A deal desk in your assistant is the repeatable work behind a quote, done by the assistant your team already uses, to rules your company wrote down: which sources to consult, what the output has to contain, and what must come back to a person before it goes out.
The distinction matters because it is more than a chatbot. A chatbot answers from whatever it can find, and two people asking the same question get two answers. A deal desk works to a procedure, and the procedure belongs to the company rather than to whoever typed the prompt.
Nerd holds what your company published: the steps, the source each fact has to come from, the document and wording requirements, and which approval a request needs and where it has to be raised. Your assistant asks Nerd how this company does the job, then does the reading and the writing itself, under each person's existing permissions. Nerd does not connect to customer source systems, hold customer source-system credentials, or price quotes. It serves published guidance without running a model of its own.
Three. A model never authors a price; the playbook names the system the number comes from. Nothing commits; a person approves, in your own system, and Nerd holds no record that they did. And one person's way of working never becomes company policy: a change proposed from a real exception is published by a leader, or it is not published at all.
Teams already work in an assistant like Claude or ChatGPT every day. What they do not have is any way to make it answer the same way for everyone, which is a question about company knowledge rather than about the model.
Start with one repeatable workflow. Record how it is done today, review the playbook that comes out of it, publish it, and use it on the next real request. Expand once the first one is right.