Search "AI agent for website" and you get four completely different products back, each happy to let you believe it is the same thing as the others. It is not. One helps shoppers while they are still on your page. One answers support questions. One chases leads by email after people leave. One belongs to the shopper, not the store. Pick the wrong one and you will end up solving a problem you do not have.
Here is how to tell them apart, and how to decide which one to install first.
What are the four kinds of AI agent?
| Category | Acts | Surface | Best for | Example vendors |
|---|---|---|---|---|
| Onsite Conversion Agent | During the website session | On-site widget + animated character | Marketing site conversion | Yokaify, Perspective AI |
| Support Agent | When asked | Chat bubble + help center | Post-purchase support | Intercom Fin, Plain AI, HubSpot AI |
| AI Sales Agent | After the visitor leaves | Email + SMS + DM | Outbound B2B / cart recovery | Reply.io, Lemlist, Outreach AI |
| AI Shopping Assistant | Buyer-side, multi-site | Browser extension / app | Comparison shopping | Perplexity Shop, Operator browsers |
The four share the word "agent" and almost nothing else. The common mistake is comparing across categories — judging an onsite conversion agent against an AI sales agent the way you would compare two chatbots. They are not substitutes. They do different jobs at different moments, and the one you need depends entirely on where you are losing people.
How does an AI agent actually work?
Strip away the marketing and all four do the same handful of things, just on different inputs.
It takes in a signal. For an onsite agent, that signal is real-time behaviour: scrolling, dwelling, hovering, heading for the exit. For a support agent, it is the question someone types. For a sales agent, it is an event from your CRM, like an abandoned cart or a form fill.
It decides whether to act, and when. A good agent stays quiet far more often than it speaks, and that restraint is most of what separates a helpful one from an annoying one.
It grounds its answer in your real content. Rather than guessing from whatever the underlying model absorbed in training, it draws on your own product pages, FAQs, and policies, so it talks about your actual catalogue, prices, and rules instead of making things up.
Then it responds in your brand voice and does something useful: a reply, a one-tap coupon, a side-by-side comparison, a booking link, a follow-up email. The differences between the four categories live mostly in the first and last steps — what each one listens to, and what it is allowed to do. The middle is the same idea pointed at a different moment.
What does an AI agent cost in 2026?
Pricing splits cleanly along category lines.
| Category | Pricing model | Typical SMB cost | Typical enterprise cost |
|---|---|---|---|
| Onsite Conversion Agent | Flat per site | see /pricing | $500-$2,000/mo |
| Support Agent | Per resolution | $0.99 per resolution | Volume contracts |
| AI Sales Agent | Per lead / per seat | $50-$200/seat/mo | $100/seat × volume |
| AI Shopping Assistant | Free (consumer-side) | $0 | N/A — buyer-side |
The two models reflect two different cost structures. A flat per-site price fits an onsite agent, where helping one more visitor costs almost nothing. Usage-based pricing fits support and sales agents, where every resolution or every lead carries a real per-unit cost. Neither is wrong. The right one depends on your traffic and on how often the agent actually has work to do. For onsite plans, see /pricing.
How do you choose which one to install?
Six questions, in order. The first one that gives you a clear answer points you to the category.
1. Where is the conversion bottleneck? If people arrive, browse, and leave without converting, the answer is an onsite conversion agent. If they convert but churn fast, no agent will help — fix retention first.
2. What is the main question your agent will answer? "What should I buy, which plan do I need, what do you offer?" is an onsite conversion agent. "Where is my order, how do I return this, why is it broken?" is a support agent. "Follow up with the lead I captured last week" is an AI sales agent.
3. Is the visitor already on your site? If yes, you want an onsite conversion agent or a support agent. If no, you want a sales agent.
4. What is your average order value? Above $200, an AI sales agent for after-session follow-up usually earns its place alongside the onsite agent. Below $50, that follow-up is harder to justify, so put your effort into the onsite layer. Between $50 and $200, run the onsite agent first, measure for 60 days, then decide.
