What an AI chatbot for business costs, how WhatsApp message fees work in 2026, and what to check before buying: handover, PDPA and wrong-answer risk.
Quick answer: An AI chatbot for business is worth buying when it answers from your own content, hands over to a person cleanly and costs less per resolved conversation than your current support. Expect a platform fee, usage charges, set-up work and upkeep. Test it on your real questions first, because a wrong answer is legally yours.
A modern chatbot reads a question in ordinary language and writes an answer from material you give it, such as your help pages, price lists and policies. A rule-based bot only follows a fixed menu or keyword script. That difference is why the one word "chatbot" now covers products with very different costs and risks.
In practice an AI chatbot for business can answer the long tail of questions nobody scripted, work in several languages, look up an order or book a slot through a connected system, and collect lead details before passing them on. Klarna's press release from February 2024 reported that its assistant had held 2.3 million conversations in its first month, two-thirds of its customer service chats, with errands resolved in under two minutes against eleven before. These are company-reported figures from a very large fintech: a ceiling, not a forecast for a ten-person firm.
What it cannot do is know things you have not told it. It does not check a fact before stating it, it cannot approve a refund unless you wire in that permission, and it will sometimes sound sure while being wrong. Vectara's public hallucination leaderboard, updated September 2026, measures how often models add unsupported content when summarising a short document. Rates run from about 1.8% to 24.2%, with most models between roughly 3% and 15%. That test covers summarisation only, so it is a rough guide, but it shows the error rate is never zero even on an easy task.
The honest answer to "chatbot vs live chat" is that they do different jobs. Live chat puts a person in a window on your site, so it is accurate but limited by staffing hours. A bot is always on but only as good as its content and handover rules. Most small firms end up with a bot in front and people behind it.
| Rule-based bot | AI chatbot | Live chat (people) | Human team (phone, email) | |
|---|---|---|---|---|
| Best for | Fixed menus: opening hours, track my order | Open questions answered from your content | Complex or emotional enquiries on your site | Disputes, negotiation, high-value sales |
| Handles unscripted questions | No | Yes, with a risk of wrong answers | Yes | Yes |
| Available out of hours | Yes | Yes | Only if staffed | Only if staffed |
| Main failure mode | Dead ends and menu loops | Confident wrong answers | Slow replies, queues | Cost and capacity |
| Cost pattern | Low, mostly build | Platform fee plus usage plus upkeep | Per-seat plus salaries | Salaries |
| Setup effort | Script every path | Prepare and test your content | Hire and train | Hire and train |
Decision matrix based on how each option behaves in practice; not a ranking, because the right mix depends on your enquiry types.
Customers are not uniformly keen. A Gartner survey of 5,728 customers run in December 2023 found that 64% would prefer companies did not use AI in customer service. It is a stated preference from one older survey, but it supports a practical rule: always offer a clear route to a person. Our AI chatbot and conversational agent development work is built around that rule.
Start where your customers already message you. For most firms in India and Singapore that is WhatsApp, with a website chat widget second. If you are weighing an AI chatbot for small business use, one channel done properly beats three done thinly.
A website bot is simple to launch, and the channel adds no per-message fee, because you pay your platform, not Meta. It also depends on a script running on your page, which brings its own risks. On one client site, a recurring vertical line appeared on every page for days and matched no CSS. The cause was a transparent third-party chatbot iframe expanding over the page whenever its teaser bubble appeared.
A WhatsApp chatbot for business reaches people in the app they check anyway, and it can send order updates and reminders as well as answer questions. It comes with rules and fees set by Meta, covered below. Our explainer on what a WhatsApp automation platform is and our comparison of WhatsApp CRM software for small business cover the tooling side.
