An AI Sales Assistant That Helps Website Visitors Choose

Most visitors who leave are unsure, not uninterested

Watch a session recording of a visitor who didn’t buy and you will often see the same pattern: they open three product pages, flip between them, scroll the comparison chart twice and leave. They weren’t rejecting you. They couldn’t work out which option was theirs.

Take a company selling water filtration equipment in three lines: residential units, light commercial systems and whole-building installations. A coffee shop owner lands on the site and asks, in effect, “Which one do I need?” The product pages answer in flow rates, grain capacities and port sizes. The owner knows how many drinks they sell on a busy morning and that their espresso machine’s manufacturer warned about scale. A human salesperson bridges that gap in two minutes of conversation. A website sales assistant is built to perform the same translation, from the buyer’s situation to your specification.

That makes it a different tool from its neighbors. A support assistant helps people who have already bought. A lead qualification chatbot decides whether a prospect deserves a rep’s time. A sales assistant’s job is narrower and sits closer to the purchase: help this particular buyer choose correctly, then get them to checkout or to your team with that choice already made.

Writing down what your best salesperson knows

Point an AI model at your product pages and you get an assistant that can recite specs. It cannot sell, because the knowledge that makes a salesperson good is mostly absent from your website. It lives in their head: which customers the entry model disappoints, which add-on nearly everyone ends up needing, which question signals a job that needs a site visit.

So the most important deliverable in a sales assistant project is a document we call the fit sheet, built largely from an interview with whoever sells best for you. It holds:

• Fit rules in plain if-then form, such as: a café above a certain daily volume of espresso drinks needs at least the light commercial line.
• Disqualifiers that send a buyer to a person; well water with heavy iron content, for instance, changes the whole system design.
• Honest weaknesses of each option, so the assistant can say what a product is not good at.
• Common confusions buyers arrive with, like two model names that differ by a single letter.
• Things you don’t sell, and where to send people who need them.

The fit sheet sits alongside structured data that the assistant reads directly:

Keeping the fit sheet as a readable document rather than hidden prompt text means your team can see exactly what the assistant believes and correct it without calling a developer.

Ask only the questions that change the recommendation

Salespeople learn to avoid interrogations, and an assistant should too. Our rule for discovery questions is blunt: if the answer would not change what we recommend, the question gets cut. Most product decisions hinge on two to four facts.

For a company selling scheduling software in three plans, the deciding questions might be how many staff members take bookings, whether clients must pay a deposit when they book, and whether the business runs more than one location. Everything else, like industry or how they found you, may interest marketing but has no bearing on which plan fits. Those three questions come with quick-reply buttons for the common answers and a free-text box for the rest, and a visitor who would rather skip ahead can ask for the comparison outright.

The assistant also uses what the page already tells it. Someone who opened the chat from the multi-location plan page does not need to be asked whether they have more than one location; the assistant can confirm that assumption in passing. Small touches like this separate a conversation that feels attentive from one that feels like a form with a chat bubble drawn around it.

One recommendation, with the reason attached

Once it has what it needs, the assistant commits: one primary option, occasionally one alternative, and never a list of eight products to sort through, since that is the exact problem the visitor came in with. Every recommendation carries its reason in terms of what the visitor said: “Because two of your staff take bookings and you want deposits, the middle plan covers you; the entry plan doesn’t support deposits.”

It explains the trade-off honestly as well. If the recommended option costs more than the visitor hoped, the assistant can describe what they would give up by going lower, drawing on the honest weaknesses in the fit sheet. It does not hide the cheaper option to push the bigger sale.

Sometimes the right answer is that you are not a fit. A custom cabinet shop’s assistant, asked for budget flat-pack cabinets delivered next week, should say plainly that the shop builds to order and describe what that involves, rather than steering the visitor into a consultation that wastes both sides’ time. Buyers remember the business that told them the truth, and your salespeople stop spending afternoons on calls that were never going to close.

Prices, discounts and promises follow your rules, not the model’s

Language models are agreeable by nature, and an agreeable model with no rules will happily invent a discount to keep a buyer happy. Pricing is therefore the most tightly controlled part of every sales assistant we build. You choose one of three policies:

1. Published prices. The assistant quotes prices only from your price feed, exactly as listed, and never calculates a figure you have not approved, such as an improvised bundle total.
2. Drivers, not numbers. For custom or quoted work, it explains what moves the cost, such as dimensions, materials or site conditions, and routes the buyer to someone who can quote.
3. No pricing talk. Price questions go straight to a person, which suits businesses whose pricing depends on negotiation.