5. How big is your support inbox? If your team answers more than 50 chats or emails a day, a support agent can deflect 60-80% of them and pay for itself in saved hours. If you are answering fewer than 10 a day, the maths is weak — reactive chat or a solid FAQ page is enough.
6. What is your team's technical capacity? For the SMB tier, every named agent installs with a script tag, so the effort is roughly equal. For custom work (multi-site, white-label, API-driven), the gap between vendors is wide — read the API docs before you sign.
The 2026 vendor map
Sorted by category, then by rough 2026 market share within each. Inclusion is not endorsement.
Onsite conversion agents
- Yokaify — an animated character that reads behaviour and steps in at the right moment, with answers grounded in your own site content. Free tier; see /pricing for SMB-to-mid-market.
- Perspective AI — form-replacement and concierge framing. Strong in B2B SaaS demo-request flows.
- Conversion.ai — uses the "Conversion Agents" name (May 2026) but the product is marketing automation, not in-session help.
- Drift — distressed after the Salesloft acquisition and a 2024 OAuth breach. Still in-market, but losing share.
Support agents
- Intercom Fin — AI add-on to the Intercom Messenger. Strongest help-centre integration.
- HubSpot AI — bundled with HubSpot Service Hub. A strong CRM-native option.
- Plain AI — a modern alternative, good for technical SaaS support.
- Zendesk AI — the incumbent, with a large enterprise install base.
AI sales agents
- Reply.io — outbound sequences with AI-generated personalisation.
- Lemlist — similar, more mid-market focused.
- Outreach AI — enterprise outbound and sales engagement.
- Apollo — AI BDR with built-in lead enrichment.
- Klaviyo AI — ecommerce-native, cart-recovery focused.
- 11x — a fully autonomous AI SDR.
AI shopping assistants (buyer-side)
- Perplexity Shop — shopping driven by queries inside the Perplexity assistant.
- Operator browsers — Anthropic, OpenAI, and Brave variants.
- Walmart, Target, Shopify protocol — the emerging agentic-commerce protocol set.
The buyer-side category does not affect most merchants directly today. That will change. As agentic commerce protocols mature over the next couple of years, the stores that already publish a clean product feed and structured data will have a much cheaper transition than the ones scrambling to retrofit it.
How do you roll one out in the first 30 days?
Day 1 — pick a category. The six questions above answer this.
Days 2-3 — install. For Yokaify and most peers, this is a single script tag and a quick auto-scan of your site. More complex tools (Outreach, Apollo) take one to two weeks of CRM integration.
Days 4-7 — configure. Set your brand colours and the character's personality, then add guardrails ("never quote competitor prices," "always send billing questions to support"). Review the auto-generated knowledge base and correct anything the agent should handle differently.
Days 8-14 — run a holdout. Keep the agent live for half your visitors and off for the other half. Measure conversion rate, time on site, support volume, and (for support agents) deflection rate.
Days 15-30 — read the result and tune. Pull the lift number from the holdout, refine the rules, and widen the set of moments the agent is allowed to step in. From here it mostly runs itself.
By day 30 you have a clean, measured lift number. That number is the input for every decision after it: whether to upgrade your plan, add a second agent, or expand where the first one operates.
Further reading
- GuideHow onsite conversion agents workThe deepest category in this guide.
- GuideThe proactive chat playbookThe behavior-driven side of onsite agents.
- GuideConversion-rate optimisation referenceWhere AI agents fit in the broader CRO playbook.
- ToolCart abandonment calculatorModel the recovered revenue from in-session intervention.
Frequently asked questions
Software that watches how a visitor behaves, decides whether and how to engage, and then acts. The 2026 category breaks into four sub-types: onsite conversion agents, support agents, AI sales agents, and AI shopping assistants.
Last updated May 31, 2026.