AI chatbot cost has four parts: the platform, the usage charges, the build and integration work, and upkeep. Vendors price the first two in different ways, so compare them on your own expected volume rather than on the headline number.
| Pricing model | Published structure (pages checked October 2026) | What to watch |
|---|---|---|
| Per outcome | Intercom lists its AI agent from $0.99 per outcome, charged once per conversation, with no seat, set-up or platform fees on the standalone option. | An outcome includes a customer who simply stops asking after the answer, and a completed workflow including handoffs. Your bill tracks behaviour, not proven resolution. |
| Per seat plus per resolution | Zendesk Suite plans run $55 to $115 per agent a month billed annually ($69 to $149 monthly), with 5 or 10 automated resolutions included per agent. Extra resolutions are $2.00 pay-as-you-go. | The page showed committed resolutions in pounds and the rest in dollars, so confirm the billing currency. You pay for seats you may already have. |
| Per conversation tier | Tidio has a free plan with 50 conversations, a Starter plan at $24.17 a month for 100 billable conversations, and its AI agent from $32.50 a month for 50 AI conversations. | Allowances are small. Check what happens when a busy month passes the cap. |
Vendors are named only as the source of the published figures, not as recommendations. Prices change often; confirm on the vendor page before budgeting.
If the bot lives on WhatsApp, Meta charges separately from your software vendor. Since 1 July 2025 Meta charges per message rather than per conversation, and only delivered template messages are billed. Replies inside the 24-hour customer service window that opens when a customer writes to you are free, and utility templates sent in that window are free too. The same rules are summarised on WhatsApp Business's pricing page.
| Message type | How Meta charges (official pricing documentation) |
|---|---|
| Reply to a customer inside the 24-hour service window | Free (non-template messages) |
| Utility template inside an open service window | Free |
| Any message within 72 hours of a click-to-WhatsApp ad or Facebook Page button entry | Free during the entry point window |
| Marketing, utility or authentication template outside a window | Charged per delivered message, by category and country |
Source: Meta's pricing documentation for the WhatsApp Business Platform. Per-message rates are set per market in Meta's rate cards.
Three dated points matter if you sell in Singapore or India. Meta's update notes record that India billing moved to rupees from 1 January 2026, and eligible India customers must migrate to INR by 31 December 2026 or Meta stops delivering messages from non-INR accounts from 1 January 2027. Singapore dollars became a billing currency on 1 April 2026, and Singapore utility and authentication rates rose on 1 July 2026. Meta may change prices only on the first day of a quarter, so you can plan around those dates. The rate figures sit in Meta's rate card tool, and we have deliberately not quoted them here because they change.
For clients invoiced in rupees we add India's 18% GST on Meta's message fees as a separate line, and keep our base rates identical to Meta's rate card so they can compare line by line.
The platform fee is rarely the biggest number. Connecting the bot to your order system, CRM or booking calendar, preparing and cleaning the content it answers from, writing handover rules and testing all take real time. After launch someone must review conversations, fix gaps and keep the content current. Our AI CRM and AI e-commerce automation services cover the integration side, and application maintenance and support covers upkeep. We do not publish prices because scope drives them; we confirm both scope and price in writing after a discovery call.
Meta bars AI companies from using WhatsApp as a distribution channel for general assistants. Section 4.7 of Meta's terms for the WhatsApp Business Platform says AI providers are "strictly prohibited from accessing or using the WhatsApp Business Platform" where such technologies are the "primary (rather than incidental or ancillary) functionality". Meta decides what counts as primary. The terms were last modified on 23 September 2026.
Our reading is that a shop's order-status or booking bot is ancillary to the shop's own business and is not what the clause targets. Meta decides case by case, so keep the bot narrowly about your own products and services, and ask your provider to confirm.
These failures come from our own WhatsApp CRM and chatbot tooling for e-commerce clients, in production rather than in a demo.
We also build chatbots as configuration rather than one-off builds: one multi-tenant platform hosts five specialised agents, so a new client's bot is set up rather than rebuilt. The published AI chatbot and support automation case study reports 70% fewer support tickets.
From our work: the AI command box in our WhatsApp CRM can search and tag contacts, create templates and draft campaigns, but it can never message a customer on its own, because no send-capable tool exists in its tool set. Campaigns it creates are left as drafts for a person to send. The limit is structural, so no prompt or injected instruction can get round it.
Check four things, and ask the vendor to demonstrate each on your own data rather than describe it.
