AI on top of your CRM: how to draft replies and quotes with human approval

BOTrento Team3 October 20268 min read

What stays in the CRM and what AI can prepare

The most common mistake is expecting AI to "take over the CRM". It works the other way round: the CRM is the source of truth, and AI is a worker that reads from it and writes back to it under clearly defined rules.

This split prevents a second, "shadow" record from appearing. If AI gets something wrong, the mistake stays in the draft and reaches neither the CRM nor the customer without a check.

Where the delay between inbox and CRM comes from

In companies that handle event, equipment-rental or B2B service enquiries every day, the same pattern usually repeats:

• The enquiry stays in the inbox. Someone retypes it into the CRM in the evening, or never.
• Details are missing. Date, venue, headcount or technical requirements need to be chased, and the question gets postponed.
• Quotes are built by copying. A salesperson hunts for a similar older quote and retypes items and prices by hand.
• Follow-up depends on memory. There is no task in the CRM, so nobody chases the quote that was sent.

These four spots suit AI well: they are repetitive text work that is easy to check. The flow of enquiries itself is covered in more detail in how to handle sales enquiries fast without losing any.

Step by step: AI-drafted replies on top of your CRM

Start with one type of deal, for example an enquiry for sound equipment hire or for a corporate event. Add further types only once the first one works.

1. Choose one inbox and one deal type. Typically the shared enquiries address. Leave other email aside for now.
2. Define the mandatory details. List what an enquiry must contain to be priced (date, venue, scope, contact person). AI uses this list to spot what is missing.
3. Map the CRM fields. Decide which fields and stages data may be written to and what AI may only suggest. Fewer fields is better at the start.
4. Prepare the quoting material. A current price list, standard packages and discount rules. AI should work only from these, not from guesses.
5. Set up the approval step. Who approves replies, who approves quotes, and what happens when the approver is unavailable.
6. Decide what is written back. After approval: the sent email in the history, the quote attached to the deal, a new follow-up task.
7. Measure process indicators. For example the time from enquiry to first reply, the number of deals with no next step, and the share of drafts that went out without edits.

Template: reply to an incomplete enquiry

This is what a draft prepared by AI for approval might look like when an enquiry lacks details:

Dear Ms [Surname],

Thank you for your enquiry about equipment for your company event. To send you an accurate quote, we need a few more details:
– the exact date and time of the event, including set-up and take-down,
– the venue address and whether it is indoors or outdoors,
– the approximate number of guests.

Once we have these, we can usually send the quote the same day. If it is easier for you, we are happy to go through everything by phone.

Kind regards
[salesperson's name]

How to speed up the quote itself while keeping control of prices is covered in how to speed up sales proposals without losing control of your prices.

Approval: a checklist of what must not leave without a person

Approval is not a brake but a safeguard. To avoid slowing things down, it has to be simple and clear.

• Every email to a customer is approved by a person, even if the draft is flawless.
• Every price outside the price list goes to a responsible person for approval, not just to the salesperson.
• Moving a deal to "won/lost" is done by a person.
• Unusual requests (contractual penalties, atypical payment terms, legal questions) are only flagged and handed over by AI.
• The approver sees the sources: the original email, CRM data and the price-list items used, next to the draft.
• Edits feed back into the process: if an approver keeps changing the same thing, fix the rules, not just the individual draft.

The difference between a chat where a salesperson pastes text by hand and a managed process with approval and records is described in ChatGPT or Claude for your business, or a managed AI process.

General principles for connecting AI to a CRM

The exact connection options differ by CRM and by what you have switched on. So a few general rules apply regardless of the system:

Permissions and access

The connection should have its own technical account with minimal rights: read what the task needs and write only to agreed fields. Access must be removable at any time.

One direction of truth

Customer and deal data are taken from the CRM and results are returned there. If a spreadsheet and the CRM disagree, the CRM wins.

Data and privacy

Find out where data is stored and who processes the text. The language model should receive only the text needed for the task. If an external AI API is called, even a server in the EU does not make for a closed computation: the model provider processes the text. For specific questions about processing personal data under GDPR, consult a lawyer.

Audit trail

For every email and quote sent, it should be possible to trace who approved the draft and which source material it was based on.

How BOTrento Chief of Staff handles it

BOTrento Chief of Staff is a managed process for enquiries, quotes and other sales steps. It does not replace your CRM but works on top of it: it reads enquiries from the inbox, prepares a summary, a draft reply, a draft quote and the next step, and hands everything over for approval. Nothing is sent without human approval.

• A pilot on one process: a fixed scope of 1 inbox, 1 record (CRM or spreadsheet) and 1 type of deal for 30 days, with agreed process KPIs.
• Your own API keys: you enter keys from the model provider of your choice (e.g. Anthropic Claude, OpenAI or Google Gemini) and pay for AI directly to them with no markup. Keys are encrypted, only the last 4 characters are visible, and you can remove them at any time. See the help page on your own API keys.
• Data: we store it on servers in Germany (EU). The text needed for the task is processed by the provider you choose; commercial API providers do not use data for training.
• Oversight: 24/7 monitoring, incident response within 4 hours on working days.

In an internal pilot, 35 enquiries, 18 emails, 24 quotes and 15 actions went through the process in 7 days. The process demo shows what such a process looks like step by step; the tiers are laid out on the pricing page.