AI AutomationWeb Design

AI Chatbots on Small Business Websites: What They Cost and When They Pay

Kage Works13 August 202610 min read

Gartner surveyed 3,566 B2B and B2C customers in February and March 2026 and found them around three times more likely to reach for a third-party GenAI tool than a company's own chatbot when sorting out a service issue. Use of third-party tools nearly doubled in a year. Use of company-provided chatbots has stayed statistically unchanged since 2022 (Gartner).

Your customer with a question about your business has started typing it into ChatGPT rather than the bubble in your bottom-right corner. That doesn't close the question, since a chatbot on a plumber's site does a different job from a bank's deflection tool, and the small version can pay for itself on after-hours enquiries alone. It does set the order of the decision: price the volume first, then look at the tool.

Price your volume before you price the tool

Vendor pricing tells you the volume these products assume. Intercom's Fin AI Agent bills US$0.99 per resolution on a base plan of US$49 a month that includes 50 of them, counting a resolution when the customer confirms the answer helped or leaves without asking for more, charging at most once per conversation, and billing nothing for an escalation to a human (Fin). Tidio's Lyro sells packages from 50 conversations a month up to 1,000, with the entry tier around US$32.50 (Tidio).

Fifty conversations a month is under two a day. That's the floor these products are built around, and it's the number to hold your own enquiry log against before you read a feature list. Take a cabinetmaker fielding six enquiries a week as the illustration: buying the entry tier gets capacity for 50 and uses about half, and the enquiries that matter most are the ones asking for a quote, which is the answer a bot must not give.

The maths runs the other way on a busy site. A physio clinic taking forty calls a week about opening hours, parking and ACC surcharges has a volume problem worth spending on, and a bot handles the same dozen questions the receptionist fields between patients.

Air New Zealand spent a year getting Oscar to 75%

Air New Zealand launched its chatbot Oscar in February 2017. At his first birthday the airline reported conversations across more than 380 topics, close to 75% of questions answered correctly against 7% on day one, and roughly 1,000 conversations a day (Air New Zealand via Scoop). Researchers at Unitec later studied the deployment and concluded the service suited routine problems and standardised FAQ answers (Unitec ePress).

Oscar predates large language models, so treat 7% as an artefact of 2017 tooling rather than a starting point you'd hit now. The other two numbers transfer intact. Those 380 topics were a body of written answers somebody maintained, and a thousand conversations a day gave the airline the feedback that showed which of its answers were wrong.

A small site supplies neither input. Your bot answers from whatever you have published, so a six-page site with a services list and a contact form gives it a services list and a contact form, and it fills the gaps with something that sounds right. Writing the dozen answers properly is the work itself, and once they exist as pages you collect part of the benefit before any widget loads.

Labelling the bot costs you what the vendor is selling

Xueming Luo and colleagues ran a field experiment with a financial services company in China, randomising more than 6,200 customers to receive structured outbound sales calls from either a chatbot or a human worker. Undisclosed, the bot matched proficient staff and beat inexperienced staff by around four times on purchases. Disclosing the bot's identity before the conversation cut purchase rates by over 79.7%. Disclosing later in the call, and prior customer experience with AI, both softened the drop (Marketing Science, 2019).

That was a phone call, in 2019, in another market, selling rather than supporting, so leave the multiplier where you found it and keep the direction. Gartner's December 2023 survey of 5,728 customers found 64% would prefer companies didn't use AI in customer service, and 53% would consider switching to a competitor on learning that a company planned to (Gartner).

New Zealand sits at the sharp end of that. The University of Melbourne and KPMG surveyed 48,340 people across 47 countries between November 2024 and mid-January 2025, with nationally representative samples of at least 1,000 per country. 44% of New Zealanders said the benefits of AI outweigh the risks, the lowest ranking of any country in the study (KPMG New Zealand).

The Privacy Commissioner closes off the undisclosed version anyway. Transparency about what you collect and how you use it sits in IPP3, and the OPC's guidance calls out chatbots as a case needing particular care where you want to use the conversation to train or refine an AI tool (Office of the Privacy Commissioner). Chat transcripts carry names and job addresses, making them personal information under the Privacy Act 2020 and making the vendor's data handling your problem to check.

