How to Draft Inquiry Replies with AI Safely: Guardrails and a Pre-Send Checklist

BOTrento Team8 November 20268 min read

What can go wrong in an AI-drafted reply

AI drafts read fluently, and that is the risk. A wrong sentence in a polished email looks just as trustworthy as a right one. In sales replies the errors tend to fall into the same groups:

• Invented or stale facts: a price, a lead time or a spec that sounds plausible but is not in your price list.
• Accidental commitments: “we can deliver on the 14th” or “yes, that is included” when nobody has checked availability or scope.
• Liability wording: promises of results, guarantees or absolute statements (“fully compliant”, “no risk”) that your contract would never support.
• Wrong tone: too casual for a procurement manager, too stiff for a repeat customer, or cheerful about a complaint.
• Misread enquiry: the draft answers the wrong question or ignores a condition buried in the third paragraph.
• Wrong context: details from another customer's thread leaking into the reply.

None of these are reasons to avoid drafts. They are reasons to decide in advance who checks what.

Five guardrails to set before the first draft

1. Nothing is sent without a person approving it

This is the one non-negotiable rule. The AI prepares, a human sends. For a risk-based view of which emails need a quick look and which need the owner, see our article on an AI draft approval workflow for sales teams.

2. Prices and dates come from your sources, not from the model

Tell the team that any number in a draft is a placeholder until it is matched to your price list, your quote or your calendar. If the draft contains a figure you cannot trace, delete it.

3. Define commitment levels

Agree three levels of wording: information (“our standard setup usually includes…”), intention (“we will check availability and come back by Friday”) and commitment (“we confirm the 14th at this price”). Drafts may use the first two. Only a person writes the third.

4. Give it a banned-phrases list

Words such as “guarantee”, “always”, “fully”, “no additional costs” and “legally” should trigger a manual rewrite. Add your own from past contract disputes.

5. Limit what history it sees

Give the AI the thread it is replying to and relevant notes, not the whole mailbox. Our piece on using customer history in AI replies covers what to include and what to keep out under GDPR.

The pre-send checklist (about one minute per draft)

Print this or paste it into your team wiki. Every box must be ticked before a draft is sent.

1. Did it answer the actual question? Re-read the enquiry, then the draft. Check conditions, deadlines and attachments mentioned by the customer.
2. Is every price, quantity and date verified? Match against your price list, quote or availability. Check VAT wording: are amounts stated excl. VAT?
3. Is anything promised that you have not confirmed? Look for “yes”, “we can”, “included”, “by Tuesday”.
4. Is the liability language safe? No guarantees of outcomes, no legal or compliance claims, no waiving of terms.
5. Is the scope clear? What is included, what is not, and what you still need from the customer.
6. Is the tone right for this person? New contact or long-term customer? Formal or relaxed? Complaint or routine?
7. Is there anything from another customer? Names, projects, figures.
8. Is the next step clear? One action, one owner, one date you are actually able to meet.

A draft that fails points 2, 3 or 4 goes to the owner or the person who handles pricing, not back to the AI for another try.

Risk areas that need a closer look

Pricing and quotes

If a draft gives a price range, check that the range matches how you really price. Rental firms should check availability before the draft says anything about dates. Event and production teams should watch scope: a draft that says “includes technical support” may be promising more than your standard package covers.

Liability and legal wording

Anything about warranties, damage, insurance, data protection or compliance should be written or approved by someone who knows your terms. For specific cases, ask a lawyer. A short rule helps: if the sentence would matter in a dispute, a person writes it.

Tone

AI tends towards enthusiasm. That is fine for a first enquiry from a new lead, and wrong for a customer who is annoyed about a delay. Keep two or three tone notes (for example “direct, polite, no exclamation marks”) and apply them in review.

Repeatable versus messy enquiries

For identical routine questions, a fixed template may be safer than a generated draft. For detailed, messy requests, a draft saves time but needs more checking. We compare the two in email templates vs AI drafts.

Example: a draft before and after review

Suppose a customer asks about equipment for a conference in three weeks. The raw draft says:

Hi Anna, great news, we can definitely supply everything you need for the 14th. The package is fully insured and there are no extra costs. We guarantee a flawless event.

A reviewer would flag availability, “fully insured”, “no extra costs” and “guarantee”. The approved version might be:

Dear Anna, thank you for your enquiry. We are checking availability of the items you listed for the 14th and will confirm by Friday noon. Our quote will state what is included and what is charged separately, with all amounts excl. VAT. Could you confirm the venue's power access and the setup time?

The second version commits only to an action you can control, asks for missing information and moves the next step forward.

How BOTrento builds these guardrails into the process

BOTrento Chief of Staff is a managed process, not a free-form chat window. It watches enquiries and quotes in your mailbox and uploaded data, and prepares draft replies, quotes and follow-ups in four languages (CS, EN, DE, FR). It never sends anything on its own: every email and quote draft is approved by a person. It also does not confirm dates or bookings, and it works from the text of the email, not attachments.

It also catches promises such as “I will get back to you by Friday” and reminds you, and each working morning it shows a short Dnes 1+3 overview so you know what needs checking first. AI usage is included in the plan within a monthly limit, so there is no surprise charge. If you want to see how the steps fit together, look at the process demo or read what Chief of Staff does. For data questions, our GDPR checklist for AI sales tools goes through what to ask a vendor.

You can test it on your own enquiries: the trial is 14 days, free, with no card and no API key of your own. It starts when you first upload data, and you have 7 days to prepare. You can start the 14-day trial and a person from our team will get in touch within 24 hours on working days to help connect your mailbox.

Make the checks stick in your team

• Name the approver. Decide who signs off which type of reply: the salesperson for routine, the owner for pricing exceptions and legal wording.
• Review mistakes weekly. Keep a short list of what reviewers corrected, then add those items to your banned-phrases list.
• Keep sources current. A tidy price list and availability calendar matter more than the AI. Drafts can only be as accurate as what they refer to.
• Speed still matters. Careful review should not mean slow replies. Our answer on replying to sales enquiries faster covers the rest of the workflow.