Ask where each answer comes from. A good setup answers only from your approved pages and documents, can show which one, and says "I don't know" when nothing matches. The legal stakes are real. In Moffatt v Air Canada (decided 14 February 2024), a British Columbia tribunal held the airline responsible for a chatbot that wrongly described a bereavement fare, awarding about $650. The tribunal treated the chatbot as part of the airline's website and rejected the argument that the correct policy appeared elsewhere on the site. It is a small tribunal decision, not binding elsewhere.
The US Consumer Financial Protection Bureau's 2023 review of bank chatbots warns of "doom loops", where customers repeat themselves with no route to a person, and says chatbots "must comply with all applicable federal consumer financial laws". The lesson applies beyond banking: test that a customer can reach a person in one step, with the conversation passed across so they do not start again.
A chatbot collects names, phone numbers and sometimes sensitive details, so data protection law applies from day one. Singapore's PDPC published Advisory Guidelines on the use of personal data in generative AI on 20 July 2026, and your vendor should be able to explain how it meets them, including where data is stored and whether your conversations train its models. In India, the DPDP Rules were notified in November 2025 with an 18-month phased timeline, so consent notices need planning now. Our guide to data privacy in the AI era and security and compliance consulting go deeper. In the United States, a 2023 law-firm summary of FTC guidance notes that steering people against their interests through chatbots is likely unfair or deceptive.
You need to see unanswered questions, handovers, resolution and cost per conversation. In our reporting work a data connector silently ignored account filters, and averaging rates row by row gave one client a 97.8% engagement rate where the session-weighted figure was 72.3%. Ask how the vendor defines a resolution, since the three pricing models above each define it differently. Our analytics and reporting dashboards work covers this.
For a wider view of building with agents, see our practical guide on how to build an AI agent for your business, and for the tasks around the chat itself, business process automation with workflow tools, RPA or AI agents.
For a small or mid-sized business, start with a narrow AI chatbot on the one channel your customers already use, give it only the content you are willing to be held to, and keep a person one tap away. The Air Canada case shows that you own what the bot says, the CFPB review shows what happens when people cannot escape it, and our own WhatsApp work shows that most failures are plumbing and message rules, not clever AI. Buy on cost per resolved conversation at your own volume, and insist on a month of reviewed conversations before widening scope. If you want to talk it through, contact us; we have teams serving Singapore and India, and you can see the WhatsApp AI agent suite we built and run.
It is worth it when you get repeated questions on a few topics, such as delivery, hours, prices and bookings, and your content is written down. It is not worth it if most enquiries are complex or emotional. Start with one channel and three enquiry types, measure cost per resolved conversation, and widen only if the numbers hold.
Live chat connects a visitor to a person, so answers are accurate but limited by staffing hours. A chatbot answers instantly at any hour but can be wrong or hit a dead end. Most small businesses use both: the bot handles routine questions and passes anything complex or urgent to a person with the conversation attached.
Yes. Meta's terms restrict AI providers whose main product is a general assistant, not businesses serving their own customers, which is our reading of section 4.7. Keep the bot focused on your products, use an approved provider, and follow Meta's template and opt-in rules. Ask your provider to confirm the setup against the current terms.
Since 1 July 2025 Meta charges per delivered template message, by category (marketing, utility, authentication) and country. Replies within the 24-hour customer service window are free, as are utility templates inside it. The actual rates are in Meta's rate card for your market and currency, and they can change each quarter.
Yes. Public tests such as Vectara's hallucination leaderboard show every model adding unsupported content at least some of the time. You reduce the risk by limiting scope, answering only from approved content, showing sources, testing on real questions and giving customers an easy route to a person.
In the Moffatt v Air Canada decision (2024), a Canadian tribunal held the airline responsible for its chatbot's misleading answer and rejected the idea that the bot was a separate entity. It is not binding everywhere, but treat your chatbot's statements as your own, especially on prices, refunds and policies.
Yes, if it collects or uses personal data such as names and phone numbers. Singapore's PDPC published guidance on personal data in generative AI in July 2026. Tell people what data you collect and why, get consent where required, and check where your vendor stores data and whether it trains models on your chats.
It depends on how many systems it must connect to and how ready your content is, so we confirm scope and timing in writing after a discovery call.