Can you trust AI to sell your products? Yes, as long as you properly test what you expect from it

"We don't trust AI to talk about our products. And we don't have time to waste testing it."

This is the objection I hear most often from e-commerce leaders. And it's probably the one I understand best. When you sell technical products, or products where your customers' health and comfort are at stake, inaccurate information is not a detail: it means a returned product and a lost customer.

But this objection often rests on a misunderstanding. When people hear "conversational AI", many still picture the chatbot of 2018: a decision tree that recites canned answers and goes round in circles as soon as you step off the script. A modern conversational AI agent is a completely different thing.

In this article, I take the objection apart point by point: what sets an AI agent apart from a chatbot, what can really go wrong, the safeguards that build trust, how to test in under 30 days, and what agentic commerce makes possible next.

1. The misconception: "a conversational AI agent is just a chatbot"

A chatbot follows a script; a conversational AI agent understands intent, reasons over your catalogue and acts on your site. Confusing the two means judging today's technology by yesterday's flaws.

There are three generations to tell apart:

‍

Today's distrust largely comes from the first generation. In a 2021 Verint survey of 1,000 US consumers, 54.5% said they often had to ask for a human after using a chatbot, and over 60% had to re-explain their situation to the agent who took over (Verint, via Business Wire). These figures describe tools that did not understand their users.

Beware of the opposite trend, though: Gartner warns against "agent washing", the practice of rebranding plain chatbots or assistants as "agents". It estimates that only about 130 of the thousands of vendors claiming agentic AI are the real thing (Gartner, June 2025). So the right question to ask a vendor is not "is it AI?" but "what can it actually do that my old chatbot couldn't?"

2. Distrust is legitimate: what can really go wrong

A poorly supervised AI can state false things with complete confidence. Brushing that risk aside would be dishonest.

Hallucinations. A language model produces the most plausible text, not necessarily the most accurate. Even on a tightly framed task, summarising a document while sticking to its facts, leading models still make things up in roughly 3 to 12% of cases, according to Vectara's public ranking (Hallucination Leaderboard, updated May 2026). Left without safeguards on an online store, a model can invent a product feature, a compatibility or a returns condition.

Sensitive sectors. In paramedical products, cosmetics or technical equipment, a mistake costs more than a sale. It can lead to an unsuitable product, a return or a complaint.

The conclusion is not that you should give up on AI. It's that you can't deploy it like a gadget: it needs to be supervised the way you would supervise a new salesperson.

3. The real question: what level of service can you afford without AI?

The right comparison is not "AI versus perfection" but "AI versus what you actually offer your visitors today". And on most sites, what's on offer at night, at weekends or in a foreign language is often… nothing.

One of our customers sells paramedical products online to consumers. A few months after their assistant went live, the CEO asked his sales manager whether the investment was really worth it. His answer:

"Find me an advisor who can handle 1,000 conversations a month, 24/7, in any language, for the same price, and I'll cancel tomorrow."

And about mistakes:

"Yes, it sometimes gets things wrong. So do humans."

That second sentence is the most important one. A tired, new or overloaded human advisor makes mistakes too. The difference is that their mistakes are almost never measured, whereas every AI conversation can be reread, analysed and corrected.

The cost of no advice. The Baymard Institute has tracked an average cart abandonment rate of 70.19% over 14 years (Baymard Institute). Not all of that hesitation comes from an unanswered question, but with complex products, a lot of it does: which size, which compatibility, which product for my situation? Faced with those questions, visitors look elsewhere, or buy the wrong product and return it.

What customers expect. According to McKinsey, 71% of consumers expect personalised interactions and 76% get frustrated when that doesn't happen (McKinsey). A product page, however complete, doesn't ask questions or adapt to each person's needs. An advisor does.

So you don't deploy conversational AI to "do AI". You deploy it to offer a level of service that would be very hard, if not impossible, to fund any other way: extending your sales capacity without hiring new staff.

4. Building trust: the 5 essential safeguards

Trust can't be decreed; it's built with five safeguards, exactly as you would train a new salesperson before leaving them alone in the store.

1. Train the AI on your products. The assistant must answer from your catalogue, your technical sheets and your guides, not from its general knowledge of the internet. This is what reduces hallucinations the most: the AI is no longer looking for the most plausible answer, it's looking for yours.
2. Define its scope. What does it talk about, and what doesn't it? An AI sales assistant has no business giving medical advice, commenting on a competitor or promising a delivery time it doesn't know.
3. Control what it can say. Brand tone, authorised promises, mandatory statements, banned phrasing: whatever you expect from a salesperson, you can impose on an AI, and often more consistently.
4. Allow it to say "I don't know". It's counter-intuitive, but it's the most important safeguard. An AI that acknowledges its limits inspires more trust than one that answers everything.
5. Keep your customer team in the loop. When a request falls outside its scope, the assistant gives the customer the customer service contact details. But your team's real role lies elsewhere: monitoring conversations, spotting those where the assistant's behaviour needs improving, and correcting it. The goal isn't to create a "tier 1" that filters requests, but to extend your sales capacity without hiring.

