid JSON
  • the quoted item does not exist in the catalogue
  • the justification is empty
  • the proposal cites no finding at all
  • the quoted label is not among the measured findings
  • the proposal rests only on identity data (name, address, tax number)

The last rejection is in practice the most frequent and the most important. That the recipient has an address and a tax number is no reason for an offer: everyone knows that about them. An offer stands only on a finding that actually says something about them.

The facts rule and the Evidence panel

The email may contain no number, amount, competitor name, date or claim about the recipient that does not follow from the structured findings. Price ranges are not a matter of the model's judgement: they come from the service catalogue, and where we have no range for an area, the hook goes into the email without an amount. When there is no concrete finding about the recipient at all, the system writes that down instead of inventing one.

We do not write public financial data about the recipient (size, headcount, revenue) into the email. For Slovenia we have no measured free source we could read them from reliably, so the system leaves them out rather than guessing.

Beside the draft stands the Evidence panel: for every sentence it shows which analysis and which finding it came from. This is not trust decoration, it is a working tool. Before you send the email, you can see whether a sentence rests on a reading of the website, on a register record, or on your own history with that contact.

Why the email does not read like AI text

Because the banned phrases are a list, not a recommendation. The list lives in the system and not in the instruction to the model, and text that hits a banned phrase is rejected and rewritten, at most twice. The lists are built separately for each language and are not translations of each other; they are confirmed for Slovene and English, and an email in a language without a confirmed list is stopped, because an empty list would let it through unchecked.

Banned and suspect in Slovene (extract)
v današnjem hitro spreminjajočem se svetuključnega pomenacelovita rešitevbrezhibnoizkoristiti potencialdvigniti na višjo ravenupam, da vas to sporočilo najde dobrone oklevajtez veseljem vam pomagamrevolucionareninovativen pristopsinergijaoptimizirati vaše poslovanjev svetu digitalizacije
Banned and suspect in English (extract)
delveI hope this email finds you wellseamlessleverageelevateunlockgame-changercutting-edgerobusttailored solution
60 - 130
words in a first email, 40 to 90 in follow-ups
1
hook and one call to action, nothing more
0
formatting: no headings, bullet points, bold type or emoji
0
dashes: the en dash and the em dash are caught by code point

The first sentence must open with a concrete, verifiable fact about the recipient; a rhetorical opening question is banned. Before saving, the text is stripped of invisible characters and formatting residue, and the mail headers are the same as for manual sending from the same address. The signature is not generated anew: the one from the user's settings is used.

What the hook is made of

Only publicly available sources and your own data in the CRM. Each analysis runs once per domain and is then reused within its validity period at no cost.

SourceWhat is read from it
Recipient's websiteThe text of the public home page: what the company says about itself, which services it lists, what is missing.
FURS taxpayer registerWhether the recipient is registered for VAT and since when.
AJPES business registerLegal form, registering authority and registered office from the public export of the register.
VIESValidity of the VAT number for cross-border business.
History in the CRMYour previous touches, quotes and support tickets for this contact.
TED public procurementNotices on the European portal. Available with the Partner monitoring module.
AI visibilityHow visible the recipient is in language models and search engines.
Preverba360, legal partLegal gaps on the website. Price ranges come from the catalogue, not from the model.
Human approval

No email is ever sent by the system

A draft stays a draft until a person sends it. Automatic sending cannot be switched on in any setting, because this is not a setting but a decision: on a first approach to an unknown company, someone has to take responsibility for the individual message, not for a rule.

One approval
one draft. The approval is recorded with the person and the time.
Editing is visible
the system measures whether the draft was changed before sending.
The recipient sees no AI
the record of how the draft came about stays internal.
Consent and guards

Who the email may be sent to

Legal responsibility for the content and for the lawful basis of sending rests with the company sending the email. We therefore do not leave this question open: before every send, a shared guard runs and returns all blocking reasons at once, not just the first one it finds.

Email consent on the contact

The values yes, no and no data, defaulting to no data. With yes, the date, the source and the person who entered the consent are mandatory. Every change goes into the audit log.