Run the thing labelled, and budget for the labelled version's conversion rate rather than the one in the case study.

Every visitor pays the load, a fraction opens the chat

Intercom's engineering team published a write-up of shrinking the Messenger bundle, describing close to 600KB of gzipped JavaScript at one point and a rebuild that brought the boot payload to 280KB (Intercom). Lighthouse fails its third-party audit when third-party code blocks the main thread for more than 250ms during load (Chrome for Developers).

Search for the share of visitors who open a chat widget and you'll find 5% to 15% repeated across vendor blogs, none of them attaching a methodology or a sample. Treat the range as folklore. The asymmetry underneath holds whatever the true figure is: the download lands on every visitor loading your site on a phone in a Northland driveway, and the benefit reaches whoever opens the panel.

Fixing that takes an afternoon. Render a static button in your own markup and load the vendor script on first click. Our guide to why your website is slow covers the same pattern for the other scripts a marketing site collects.

Klarna ran the largest version of this and walked part of it back

Klarna announced in February 2024 that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of the company's customer service chats, doing the equivalent work of 700 full-time agents. Average resolution time fell from 11 minutes to under 2, repeat enquiries dropped 25%, and Klarna estimated a US$40 million profit improvement for 2024 (Klarna).

In May 2025 chief executive Sebastian Siemiatkowski told Bloomberg the company had pushed too far: "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality." He added that "really investing in the quality of the human support is the way of the future for us," and Klarna started recruiting agents again (CX Dive).

Both sets of numbers are true, and they describe different halves of the same queue. High-volume repeat questions went well. Disputes and hardship cases went badly enough to reverse a public strategy. Gartner's August 2026 release from that same 3,566-customer survey puts the customer-side version plainly: 87% say an option to reach a human is essential when a company uses GenAI, while 50% say GenAI makes their interactions easier (Gartner).

Returns are thinner than the pitch too. In a separate Gartner survey of 1,303 senior leaders, service and support functions put a median 12% of their 2025 budget into AI, the highest share of the ten functions assessed, and 24% of those leaders showed positive financial returns across their AI use cases.

Klarna had the friendliest queue a bot could ask for: millions of near-identical questions about payments, and engineers on staff to tune the answers. Your queue is smaller and more varied, which makes the tuning harder and the payoff smaller.

A bot earns its place in a narrow slot

The jobs that survive contact with a five-person business look dull on a demo:

After-hours capture. Somebody lands on your site at 9pm with a leaking cylinder. Taking a mobile number and two sentences about the job beats a contact form, and beats a bot that tries to diagnose the cylinder.
The questions you answer twenty times a week. Opening hours, service area, what you don't do, parking, whether you take EFTPOS on site, how long a callback takes.
Routing. Sorting an enquiry by suburb or job type and putting it in front of the right person.
Handing over. A visible path to a human from the first message, since the Gartner number says that path is what makes the rest acceptable.

Two rules cover the other side. It doesn't quote a price or promise a lead time, because a bot inventing either creates an obligation you have to honour or argue about under the Fair Trading Act. It answers nothing that isn't already written down somewhere you control. Our piece on AI workflow automation for NZ small business covers where that liability lands, including the Air Canada tribunal decision.

Four weeks of counting beats any demo

Do this before you sign up for anything.

Log every enquiry for two weeks: the time of day, and the question asked in the customer's own words.
Count how many arrived outside working hours, and how many repeated something your site already answers.
Write proper answers to the ten most common questions as pages on your site, in the customer's vocabulary. Your bot needs them, and so does the ChatGPT session Gartner found your customer running instead.
Re-count for two weeks. Enquiries that stop arriving are the ones your pages solved for free.

If what's left runs under 50 conversations a month, put the money into a better FAQ page and an enquiry form that texts your phone. Above that, weighted towards after-hours, a labelled bot with a human handoff has a case, and you hold the log to check it against a month later.

At Kage Works we build the capture path into the site itself, so an after-hours enquiry reaches your phone without a fourth subscription in the stack. A chatbot on top of that is a volume question, and two weeks of your own counting answers it better than any vendor's benchmark.