Learn
Should AI answer questions from people who have not bought yet?
Does it do X, does it integrate with Y, is it right for us. Support-shaped, but really sales.
High value and high risk — an overclaim made to close a sale becomes a refund later.
These arrive at support and are really sales. Someone has not bought yet and wants to know whether your product does the thing. A fast answer converts; a wrong answer becomes a refund and a review.
Why the economics are different
Every other queue on this site is a cost centre. This one has revenue attached, which changes the calculation: a slower, more accurate answer may be worth more than a fast deflection, and speed genuinely matters because the person is deciding now.
The specific risk
Overclaiming. A model asked whether you integrate with a system you half-integrate with will find the most encouraging phrasing in your marketing copy and repeat it. The customer buys, discovers otherwise, and you have converted a lost sale into a refund plus a bad review.
This is why constrained-generation platforms exist. Netomi and Certainly restrict output to sanctioned answers for exactly this reason, and for brands where a capability claim has legal weight that constraint is the product.
What to do
- Write a sanctioned answer set for capability questions and let the AI use only that.
- Make the AI say it does not know rather than infer from marketing copy. Inference is where overclaims come from.
- Route anything about compliance, security or contractual terms to a human, always.
- Track refunds and cancellations attributed to expectation mismatch, not just conversion.
The upsell question
Some platforms — Zowie most explicitly — recommend products inside the conversation and report revenue influenced. Whether you want that is a brand decision rather than a technical one. Upselling someone who arrived with a problem reads badly; helping someone choose between two products does not.
Measuring it properly
| Metric | Why |
|---|---|
| Conversion after AI answer | The actual point |
| Refund rate on AI-influenced sales | Catches overclaiming |
| Escalation rate on capability questions | High is good here |
| Time to first answer | Speed converts on pre-sales |
Set the escalation threshold lower on pre-sales than anywhere else. A human closing the sale beats a bot losing it, and the cost of a wrong answer is paid twice.
Tools That Actually Do This
All three emphasise sanctioned, brand-safe responses over open generation.
Netomi
Enterprise agent that markets heavily on brand-safe, sanctioned responses rather than open generation.
Certainly
Conversational AI with an e-commerce lean, though it sells across sectors.
Zowie
Resolution-focused e-commerce agent that also pushes for revenue rather than deflection alone.
Frequently Asked
Should AI answer pre-sales questions?
For documented capability questions, yes — speed converts. For anything requiring inference about what your product can do, no, because overclaims become refunds.
What is the risk with AI answering pre-sales questions?
Overclaiming. A model will find the most encouraging phrasing in your marketing copy and repeat it, turning a lost sale into a refund and a bad review.
Which tools are safest for pre-sales AI?
Netomi and Certainly restrict output to sanctioned answers rather than generating freely, which is the direct answer to the overclaiming risk.
Should AI upsell in a support conversation?
A brand decision, not a technical one. Helping someone choose between products reads fine; upselling someone who arrived with a problem does not.
How do I measure pre-sales AI?
Conversion after an AI answer, and refund rate on AI-influenced sales. The second catches overclaiming that conversion alone hides.
What should always escalate on pre-sales?
Compliance, security and contractual questions. Those answers have legal weight and should come from a person.
Is a high escalation rate bad here?
No — on pre-sales it is usually good. A human closing the sale beats a bot losing it.
Related Questions
What is customer service, and what does it actually cost?
YesThe function that answers customers after the sale. Simple to define, and the second-largest operating cost in a lot of businesses.
What are the different types of customer service?
YesEight delivery channels, each with a different cost per contact and a different answer to whether automation helps.
How does AI actually augment a support team?
YesFour levels of automation, from suggesting text to acting on your systems, each with a different cost and a different failure mode.