AI customer service software for small business: priced by billing unit, grounding and hand-off

Most lists of AI customer service software rank features. The thing that actually separates these tools for a small business is what the AI bills on, because that decides whether your invoice tracks the good answers or every attempt, and whether it can run up while you sleep. After that come two questions that decide how much you can trust it: where does it get its answers, and what happens when it doesn't have one?

This guide sorts AI customer service tools by those four things, then prices the same two teams on each: three people with 300 AI-handled conversations a month, and five people with 1,000. It names who each tool is for and where each one falls short. We make Talvo, so it gets its own section below with its gaps stated, and every competitor figure is the vendor's own published price or policy as of October 21, 2026. Recheck before you buy.

If you want the full rate card for ten vendors, read AI resolution pricing compared. If you're leaving Fin, start with Fin AI alternatives. If you're choosing live chat first and AI second, the best live chat software for small business is the better door. This post is for people who start from the AI.

The short answer, by situation

Four questions that sort every tool

Ask these before you read a feature list.

What does it bill on? A resolution, an outcome, a conversation, a session, a credit, or nothing at all. The same word means different things, which is why we compared the definitions separately.

Where do its answers come from? The best case is that it reads articles you published and says which one it used. The worst case is that it fills gaps from general knowledge. Ask the vendor to show a live answer with its source, on your content, in a trial.

What happens when it's unsure? A good agent hands off with the whole conversation. A bad one apologizes and tries again. Our hand-off guide lists the rules and a ten-message test.

Can you cap the spend? Some tools stop the AI at a limit you set. Others keep metering.

Billing units, in one table

This is the compact version. The full rate card, with every vendor's wording, lives in AI resolution pricing compared.

The pattern in the unit column is the one to remember. Fin, Zendesk, Help Scout and Talvo charge for finished jobs. Tidio, Crisp, Freshworks and Tawk.to charge for activity: what counts is the AI taking part, not the outcome. Crisp says escalated conversations still use credits, and Freshworks counts sessions by time window rather than by resolution, so a conversation the AI hands to a human can still cost you. Olark charges for nothing by the answer. Cheap per-unit activity pricing can be the expensive choice if your AI hands off a lot.

What the same team pays

Two teams, same assumptions throughout. We count conversations the AI handled, and for the vendors that bill finished jobs (Talvo, Fin, Help Scout and Zendesk) we assume 60% become billable resolutions. That 60% is our assumption for easy arithmetic, not a benchmark, and not a vendor figure. For the activity-billed tools (Freshchat, Tidio, Crisp) we count every handled conversation. Prices are list prices for the cheapest plan that includes the AI and chat, with seats priced per person where the vendor does that.

The arithmetic, for the ones that need it.