These principles echo the reference frameworks. The NIST AI Risk Management Framework, published in 2023, organises AI risk management into four functions: govern, map, measure, manage. In Europe, Article 50 of the AI Act has required since 2 August 2026 that users be told they are interacting with an AI, unless this is already obvious (Cooley). Far from being a constraint, this transparency is part of trust.

5. "No time to test": the method in under 30 days

The full testing phase fits into less than 30 calendar days and takes only a few hours of the customer's team's time. Most of the work is done by the vendor, not by you.

Here's what that month of testing looks like:

1. Week 1: scoping and setup. Importing the catalogue and product content, defining the scope, the tone and the cases where the customer should be referred to customer service. On the customer side: one scoping workshop.
2. Week 2: training and internal testing. The vendor tests the assistant on your customers' real questions and on trick cases: similar products, out-of-scope questions, sensitive requests.
3. Week 3: acceptance testing by your team. Your advisors, the people who know your products best, try to catch it out. Their feedback is used to fine-tune answers and rules. This is where the few hours required are concentrated.
4. Week 4: adjustments and controlled go-live. Fixes, final sign-off, then opening to your visitors.

What to evaluate. Evaluate your assistant the way you would evaluate a new intern: what does it bring to the business, what are people asking it, and where does it lose them?

• Its contribution to sales: assisted revenue, plus the conversion rate and average order value of the visitors it has helped.
• What customers come to ask it: a recommendation, a product question, an after-sales question… These intents show what it's really used for, and what's missing from your site.
• What makes visitors drop off: when, and why, a conversation ends without a result. That's where the most useful improvements are found.

The loop doesn't stop at launch. Test, adjust, measure, repeat. Every conversation is data: it reveals the questions your customers really ask, the information missing from your product pages and the needs you hadn't anticipated. That's where your customer team becomes essential: it monitors conversations, detects those where the assistant needs to improve, and uses these real requests to raise its own level of service too.

Trust doesn't come before experimentation. It comes from it.

6. Beyond answers: agentic commerce

Once trust is established, you can go further: from an AI that answers to an AI that acts, following your brand's own way of selling.

McKinsey defines agentic commerce as shopping powered by AI agents "capable of anticipating, personalizing, and automating every step of the process". It estimates that agentic commerce could orchestrate $3 trillion to $5 trillion in global goods sales by 2030, including up to $1 trillion in the US (McKinsey, October 2025).

This movement has two sides. On one side, consumers' agents: general-purpose assistants that search and compare on their behalf. On the other, brands' agents, deployed on their own sites. It's this second lever that lets an online retailer stay in control of the customer relationship.

Custom agents aligned with your sales playbook. An experienced salesperson doesn't just answer questions. They qualify the need, ask the right questions in the right order, know which products to put forward and when to suggest an add-on. With agentic commerce, you can create custom agents that reproduce this method:

• qualify the need with the questions your best advisors ask;
• guide the customer through the site, taking them to the right category rather than just sending a link;
• recommend the right selection, complementary products included, according to your sales rules;
• adapt the journey to each visitor's profile and context.

That's where the e-commerce user experience becomes truly personal: the AI shopping assistant no longer sells "like an AI", it sells like your best advisors. And that, once again, is what sets a conversational agent apart from a chatbot.

7. 7 questions to ask before choosing an AI shopping assistant

Before you sign, ask your vendor these seven questions. The answers will tell you whether you're dealing with a conversational agent or a rebranded chatbot.

1. What sources does the assistant rely on to answer: my catalogue only, or general knowledge?
2. How do we define what it can talk about, and what it must not?
3. What does it do when it doesn't know the answer?
4. What does it do when a request falls outside its scope, and how can my team monitor and improve its conversations?
5. How long does the testing phase last, and how many hours does it take from my team?
6. Which indicators will I be able to track, and can I reread the conversations?
7. Can it follow my own sales method, or does it apply a standard script?

Conclusion: don't trust it, test it

Yes, it's healthy not to trust an AI you haven't tested. But refusing to test it simply because you don't trust it yet means giving up a level of service your customers already expect.

A conversational AI agent is not a chatbot. It is trained, supervised, measured and improved, just like a salesperson. And one month is enough to find out whether it's up to the job for your products.

Want to check it on your own catalogue? Get in touch to set up a test.

This article builds on my LinkedIn series "7 good reasons NOT to deploy a conversational AI".

‍