What an AI assistant in a CRM does, and what it does not

Example scenario: Picture an architectural lighting company in Şişli. In the morning the inbox holds an architect asking for a project price, a contractor asking about a delivery date, a supplier's invoice, and a customer complaining about a faulty fitting. The assistant reads each one, matches the sender to the customer card, sorts the subject, and puts it on the right person's list. The sales adviser starts the day by answering the price question, not by combing through the inbox.

The assistant's value shows as the small jobs add up. Filing one email on a card is a small job; but when it has to be done every day, for every team member, on every channel, the team either skips it or does half of it. The assistant finishes that half-done work, and the card starts carrying the whole conversation.

• Reads and sorts: ties incoming correspondence to the right customer and the right person.
• Summarises: boils a long thread down to a few sentences.
• Drafts: writes the first version of a reply, a reminder, or a quote text.
• Follows up: notices an unanswered quote, an approaching end date, a missing report.
• Reports: presents what it did, what is waiting, and what it is asking, in one list.

The decision boundary is the clearest one. Giving a discount, renewing a contract, approving an employee's leave, or making a payment is the business's responsibility. The assistant prepares everything up to those steps; a person takes the step itself.

The second boundary is knowledge. If a salesperson promised a customer a delivery date over the phone and it was never written down, the assistant cannot know about it. Because the assistant reads the record, it is only as good as the record. If the same customer has been entered under several spellings, the assistant can notice and ask, but it should not decide on its own which one is right.

The third boundary is context. The assistant may classify a short, blunt email as a complaint, when that customer always writes that way and is in fact happy. The salesperson knows this. The assistant's classification is a suggestion, and the person who knows the customer should be able to correct it.

AI models sometimes state something that is not true in a fluent sentence. In a chat app that is annoying; in an email to a customer it can become a contractual problem. Two habits keep the risk in check. The first is that every answer the assistant gives rests on a record: there has to be an answer to the question of which card or which message the information came from.

The second is that anything going to a customer or supplier passes before a person's eyes. The assistant writes the draft, a person reads and sends it. If the draft is wrong, the damage stays on the screen. Keeping every step the assistant suggests on record, and reversible, is part of the same safeguard.

If an assistant reads email, it also reads customer names, phone numbers, and sometimes invoice and contract details. That is personal data. The questions to ask are simple: where is the data processed, whose permissions govern the assistant's access, and can everything the assistant did be traced afterwards. For an accounting office in Istanbul this question is not theoretical; if its clients' data is processed on another company's servers, it has to know and be able to tell them. The answer should be in writing before you choose an assistant. We describe how data is protected in Corex360 on the security page.

• Where it stops: at which step it asks for approval and at which it carries on alone.
• What it rests on: can it show which record its answer came from.
• What it reads: email, messages, call history, and which of them are connected.
• Where it runs: inside the CRM, or in a separate tool outside it.
• Can it be undone: is everything it suggested and did kept on record.

Ask whether the assistant runs inside the CRM: an assistant working inside reads the record itself, and the note it writes stays on the same card.

Asking these questions with your own data gives the clearest answer. Have the assistant read a few emails from your own inbox and see which card and which person each one lands with. An email that lands in the wrong place is not a bad sign either; it is a good chance to see how the assistant is corrected.

X AI is the assistant that works inside Corex360. Once corporate email is connected, it reads incoming mail, matches the sender to the customer card, and sorts the subject: quote requests go to sales, invoices to accounting, complaints to the person responsible. When an unanswered quote, a contract nearing its end, or a missing field report breaks a rule, it opens a task and assigns an owner.

Every morning it brings one report to your phone showing what it did, what is waiting, and what it is asking you, and it gives a summary when the day closes. At sending, discounts, payments, leave, and permissions it stops and waits for your approval; every suggestion is logged and reversible. You can see how X AI takes over the day's follow-up in our page on handing the process to X AI.