How an AI support chat affects conversion and the load on operators

An AI support chat changes two things most: how fast the first answer comes, and how many repeated questions still reach your operators. Published research shows a clear gain in support productivity and a large share of questions that arrive after hours; independent proof that the chat itself raises conversion is scarce, so measure the effect on sales on your own site, for example with Tolkly's numbers, rather than borrowing it from other companies' case studies.

Below: what is known from research with a stated method, where vendors' figures need care, and how to check the effect yourself. Industry data and claims about Tolkly are kept apart: about Tolkly we say only what the product does, not what results it promises.

What is known about AI's effect on support work?

The most reliable study on the subject is «Generative AI at Work» by Brynjolfsson, Li and Raymond. The authors followed the staged rollout of an AI assistant to 5,172 support agents at one company.

Issues resolved per hour rose by 15% on average (the published version in The Quarterly Journal of Economics, 2025; 14% in the 2023 NBER working paper).
Novice and low-skilled agents gained much more: 34% in the 2023 NBER version.
Customers became more polite and asked for a manager less often.

An important caveat: in this study the AI suggested answers to agents; it did not answer customers itself. It is evidence of how AI speeds people up, not of how many questions a bot closes without them.

How many questions can a bot close without an operator?

Here almost every figure comes from vendors, and each counts differently.

Klarna, February 2024. The company reported that its AI assistant handled 2.3 million conversations in its first month, two thirds of its support chats; resolution time fell from 11 minutes to under 2, and repeat inquiries fell by 25%. These are the company's own figures, not independently checked.
Klarna, May 2025. The CEO told Bloomberg that the focus on cost had lowered quality, and the company started hiring people for support again. The lesson is not that AI does not work, but that a customer needs a clear way to a person.
Intercom Fin. fin.ai states an average resolution rate of 76% across more than 12,000 customers (as of November 2026). But Intercom also counts as resolved a conversation after which the customer simply left without asking for further help («assumed resolution»). A customer who left may not have got an answer, so such a figure overstates the share of questions actually solved.

When you compare vendors' figures, first ask how they define «resolved». In Tolkly, only what the visitor confirmed counts: the «It helped» button under an AI answer, or «Yes» to «Did we solve your question?» after an operator closes the conversation. Silence does not count.

How many questions arrive after hours?

More than it seems. Two recent publications with a stated method:

Gorgias Ecom Lab, July 2026. For the median online store on Gorgias, 53% of pre-purchase questions arrive outside business hours (Mon–Fri, 9:00–18:00 in the store's local time). The median first human reply is about 2.2 hours within business hours and about 15 hours outside them. The authors say plainly that the link with revenue is associative, not causal.
Zendesk CX Trends 2026 (a survey of 6,182 consumers in 22 countries, June 2025). 74% of consumers now expect support to be available around the clock.

If half of pre-purchase questions come at night and on weekends, and the answer comes 15 hours later, the buyer has time to leave. An AI chat closes this gap for questions whose answer is already on the site: delivery, payment, returns, timing, terms.

Does an AI chat help conversion?

The honest answer: we found no independent studies that prove a causal rise in conversion from an AI support chat as such. There is indirect evidence:

Pre-purchase questions often arrive after hours (Gorgias, 2026).
AI speeds people up and makes support more even (Brynjolfsson et al., 2025).
Customers are wary of AI: in a Gartner survey (5,728 people, December 2023, published July 2024), 64% would prefer that companies did not use AI in customer service, and 53% would consider switching to a competitor. The main fear is that it will be harder to reach a person.
Disclosing that a bot is talking can lower sales: in a field experiment by Luo et al. (Marketing Science, 2019, over 6,200 customers, outbound sales calls), disclosure before the conversation cut purchases by more than 79.7%.

Point 4 is no reason to hide the AI. In Tolkly every conversation starts with a mark that an AI answers, and that will not change: model usage policies require it, and it is simply fair to the visitor. What points 3 and 4 do show is that an AI chat's value for conversion depends on how far its answers can be trusted and how fast the visitor reaches a person when the bot cannot help.

