To choose an AI support chatbot, test it on your own content and judge it on seven things a demo will not show you: a resolution rate with its denominator, what it answers from and how that stays current, whether it shows its source and says when it does not know, where the handover goes, how long setup takes and which channels it covers, security and data residency, and the price at your real volume. The first three decide most purchases. Run your 50 hardest real questions through each finalist and keep the tools that answer from the right page, cite it, and admit the gaps. Price the survivors at your busiest month.
Why the demo is the wrong test
Most buyers pick the tool with the cleanest demo and a monthly price that fits. The demo answers ten questions the vendor chose, from content the vendor cleaned. Six months later your customers are still asking your team about things that are on your website, and every handover arrives without the conversation attached.
The complaints show where it goes wrong, and it is rarely the AI. Across 1,309 complaints in reviews of 18 vendors, pricing was the biggest theme (211), then setup (197), UI and bugs (166), trust (128) and stale knowledge (38). Every tool on your shortlist can write a fluent answer. What separates them is what the bill does at volume, how long your content takes to get in, whether you can trust what comes out, and whether the tool notices when that content changes.
Vendors also stay quiet on the promises that would help you most. In September 2026 Wonderchat read every promise on twelve competitor homepages. On 10 September, none of the twelve promised to show the questions it couldn't answer. That is a question worth putting to every vendor on your list, in writing.
Seven checks before you sign
1. Resolution rate, and the denominator behind it
Vendors report one of two numbers. Deflection rate is the industry's term for conversations that never reached a person, and its denominator hides every customer who gave up, got a link to the help center homepage, or emailed you instead. Resolution rate counts the conversations where the question was actually answered. Deflection rate is the number a vendor reports when resolution rate would be embarrassing. A rate without a denominator is a guess wearing a number.
A real resolution figure also has a spread across customers. Any vendor's headline number sits somewhere on that spread, and your content decides where.
Ask: resolved out of which conversations, over what period, and is the figure a median across customers or the best case? Can the agent keep the thread when a customer asks a follow-up? Red flag: a demo made only of single-question, single-answer exchanges.
2. What it answers from, and how it stays current
An AI agent is only as good as the content behind it. Many failures are gaps in the documentation, not bad AI. As one buyer put it: "Most companies don't realize how much clean up is required in order to implement an AI chatbot."
Out-of-date content is the other problem, and it is harder to spot. Wonderchat's support team puts it this way: the failure seen most often is not invention; it is retrieval pulling an archived policy and answering from it correctly. The citation is real and the answer is out of date.
Ask: which sources can it read (help center, docs, website, files, past tickets), how often does it re-read them, and can you exclude archived pages? Red flag: a manual update process, or a page limit smaller than your help center.
3. Whether it shows its source and admits what it cannot answer
An answer without a source asks the customer to trust the bot. An answer with a link lets them check it, and lets your team audit it. What the tool does when the content runs out matters just as much: a good agent says so and hands over, a bad one guesses.
The strongest version of this check is whether the tool reports its own misses. A list of the questions it could not answer is the most useful report a support tool can give you, because every line is a page your team should write next.
Ask: does every reply show its source? What happens when the answer is not in your content? Can you see the questions it could not answer? Red flag: any promise of zero errors. No honest vendor makes it.
4. Where the handover goes, and what arrives with it
Handover is where most tools fail in front of the customer. One developer described the hard part well: "The part that took the longest to get right was the handoff, getting the bot to recognize it's out of its depth, collect the visitor's contact info, and route the conversation to a human without making the experience feel broken." The first version simply told users to contact support." The customer came for an answer and left with homework.
Check two routes. The first is a live chat inbox, so someone on your team can take over the conversation where it is. The second is your help desk. If your team works in Zendesk or Freshdesk, the conversation should arrive there as a ticket with the full transcript, not as a blank ticket with an email address. An AI agent that sits on top of your help desk is a small project. One that asks you to replace the help desk is a much bigger one.
Ask: does the person who takes over see the whole conversation? Can the agent collect a name and email first? Red flag: "we integrate with Zendesk" with no demo of the ticket that arrives.
5. Setup time and channels
Setup was the second-biggest complaint theme, so ask for a date, not a promise. Small help centers go live quickly and large ones do not. 80% of Wonderchat agents that go live do so the same day, but knowledge bases over 500 pages take a median 15 days to go live. If you are switching tools, budget for rules as well as content. Teams migrating off a chat widget hit the same wall: the export carries the transcript but not the routing rules, so every rule is rebuilt by hand. With Drift retiring on 31 January 2027, the Drift migration guide shows what that rebuild involves.
Then match channels to where your customers actually ask. One agent trained on one body of content and published to your website, Slack, WhatsApp or Microsoft Teams beats a separate bot per channel that drifts out of sync. Confirm which channels sit on the plan you are pricing, because most vendors gate them.
Ask: what is the median go-live for a content set your size? Red flag: each channel sold as a separate product.
6. Security, data residency and access control
If you sell into finance, healthcare, legal or the public sector, this check comes first. Ask for SOC 2 Type II and GDPR documentation, ask where data is hosted and processed, and ask whether a HIPAA BAA is available if you handle health data. A vendor that offers EU data residency should be able to name the region. Larger teams also need single sign-on and PII redaction, which usually sit on the top plan.
Ask: is there a trust center with current reports? Red flag: vague answers about where customer data goes.
