Guide
How to Choose an AI Support Chatbot for a Small Business
A price-aware look at helpdesk bots, website widgets, and FAQ pages, plus a ticket-and-time test for when 24/7 AI support is worth paying for.
An AI support chatbot is worth paying for when the same questions arrive after hours and its mistakes are unlikely to trigger expensive disputes or lost clients. For a freelancer or a shop of a few people, 24/7 support usually means a widget that answers from your FAQ. It can safely repeat approved, version-controlled public policies, such as published shipping times or standard refund terms, while a person handles exceptions, disputes, and account-specific decisions. Before you buy, count how many repeat questions you already answer, whether a human must take over, and where the chat should live: your site, email, or WhatsApp. Confirm current pricing on each vendor’s site; per-seat, per-resolution, and AI-message add-ons move often.
The four real options
Helpdesk platforms with AI (Zendesk, Freshdesk, Intercom, Help Scout). These sit on top of a shared inbox. The bot drafts replies, suggests macros, or answers from a help center you maintain. Intercom leans toward a website messenger and lead capture. Zendesk and Freshdesk are classic ticket queues, but their AI packaging differs: Zendesk includes AI Agent access with Suite and Support plans, measures usage by automated resolutions, and sells Copilot as an additional product. Help Scout is lighter and often a better fit for a small team that still wants a human tone. On the pricing page, check the included AI usage, overage unit, and any separate assistant add-on — those details can change the bill more than the headline tier does. The strength is handoff: a bot can greet, then a person sees the thread. The trade-off is cost and setup. You pay for seats and you still write the articles the bot reads. As a conservative screening example, a full helpdesk may be worse value than a good FAQ and a shared inbox when you get only ten emails a week. Ten is not a universal cutoff; set yours from your ticket mix, tested no-edit rate, risk, and cost of a wrong answer.
Website customer-service widgets (Tidio, Tawk.to, Crisp). These embed on your site, answer FAQs, and capture contact details. Tidio and Crisp mix live chat with a bot. Tawk.to has long offered a free live-chat tier, though what stays free changes — check before you build on it. Plans are typically a monthly widget fee plus a cap on bot replies or stored contacts; find that cap before you launch, because it is what you actually outgrow. The strength is after-hours coverage on the page where the question happens. The trade-off is quality. A widget trained on a thin FAQ can invent policies you did not write. Give customers an obvious “talk to a person” path, show business hours, and route account-specific billing questions, exceptions, and complaints to a person.
Social messaging automation (ManyChat). ManyChat is built mainly for conversations on Instagram, Messenger, WhatsApp, Telegram, and related messaging channels rather than as a website support widget. It can answer common questions and collect contact details inside those channels. Choose it when customers already message you there, then check the channel rules, contact limits, and human-handoff path before paying.
A public FAQ plus email, with a chatbot only as a drafter. You publish ten honest answers — hours, shipping, refunds, how to book — and you answer the rest by email. ChatGPT or Claude can draft a reply from your FAQ paste, but it does not face the customer until you send it. The marginal cost is close to zero if you already pay for a chatbot. The trade-off is speed: no one gets an instant answer at 11 p.m. For many freelancers that is the correct trade. A bot that is wrong at 11 p.m. is worse than silence until morning.
Who should pick which
- A shop or studio with the same ten questions every week and a site that already gets traffic: a widget like Tidio or Crisp is worth testing, trained only on answers you would publish, with human handoff for account-specific money questions, exceptions, or complaints.
- A small team already drowning in a shared inbox: a helpdesk with AI drafts can pay off, but buy it for tickets and assignment, not for the novelty of a bot. Help Scout or Freshdesk is usually enough; Intercom is easier to overbuy.
- A freelancer whose “support” is a handful of client emails: stay on email and a FAQ. Do not put an unsupervised bot on a client portal.
- Not for you yet: if you cannot list the ten questions you actually get, you do not have a bot problem. You have a documentation problem. Write the FAQ first.
Let the bot repeat only approved, version-controlled public refund and pricing policies. Send exceptions, disputes, account-specific decisions, and legal or medical advice to a person. Log transcripts for a month and read the misses. A cheap bot that trains customers to ignore you is a brand cost, not a saving.
Before launch, complete a privacy and security check. Minimize the personal information sent to the bot, mask sensitive details such as passwords, payment data, and health information, and confirm whether the vendor uses chats for model training and whether you can opt out. Set retention and deletion periods, review the data processing agreement (DPA) and subprocessors, limit transcript and admin access by role, and update your privacy notice so customers know how the bot handles their data.
A test for whether it’s worth paying
For two normal weeks, tag every inbound question: repeat FAQ, account-specific, or complaint. Time how long each repeat FAQ takes to answer, including finding the policy. Then run the same FAQs through the candidate bot and sort the replies into three piles: correct as sent, incomplete but harmless, and wrong. Only the first pile counts as time saved.
Put a number on it: monthly repeat FAQs × no-edit rate × minutes saved per no-edit answer ÷ 60 × your hourly rate, minus the cost of repairing wrong answers, minus the plan fee. For a conservative example, estimate each repair at 1.5× your normal answer time to allow for correcting the reply and rebuilding trust. Replace that multiplier with your measured repair time and raise it when mistakes carry greater financial, privacy, or safety risk.
Work an example at $50 an hour. You get 80 repeat questions a month, and each normally takes 4 minutes. The bot returns 50 correct as sent, for a no-edit rate of 50 ÷ 80, or 62.5%. The formula gives 80 × 62.5% × 4 minutes ÷ 60 × $50, or about $167 in gross time value. Of the remaining 30, say 25 hand off cleanly — you answer those in the 4 minutes they normally take, so they add no repair cost — and 5 go out wrong. Using the conservative 1.5× assumption above, those five take 6 minutes each to repair, or 30 minutes worth about $25. Net time value is roughly $142 a month before the fee. That $142, not the $167, is the number to hold a quote against.
The weak case: 12 repeat questions a month, 3 minutes each. Even if the bot handled all 12 with no edit — a deliberately generous hypothetical ceiling — that is 36 minutes, or $30 of time value, and it still has to cover both the plan fee and the hours you spend writing the articles the bot reads.
As a conservative example gate, pay only when at least two-thirds of replies need no edit for two consecutive months and measured net time value is at least twice the fee. These are not universal rules. Set the minimum no-edit rate, test period, and value-to-fee multiple from your own error costs, risk level, and observed results. With the 80-question sample above, two-thirds means at least 54 clean answers out of 80; at $142 in net monthly time value, the twice-the-fee assumption limits the plan to about $71 a month. Put the renewal date in your calendar, run another 80-question sample the week before it, and cancel if the result falls below the thresholds you set. Below your bar, publish the FAQ and draft replies in a chatbot you already pay for. Support bots sell after-hours answers to questions you already know.