The AI Chatbot That Lowers Your Conversion: When Not to Deploy a Sales Agent

A chatbot suggestion lifts purchase intention by 44.7% for products that are hard to evaluate. The same chatbot, placed in front of a simple product with a small assortment, drags purchase intention down. Same technology, same interaction pattern, opposite revenue outcome.

That finding comes from a peer-reviewed study across four experiments and 1,339 participants, published in the Journal of Business Research (Anggraini, Demoulin & De Kerviler, University of Lille, April 2026, DOI 10.1016/j.jbusres.2026.116182). It is the clearest evidence to date that “should we add a sales chatbot” is not a yes-or-no question. It is a deployment question, and the wrong answer costs you sales you would otherwise have closed.

The finding that should change your roadmap

Researchers varied two things: how hard the product is to evaluate (many sizes, styles, or fit variables versus a straightforward spec) and how large the assortment is (a handful of options versus ten or more). Then they measured purchase intention with and without a chatbot recommendation.

For hard-to-evaluate products, the chatbot helped. Purchase intention rose 44.7% when the agent narrowed choices and translated fit variables into a recommendation. That is the scenario every vendor demo shows you: a shopper overwhelmed by options, an assistant that resolves the overwhelm.

For easy-to-evaluate products with a small assortment, the chatbot hurt. Purchase intention dropped. The product was already simple enough to decide on its own. Inserting an agent between the customer and a three-option choice added a step, a delay, and an implicit signal that the decision was harder than it actually was. Customers did not experience help. They experienced friction dressed up as help.

For easy-to-evaluate products with a large assortment (ten or more options), the chatbot helped again, but for a different reason: at that volume, even a simple spec becomes tedious to scan manually, and the agent’s filtering function earns its keep.

The pattern is a 2x2 matrix, not a spectrum: evaluation complexity crossed with assortment size. Deployment is not “add a chatbot to the funnel.” It is “route this specific product surface into the cell of the matrix where an agent adds value, and route everything else to a human or a plain interface.”

Why this generalizes past retail

We have made this argument before in engineering contexts: govern the deployment, not just the model. Where you place an AI system in a workflow determines whether it helps or hurts, independent of how capable the underlying model is. This study proves the same structure holds in commerce, with a number attached.

It also connects to a pattern we have tracked across the buying funnel. In Walmart’s 3x Conversion Gap, we showed that AI-mediated checkout converts worse than a merchant’s own site because trust does not transfer to an unfamiliar intermediary. In the verification collapse in AI shopping, we showed that agentic shopping breaks down at the point a customer needs to confirm a claim is true. This study adds the missing piece upstream of both: even before checkout or verification enters the picture, the decision to route a customer to a chatbot at all is already a conversion lever, in either direction.

None of those posts argued agents are bad for commerce. This one does not either. The claim is narrower and more useful: agent placement is measurable, and the measurement depends on product characteristics you already know before you write a single line of chatbot logic.

The deployment rule, made operational

Before adding a chatbot to a product page, answer two questions, not one.

Is the product hard to evaluate? Sizing, fit, style compatibility, technical specs a lay customer cannot parse alone. If yes, an agent that translates ambiguity into a recommendation is doing real work.

How large is the assortment? A handful of clearly differentiated options does not need mediation. Ten or more does, even if each option is individually simple to understand, because the cognitive cost shifts from evaluating one option to comparing many.

Cross these two axes and you get four cells, not two:

Hard to evaluate + any assortment size: deploy the agent. This is where the 44.7% lift comes from.

Easy to evaluate + small assortment: do not deploy an agent here. A clean product grid or a three-line comparison table will outconvert a chatbot every time in this cell.

Easy to evaluate + large assortment: deploy the agent as a filter, not an advisor. The job is narrowing volume, not resolving ambiguity.

The failure mode we see most often in the field is deploying the agent everywhere because it is available everywhere. That treats the chatbot as a feature to ship once and forget. The study treats it as a routing decision to make per product line, revisited whenever the assortment changes size.

Do this now

Audit your product catalog against the 2x2 before your next roadmap cycle. For each product line, classify evaluation complexity and assortment size, then check whether your current chatbot deployment matches the cell. If you have a chatbot live on a small-assortment, easy-to-evaluate line, that is not a neutral feature. It is measured revenue loss, and the fix is not a better prompt. It is removing the agent from that surface and giving the customer a clean, static path to checkout.

---

This analysis synthesizes When AI Chatbots Boost vs Hurt Sales (Science Says, summarizing the Journal of Business Research, July 2026).

Victorino Group helps teams decide where an AI agent belongs in the customer journey and where it costs you the sale. Let’s talk.