Promotions come from a list with start and end dates; if a code is not on that list, it does not exist as far as the assistant is concerned. Delivery or lead-time statements come only from real inventory or scheduling data. Warranty and return terms are quoted from your published policy, word for word where precision matters. Comparisons with competitors are limited to statements you have approved, because a confident but wrong claim about another company’s product is a liability, not a sales tactic.

Before launch we test all of this with scripted buyers who push hard: “Your competitor is cheaper, match it,” “Knock off ten percent and I’ll order now,” “Promise me it arrives by Friday.” The assistant has to hold its line politely every single time.

Inside an online store: search, compare and add to cart

On an ecommerce site the assistant can do more than advise. Through the store platform’s API it can search the catalog in plain language (“a replacement cartridge for the under-sink unit I bought two years ago”), check stock, place two products side by side and add the chosen item to the visitor’s cart.

Compatibility is where this earns its keep. Plenty of returns happen because a buyer ordered a part that doesn’t fit what they already own. When you maintain a compatibility table, the assistant checks it before recommending anything and says clearly when it cannot confirm a fit.

Cross-selling follows rules you write. We separate required accessories, such as the mounting kit a unit cannot be installed without, from optional extras. Mentioning the first kind is a service to the buyer, so the assistant always does it. The second kind is offered once at most, and not at all when the buyer has signaled a tight budget.

Payment never happens in the chat. The assistant sends the buyer to your store’s normal checkout, where card details are handled by your payment processor exactly as they are today.

Handing a buyer to your team with the notes already written

Many purchases, especially large, custom or business-to-business ones, need a person to finish them. The assistant hands off when the fit sheet says to, when an order exceeds a size you set, when the visitor asks for a human, or when the conversation keeps circling in a way that suggests the buyer is confused.

The handoff creates a record in your CRM with a note your rep can read in under a minute: what the buyer is trying to accomplish, the option the assistant recommended and why, each discovery question with its answer, worries or objections the buyer raised, and precisely what the buyer was told would happen next. If you like, the assistant books a call straight onto the rep’s calendar, respecting the territory or round-robin rules you already use.

That note is what changes a rep’s day. Instead of a form submission that says “interested in pricing,” they open a conversation already knowing the buyer runs a two-location clinic, needs deposits and is nervous about migrating existing appointments. The first call starts at the real question.

Tracing revenue back to conversations, honestly

Every conversation gets an ID that travels with the cart or CRM record, so an order or closed deal can be traced to the chat that preceded it. The tempting next step is to compare the conversion rate of visitors who chatted against those who didn’t and declare victory. That comparison overstates the effect, because people who open a sales chat are usually more motivated to begin with.

A fairer test is a holdout: the assistant is shown to most visitors and hidden from a random slice, and the two groups are compared over the same period. We set that up when your traffic is high enough for the comparison to mean something, and we tell you plainly when it isn’t.

Alongside revenue we track the signals that explain why the number moved: how often buyers accept the recommendation, how many handoffs become real opportunities in your CRM, the share of conversations ending in “not a fit,” questions the fit sheet couldn’t answer, and conversations where a buyer left right after a price appeared.

A generic sales widget, or an assistant built on your sales process?

A generic chat plugin reads your website text and answers in a pleasant tone. For a business with a handful of simple offerings that can be enough, and we will say so. It falls short when choosing correctly depends on rules nobody wrote on the website, when prices must come from a live feed, when the handoff has to land properly in your CRM, or when a single invented discount would be expensive.

EVOTECH IT LLC builds the custom kind. Our US-based team works from the Houston area and delivers remotely nationwide, starting with a free consultation by phone or video. From there we interview your top seller to build the fit sheet, connect the catalog, price data, cart and CRM, test the discovery and handoff flows against scripted buyers, launch with measurement in place and go through transcripts with you on a regular schedule. Before anything is built, you get a fixed-scope written quote listing what will be delivered. With over two decades in business and a 5.0 Google rating, we would rather lose a project than sell you a build your sales process doesn’t need.

For how an assistant fits into a wider funnel, see our page on chatbots for sales funnels; for the technology underneath, our AI agent development service.

Related services

Frequently asked questions