How does an AI chat lower the load on operators without losing customers?

A working setup looks like this:

The bot answers only from your pages and rules. In Tolkly every sentence of an answer links to its source and quotes it word for word; code checks that the quote really is on that page. An unconfirmed sentence is not shown.
No answer means an honest refusal. The bot writes «I did not find this in the help pages and rules of the site» and offers «Hand over to an operator». A refusal does not count against the limit.
A question about the visitor's own order goes to a person. If the visitor asks about their order, payment or account, the bot answers the general part from your rules and hands the conversation to an operator with the order number.
A request for a person is met at once. The bot recognizes it by meaning, not by keywords, and hands the conversation over without a word.
Operators answer where it suits them: in a topic of the company's Telegram group or in the web inbox.
At night, with no operators, the chat asks the visitor to leave a contact: Telegram, email or phone.

This way operators get fewer repeated questions, and the visitor is not stuck with a bot that cannot help.

PaywireBy the site's rules

Hello! I am the AI assistant of Paywire: I answer from the service’s rules or bring in an operator.

Sent it an hour ago, payment P-48213 has still not arrived

I cannot see your payment: an operator will check it. Bringing one in.

The operator has order P-48213

The conversation went to an operator

Operator Anna joined the chat

I see P-48213: the recipient’s bank is checking it, up to 24 hours. I will write here when it arrives.

Runs on Tolkly

A payment question goes to an operator at once, the number in the conversation Example: the service is invented.

How do you measure an AI chat's effect on your own site?

Other companies' percentages say little about your customers. Count it yourself over 30 days before and after launch:

Metric	How to count	What it shows
Time to first answer	the median over all conversations, day and night apart	whether the after-hours gap is closed
Share of conversations handed to people	handoffs / all conversations	how much load is taken off operators
Confirmed resolutions	visitors' «It helped» and «Yes» / all AI answers	real use, without «silence = resolved»
Handoff reasons	a request for a person, an order question, your topic, a refusal	which pages to write
Conversion of visitors who used the chat	orders from those who wrote in the chat / all who wrote	the effect on sales; compare with the same period last year

The bot's refusals are a useful signal: each one shows a question your site does not answer. Write the page, and the bot starts answering with a quote. In Tolkly, operators' answers also become knowledge drafts: the owner approves them, and the bot then answers such a question itself, citing «Support answer».

Frequently asked questions

How many requests can an AI chat close?

It depends on how many answers are already on your site and on how «closed» is counted. Vendors publish figures from half to three quarters of conversations, but often count as resolved a conversation after which the customer simply left. It is more reliable to check on your own questions: in Tolkly only what the visitor confirmed counts as resolved.

Does an AI chat raise conversion?

Independent studies of a causal effect are scarce. It is known that about half of pre-purchase questions arrive outside business hours (Gorgias, 2026), and a human answer then takes hours. Measure the conversion of those who wrote in the chat before and after launch.

Will customers be put off because an AI answers?

Some customers are wary of AI: 64% in Gartner's 2024 survey would prefer support without AI, and the main fear is not reaching a person. So the bot must refuse honestly where it does not know the answer and hand the conversation to an operator quickly. Tolkly does both.

Do you still need operators with an AI chat?

Yes. Questions about a specific order, disputes and anything your rules do not cover need a person. AI takes the repeated questions whose answers are already on the site.

How fast can you try an AI chat on your own site?

With Tolkly, in a minute and without signing up: type your site's address in «Try it on your site» on tolkly.io, and the chat answers from your pages, in your colours.

Try it on your site

Open tolkly.io/#try and type your site's address: in a minute the chat answers from your pages, with a quote under every statement. No sign-up needed. The first month of work is free at tolkly.io: the widget on your domains, the inbox and operators in Telegram from day one.