The value no closes sending

Even if your company has not set its own rule yet. Until then, the value no data is a warning, not a block.

Exclusion list and unsubscribes

Addresses are compared in normalised form, so an unsubscribe does not fail over capital letters or spaces. The unsubscribe link is on by default.

An existing customer is measured

Not from a label someone sets, but from issued invoices. Drafts, cancelled and unnumbered documents do not count.

Statement of responsibility

The generate button is unavailable until the company accepts the statement. The acceptance is recorded with the user, the time, the IP address, the language and the version of the text.

The processing of personal data is governed by a data processing agreement. Read the data processing agreement.

Languages

Three independent languages

The language of the interface you work in, the language of your sales profile and the language of the email are three separate settings. You can work in the English interface, keep the profile written in Slovene, and send the email in English because the recipient is foreign. The email is written only in languages your company has in its profile: a language you cannot read would be a promise you cannot check.

Each language has its own banned and suspect phrases. The English list is not a translated Slovene one and the Slovene is not a translated English one, because the two languages smell of AI through different signs. That is exactly why a list is never created by translation: only a language whose set we have confirmed with a native speaker may enter the profile. Today those are Slovene and English, while German, Croatian and Polish follow once their sets are confirmed. An email in a language without a confirmed set is stopped; an empty list would be the appearance of a check, not a check.

Frequently asked questions

What is a personalised email and how does it differ from a mail merge?+

A mail merge inserts a name and a company into a template, and the text stays the same for everyone. A personalised email is built the other way round: first the concrete findings about the recipient are measured (what their website says, their VAT registration status, what the public register shows), and only then is the text written around one of those findings. If there is no finding, the email stays general and says so.

Can the AI make things up?+

It must not, and the code checks that. The email may contain no number, amount, competitor name, date or claim about the recipient that does not come from the structured findings of the analyses. Price ranges are taken from the service catalogue, not from the model. Every sentence states its source in the Evidence panel, so before sending you can see what it rests on.

Is the email sent automatically?+

Never. Every email is reviewed and approved by a person, and automatic sending cannot be switched on in any setting. A draft stays a draft until someone sends it from the editor, and the system records who approved which draft.

Why does the email not read like AI text?+

Because the banned phrases are a list, not a recommendation. The text goes through a style linter that rejects phrases such as 'I hope this email finds you well', 'seamless', 'tailored solution' or 'feel free to', rejects headings, bullet points, bold type, emoji and exclamation marks, rejects the en dash and the em dash, and requires the first sentence to open with a concrete, verifiable fact. The lists are confirmed for Slovene and English so far. A first email runs from 60 to 130 words, with one hook and one call to action.

What data about the recipient does the system read?+

Publicly available ones: the text of their website, VAT registration status from the Slovenian tax register FURS, legal form and registered office from the public export of the AJPES business register, the validity of the VAT number in VIES, public procurement notices from TED, and your own relationship history with that contact from the CRM. Public financial data (size, headcount, revenue) is deliberately kept out of the email, because for Slovenia we have no measured free source for it.

How are email consent and GDPR handled?+

Every contact has an Email consent field with the values yes, no and no data, defaulting to no data. With yes, the date and source of the consent and the person who entered it are mandatory. The value no closes sending even if the company has no rule of its own. The sending guard also checks the exclusion list, unsubscribed addresses and internal rules, and returns all blocking reasons at once, not just the first. Legal responsibility for the content and for the lawful basis rests with the company sending the email.

What language is the email written in?+

In the recipient's language, chosen from the languages your company has in its sales profile. Only a language with a confirmed codebook of banned phrases can enter the profile, and today those are Slovene and English. The interface language, the profile language and the email language are three independent settings: you can work in the English interface and send the email in Slovene.

How many model calls does one email use?+

At most four: offer selection, writing, and up to two rewrites when the linter or the fact check rejects the first version. With a manually chosen offer there are three. Every call is written to the usage ledger, so the cost of an individual email is visible and is not an estimate.

See one draft on your own contact

The personalised email is part of the AI CRM agent in Entexia. In the demo we take a real contact of yours, prepare a draft and go through the Evidence panel together. Sending remains your decision.