• Talvo, 3 people and 300 handled: 60% is 180 resolutions. Starter (up to 5 teammates, 100 resolutions included) is $49 + (180 − 100) × $0.49 = $49 + $39.20 = $88.20 monthly, or $39 + $39.20 = $78.20 annually. Growth would be $149 or $119, so Starter wins. 5 people and 1,000 handled: 600 resolutions. Starter would be $49 + 500 × $0.49 = $294. Growth ($149, 500 included) is $149 + 100 × $0.49 = $198 monthly, or $119 + $49 = $168 annually. Scale is $299 or $249, so Growth wins.
• Fin on Intercom Essential: seats are $39 monthly or $29 annual per person. 3 × $39 = $117 plus 180 × $0.99 = $178.20 gives $295.20; annually 3 × $29 = $87 plus $178.20 gives $265.20. For five people: $195 + $594 = $789, or $145 + $594 = $739.
• Help Scout Standard: $30 monthly or $25 annual per user. 3 × $30 + 180 × $0.75 = $90 + $135 = $225; annually $75 + $135 = $210. Five people: $150 + $450 = $600, or $125 + $450 = $575. New accounts get three months of unlimited AI resolutions.
• Zendesk Suite Team: $69 monthly or $55 annual per agent, with 5 resolutions included per agent a month. Monthly, pay-as-you-go at $2.00: 3 × $69 = $207, plus (180 − 15) × $2 = $330, is $537. Annual with committed resolutions at $1.50: $165 + 165 × $1.50 = $412.50. Five people: $345 + 575 × $2 = $1,495, or $275 + 575 × $1.50 = $1,137.50.
• Freshchat Growth: $23 monthly or $19 annual per agent, plus Freddy at $49 per 100 sessions. 300 sessions is $147, so 3 × $23 + $147 = $216, or 3 × $19 + $147 = $204. 1,000 sessions is $490, so 5 × $23 + $490 = $605, or 5 × $19 + $490 = $585. The first 500 sessions are free once per account, which we leave out.
• Tidio Growth plus Lyro: Lyro is $210 for 300 conversations and $700 for 1,000 monthly, or $175 and $583.33 on annual billing. We add the Growth plan for 600 and 1,000 human-handled conversations ($129 and $179 monthly), an extra assumption of ours that moves this row more than any other.
• Crisp: Mini is $45 for 4 seats with $5 of credits, and Essentials is $95 for 10 seats with $25. At Crisp's $0.05 to $0.10 per conversation, 300 conversations use $15 to $30, so the bill is $45 + ($15 to $30 − $5) = $55 to $70. For 1,000: $95 + ($50 to $100 − $25) = $120 to $170. Crisp doesn't publish its pay-as-you-go rate, so we assumed the same price per conversation. Its per-conversation figure is also Crisp's own estimate.
• Olark: $400 a month billed annually includes one human seat; additional agents are $29 each a month. Three people: $400 + 2 × $29 = $458. Five: $400 + 4 × $29 = $516. Unlimited conversations, no usage fees.

Read the table for shape, not for a winner. Crisp is the lowest line on paper, and it's also the one whose AI cost rests on an estimate and an unpublished rate, and whose Mini plan doesn't include Crisp's knowledge base. Zendesk's bill comes from a per-resolution rate three to four times Talvo's and a seat price that starts higher. If your resolution rate is nearer 100% than 60%, the per-resolution lines grow by about two-thirds of their AI portion.

The calculator uses list prices for Talvo, Intercom Essential, Zendesk Suite Team, Help Scout Standard and Crisp, with AI resolutions as the second number, so it will differ a little from the table above, which adds estimates and footnotes.

Grounding and hand-off, tool by tool

Each summary below comes from the vendor's own pricing or help pages.

Fin resolves against your content and escalates when a customer clearly asks for a person, shows strong frustration or loops, according to its help page. It escalates once per conversation, and if no teammate is available the customer gets a refusal with no fallback, so staffing the hand-off matters. Its pricing page says you pay nothing if Fin can't answer or the customer asks to speak to a human, though a completed Procedure that ends in a hand-off does count as an outcome.

Help Scout voids its charge when the customer uses escalation, keeps searching the knowledge base or says they still need help. That bills closer to "the customer didn't object" than to "the answer was right", but a charge that goes away on escalation lines up your cost with hand-offs.

Zendesk sorts AI conversations into tiers, and its help article describes an LLM reviewing the transcript to mark a resolution as verified. The article implies, without saying it in one line, that only verified resolutions count against the allowance, which would make it one of the strictest definitions in this group. It also lets you pause the AI at the limit.

Crisp, Tidio and Freshworks bill the activity. Crisp says escalated conversations still use credits. Tidio doesn't count a conversation where Lyro hands off immediately, but one where Lyro replied first does count. Freshworks counts a session per 24-hour window on chat and doesn't tie it to a resolution. That's not wrong, but it means you can pay for the misses.

Chatbase and similar standalone bots answer from your content, but the hand-off varies. Chatbase has its own helpdesk inbox on Standard and up, and can escalate into Zendesk, Freshdesk, HubSpot and other tools, often by opening a ticket. SiteGPT's Starter has no live takeover. Our pricing guide for AI website chatbots prices them at the same volumes and lists what they don't include. At 1,000 conversations, assuming four bot replies each, Chatbase's Standard plan is $150 monthly ($120 annual) on a 1-credit model, SiteGPT's Starter is $59 ($39), and CustomGPT.ai is $499 ($449).

