AI Chatbot Response Time - We Timed 4 Support Bots (2026 Benchmark)

Every AI chatbot vendor says its bot answers "instantly". Very few publish a number, and we found none that was measured independently. So we measured it ourselves: the same five questions, typed into the support bot each vendor runs on its own website, timed from the moment the visitor presses Enter.

This page reports what we found, including where our own bot came out behind.

Key findings

• Typical first text took 4-6 seconds. Only Chatbase ever showed text in under 3 seconds (2 of its 15 answers).
• Chatbase showed the first words soonest (median 3.9 s) because it streams the answer word by word. It was also the least consistent: two of its fifteen answers took 9-11 seconds to start.
• Intercom Fin took the longest to start (median 5.8 s), then streamed the rest quickly. Its full answer was ready after a median 7.1 s.
• ELIA delivered the complete answer in a median 4.9 s, faster than Chatbase (6.4 s) and Fin (7.1 s) to a finished reply - but without streaming, so the visitor sees nothing until the whole answer arrives.
• LiveChat's bot started replying in a median 4.8 s, but never answered the question. Every reply asked for the visitor's email address first.
• Four of the seven vendors we looked at could not be tested anonymously on their own website. More on that below.

Results

Median over fifteen answers per bot (five questions, three sessions). Range in brackets.

Chatbase was fastest to first text and Fin slowest in all three sessions; LiveChat and ELIA were within a second of each other, and which of them came first varied by session. Per-session medians for first visible text: Chatbase 3.8 s, 4.8 s and 3.3 s; LiveChat 4.8 s, 5.5 s and 4.8 s; ELIA 5.6 s, 4.9 s and 4.7 s; Fin 5.8 s, 5.5 s and 5.9 s.

Two numbers matter, and they measure different things:

• First visible text is when the visitor sees the bot start answering. This is what decides whether the wait feels long.
• Complete answer is when the visitor can read the whole reply. It depends on the answer's length, so a bot that writes more will finish later.

A streaming bot can show text early and still finish late - Chatbase's slowest answer (621 characters, on pricing) took 14.6 s to complete. A non-streaming bot shows everything at once, so both numbers are the same.

What the vendors say

Only two vendors we looked at publish a speed figure, and both are vendor claims:

• Intercom says its Fin Apex model "delivers responses in 3.7 seconds" (VentureBeat, March 2026). In our test, Fin on intercom.com took a median 5.8 s to show its first text. The two numbers likely measure different things - a model benchmark versus the full path from the visitor's keyboard through the Messenger - so this is not a contradiction, but it shows the gap between a model figure and what a visitor experiences.
• Smartsupp says its AI can "answer in 1.5 seconds" (smartsupp.com). We could not test this: smartsupp.com does not run its own chat widget.

We found no independent latency measurements for any of these products.

Which bots we could not test

We tested only bots that a visitor can reach anonymously on the vendor's own website. We did not enter personal data anywhere.

Testing these on their customers' websites would measure the customer's setup, not the vendor's, so we left them out.

Does 5 seconds matter?

Usability research has long put the limit for keeping a user's attention at about 10 seconds, and the limit for an uninterrupted flow at about 1 second (Nielsen Norman Group). Every bot we tested sits between the two: fast enough not to lose the visitor, too slow to feel instant.

Faster is not automatically better, though. A 2026 study from NYU found that chatbot replies delivered after 2 seconds were rated less thoughtful and useful than the same kind of replies delivered after longer delays of 9 or 20 seconds (NYU Tandon, April 2026). For customer support, a correct answer in 6 seconds beats a wrong one in 2.

What clearly matters is the empty wait. Showing the first words early - streaming - makes the same total time feel shorter. That is where Chatbase leads this test and where ELIA currently does not.

Method

Timing was recorded inside the browser page with a script that watches the chat window for changes, with millisecond resolution. In run 3 a script drove the browser (desktop Chrome with a fresh profile) and used the same timing method. For LiveChat, whose chat runs in a cross-origin frame, timing in runs 1 and 2 started from the widget's own "message sent" event instead of the Enter key, which may slightly favour it, and saw each reply only as a whole. In run 3 the script watched inside that frame: it timed LiveChat from the Enter key and saw it reveal each reply word by word over about a second. Intercom's Messenger changes its layout one to three seconds after an answer finishes; that change does not touch the answer text, so it is not counted. Twice in run 2, Intercom's widget did not send a question as typed (once not at all, once merged with the previous question); we discarded those samples and asked again.

Limitations - please read before citing.

• Fifteen answers per bot from three sessions is a small sample. Treat differences under about one second as noise.
• Each vendor's own bot reflects that vendor's own configuration - model, knowledge base, prompt - not necessarily what a customer of theirs would get.
• Response time varies by time of day and server load. One location and one network were used.
• The bots gave answers of different length and quality. This test measures speed only; we did not score correctness.

What this means if you are choosing a chatbot

1. Ask for a measured number, not "instant". Try the vendor's own bot on its website and count. Anything in the 3-7 second range is normal today.
2. Look at first visible text, not just total time. A bot that streams feels faster at the same total time.
3. Check what the bot does before answering. A bot that answers quickly but asks for an email first has not saved the visitor any time.
4. Weigh speed against answer quality. A second or two rarely loses a customer; a wrong answer does.

How to cite this data

ELIA Assistant, "AI Chatbot Response Time - We Timed 4 Support Bots (2026 Benchmark)", 28 October 2026, https://elia-assistant.com/blog/ai-chatbot-response-time-benchmark-2026. Please mention the sample size and test dates when quoting individual figures. The raw timings for every answer are available as CSV.

We plan to repeat the test and add bots as vendors make them available. To see how ELIA answers your own customers, try ELIA Assistant free for 7 days.

Related reading: AI customer support by the numbers, Best AI chatbot for customer service, Best AI chatbot 2026.