The right add-on, at the right moment:
never a generic pitch

The process today

Most online stores run one upsell mechanism for every customer: a banner, a pop-up, a fixed “customers also bought” row, shown the same way regardless of what the person is actually looking at. Teams that test this usually find a large share of customers simply tune it out, because the suggestion was never built around their specific order.

The second problem is timing. A relevant add-on suggested at the wrong moment, before a customer has decided on the main item, or after checkout when the cart is already closed, reads as noise rather than help. The same suggestion, placed right after a customer confirms the item they actually want, lands completely differently.

The third cost is overreach. A team that notices upselling works sometimes pushes too hard, suggesting an add-on in every message, which trains customers to ignore the business’s messages in general, not just the upsell ones. The signal gets lost because there was never a limit on how often it fires.

What the agent does

The agent reads what a customer is asking about or has already put in the cart, checks it against bundle and relevance rules your team built from the actual catalogue, and suggests one specific add-on at the moment it fits, usually right after the customer confirms their main choice in chat or at checkout. The suggestion names the actual product and the actual reason it pairs with what they are buying, not a generic “you might also like” line.

If the customer declines or ignores the suggestion, the agent does not repeat it in the same conversation. One relevant offer, clearly stated, is the rule, not a sequence of nudges. For stores that sell through chat as well as a storefront, the same logic runs in Telegram, WhatsApp or web chat, so a customer asking a support question about one product can still get a relevant suggestion without it feeling like a sales interruption to their actual question.

Post-purchase, the agent can send one follow-up a few days later for consumables or recurring items, a refill, a companion product, timed to when the customer would plausibly be ready for it rather than immediately after the first order lands. Every suggestion made, and whether it was accepted, is written to your CRM so your team can see which combinations actually convert and retire the ones that do not.

What stays with humans

The bundle and relevance rules are built and approved by your team, not guessed by the agent, and any pricing change, discount, or custom bundle outside the approved list goes to a person. The agent suggests; it does not negotiate price or create a one-off deal on its own.

Guards

Every suggestion is logged against the rule that triggered it and whether the customer accepted it, so your team can audit what is actually working. New bundle rules go through a review step before going live, and a frequency cap keeps the agent from suggesting more than once per conversation regardless of how the rules are written. A kill switch turns suggestions off instantly if something looks off in the data.

Price and timeline

Running cost is usually $20 to $70 a month in model usage, depending on conversation volume.

Related

Pair this with abandoned cart recovery to catch the customers who did not finish checking out at all, and with win-back of inactive customers once a customer has gone quiet for longer than a single cart. See the full package breakdown on the AI agents service page. For how targeted selling moved real revenue, see the Balkans supplements store case study, where average order value rose after the funnel was rebuilt, and the sports nutrition relaunch case study.

Want relevant add-ons suggested automatically, without the pitch fatigue? Get in touch and we will look at your catalogue for the obvious pairings first.

Tired of doing this by hand? We can take the whole routine off your team, not just this step: Routine takeover, from $400 →