For citations, what a vendor shows in the widget is what counts, so ask to see it in a demo. Talvo's version is in the next section.

Seven small-team options, one weakness each

Tawk.to belongs near this list for a different reason: its core live chat is free with unlimited agents, and its AI Assist is a paid add-on in credits. On the Auto model one reply costs 3 credits, so the 1,000 credits in the $29 tier is about 333 replies (our arithmetic). That's workable for a low-volume site, and Tawk.to's Security page lists no SOC 2 or ISO 27001 certification.

Talvo for AI customer service

Talvo is a shared inbox, live chat widget, email, help center and AI agent in one workspace. The price is per workspace, not per seat. Plans are Starter at $49 a month ($39 on annual billing, up to 5 teammates, 100 AI resolutions included), Growth at $149 ($119, up to 15 teammates, 500 included) and Scale at $299 ($249, up to 25 teammates, 2,000 included). Resolutions past the included amount are $0.49 each.

How it answers: only from your published articles, and it shows the article it used. If the articles don't cover the question, there are no published articles or the AI call fails, it hands the conversation to your team with the thread intact, and an internal note says why. A reply to a real question that cites no article becomes a hand-off. You can also list topics, such as refund or cancel, that always go to a person. A hand-off is never billed as a resolution.

At the volumes above, that comes to $88.20 for three people and $198 for five on monthly billing. It also gives you the inbox, help center and live chat, which the standalone bots lack. At 3,000 conversations (1,800 resolutions at 60%), Scale's 2,000 included resolutions cover it at $299 monthly or $249 annually.

Hand-offs land in the shared inbox, where the thread is intact for whoever picks it up.

Where the others beat us: Help Scout has a cap, Zendesk has a pause at the allowance, and Fin has spending limits and alerts. Talvo has none of those. If a predictable maximum matters more than a lower rate, that is a real point against us. The Free plan is also a trial more than a plan: 2 teammates and 25 AI resolutions a month during the beta. How AI auto-resolve works and what counts as a resolution are in the help center.

Free AI options, and where they stop

A week-long trial on your own articles

The only test worth running uses your content and your real questions.

1. 1Pull 30 real questionsTake the last month of conversations. Mark 20 your articles answer and 10 they don't.
2. 2Point each tool at the same articlesSame content, same day, so the comparison means something.
3. 3Run all 30 through each toolNote which answers cite a source, and which ones are right.
4. 4Check the 10 you don't coverA good tool hands off. A weaker one improvises. This is the test that matters most.
5. 5Read the billLook at what each tool counted as billable, including hand-offs and silent customers.

When the uncovered ten go wrong, the fix is an article, not a setting.

Then read what each tool counted as billable. In Talvo, the analytics overview shows AI resolutions next to CSAT split between AI-resolved and human-resolved conversations.

Fin AI alternatives has a longer version of this test, and knowledge base chatbot explains why the ten uncovered questions are the ones that matter.

The risk side, in one paragraph

A tribunal in British Columbia held Air Canada responsible for what its website chatbot said, rejecting the idea that the bot was a separate entity. The ruling doesn't say what kind of chatbot it was, and it isn't legal advice or a US precedent, but it is the clearest case of why your AI should answer from what you've published and hand off the rest. We summarize the decision and give a checklist.

Which to pick

Pick the tool whose meter you can live with and whose hand-off you've tested. If you're already on Zendesk or Intercom and the product works, the AI may not be the reason to move. If you want email-first with a cap, Help Scout is the careful choice. If you'd rather not run a meter at all, look at Olark. If you want the inbox, help center and AI in one workspace at a lower per-resolution rate than the big suites, with the gaps above, try Talvo's free plan against your own articles.