7. The price at your volume, not the list price
Pricing was the biggest complaint theme of all, and the list price is only the start. What matters is the bill in your busiest month under each model:
Per resolution: you pay for each conversation the AI resolves. Often cheapest at low volume, and the bill rises as the AI does more of the work.
Per seat: you pay for everyone who logs in, including colleagues who only take the occasional handover.
Flat tiers: a monthly fee for a set allowance of usage. Easier to budget, but check exactly what the allowance counts.
Then add the costs that never appear on the invoice: integration time, the hours spent cleaning content, and the cost of fixing wrong answers, which buyers underestimate most. The guide to AI chatbot cost traps shows where bills grow, and the chatbot pricing breakdown compares the models vendor by vendor.
Ask: what would last month's conversation volume have cost? Are there setup fees? Red flag: a quote that does not show what happens when your volume doubles.
What changes with the size of your team
The seven checks apply to everyone, but their order changes as you grow.
Stage | Check first | Typical setup |
|---|---|---|
Founder or small team | Setup time and price at low volume (checks 5 and 7) | An AI agent with email handover, on its own or on top of a simple help desk such as Help Scout or Freshdesk |
Growing support team | Handover, help desk and CRM connections, and the list of unanswered questions (checks 3 and 4) | An AI agent handing tickets into Zendesk or Freshdesk and leads into HubSpot |
Enterprise | Security, single sign-on, PII redaction and data residency (check 6) | An AI agent on an enterprise plan, or a full service platform such as Salesforce Service Cloud or ServiceNow when the whole operation is moving |
Small teams should also look at the sales side. The same agent that answers support questions can qualify a visitor who asks about pricing and pass the lead to the right person, which matters more when the founder is also the sales team.
How Wonderchat answers the seven checks
Resolution: about 75% of evaluated conversations are resolved without a person. Jortt, a Dutch accounting platform, resolves about 92% of roughly 30,000 monthly inquiries without a person.
Content: the AI Customer Service Agent answers from your help center, docs, website, files and past tickets, in 40+ languages.
Sources and gaps: it shows the source on every reply, says so when the content does not cover the question, and puts that question in the review inbox so you answer it once.
Handover: with the whole conversation, as a ticket or by email follow-up on every plan, in live chat from Basic, and into Zendesk or Freshdesk if you keep them. Use its help desk, or keep yours: tickets with the full chat come on every plan.
Setup and channels: 80% of Wonderchat agents that go live do so the same day. Every channel comes on every plan, Free included: the website widget, search bar, email, Slack, WhatsApp, Microsoft Teams, phone calls and more, and no channel costs extra.
Security: SOC 2 Type II, GDPR, EU-hosted in Frankfurt, HIPAA BAA available. SAML SSO and private crawling come with Enterprise, and PII redaction is available.
Price: a Free plan at $0 with 50 AI replies a month, refilled monthly, with no credit card. Paid plans when you need more AI replies: Starter 500, Basic 3,000, Scale 9,000, Enterprise custom. The meter is AI replies, with no per-seat pricing. Details are on the Wonderchat pricing page.
For the sales side, the AI Sales Agent asks your qualifying questions and routes each lead to the right person, and HubSpot contact sync comes with Basic.
Where Wonderchat loses
If most of your tickets need the agent to do something rather than answer something, such as issue a refund, change a booking or update an account in your own systems, Ada, Decagon or Sierra is the better choice. They take multi-step actions; Wonderchat answers from your content and hands the rest to a person. If your team already runs on Fin, formerly Intercom, its ecosystem is the deepest of any tool here (Salesforce completed the acquisition in September 2026), and the Fin pricing breakdown shows what that costs. And at low volume, Featurebase's per-resolution billing can cost less than a flat monthly plan.
Verdict
Pull your 50 hardest real questions from last month's tickets, follow-ups included, and run them through each finalist on your own content. Keep the tools that answer from the right page, show the source, and say so when the page does not exist. Then price the survivors at your busiest month. For most small and mid-size teams with a help center and a help desk already in place, that test favors an AI agent that answers from documentation and hands over into the help desk you have, at a flat monthly price. Wonderchat is built for exactly that. If your tickets need actions in your own systems, the test will point you to Ada, Decagon or Sierra instead, and you should follow it. For a shortlist to test, see the best AI customer service chatbots compared side by side.
Frequently asked questions
What is the difference between resolution rate and deflection rate?
Deflection rate counts conversations that never reached a person, including customers who gave up. Resolution rate counts conversations where the question was actually answered. Always ask for the denominator: resolved out of which conversations, and over what period.
Does an AI support chatbot replace a help desk?
Usually not. For most teams the best fit is an AI agent that answers first and hands unresolved conversations into the help desk you already use, such as Zendesk or Freshdesk, with the transcript attached.
Which pricing model is cheapest for an AI support chatbot?
It depends on volume. Per-resolution billing is often cheapest at low volume and grows with every conversation resolved. Flat tiers are easier to budget. Per-seat pricing charges for people, not usage. Price each model at your busiest month.
How long does it take to set up an AI support chatbot?
80% of Wonderchat agents that go live do so the same day. Large documentation sets take longer: knowledge bases over 500 pages take a median 15 days to go live. Ask every vendor for its median go-live at your content size.
How can you tell whether an AI chatbot will make things up?
Test it on questions your content does not cover. A trustworthy agent shows the source on every reply, says so when it cannot find the answer, and hands over. Be wary of any vendor that promises zero errors.
