Why most chatbots fail — and where Anserve stands
Dutch and European consumer research, industry benchmarks and health-service pilots all describe the same problem this year: most chatbots leave people without an answer, and without an easy way to reach a person instead. We read it, checked it against how Anserve actually works, and are publishing the honest version — including the parts we haven't fixed yet.
What the research says
A consumer-rights survey, two national benchmark studies, an AI-accuracy audit and independent health-service coverage — different publishers, same pattern.
| Consumentenbond Survey, ~10,000 consumers |
71% get no answer from a chatbot; 90% say it makes reaching a human harder. Of 100 companies audited, 23 hid their phone number and 35 blocked saving the chat transcript. |
| Nationale Voice Monitor 2026 National benchmark study |
Only 12% of chatbot questions answered correctly. Frustration up from 38% to 42% year over year; 40% can't get transferred to a person. |
| MT Sprout Business press |
0% of consumers name chatbot as their preferred contact method; 90% prefer the phone. Bots fail hardest on urgency, emotion and repeat issues. |
| AI brand-accuracy audit 1,917 answers, 37 Dutch brands |
Testing ChatGPT, Gemini and Perplexity: 62% of wrong answers were outdated information, not invention — AI mostly repeats yesterday's price. |
| HLN Consumer advice |
A mainstream explainer teaching readers to catch AI hallucination — a sign this scepticism has reached general news, not just tech press. |
| Totaal TV Telecom trade press |
Dutch telecom providers leaning harder on chatbots while also moving phone support to cheaper overseas call centres — automation and offshoring as parallel cost cuts. |
| Tweakers & Bright.nl NHS AI triage pilot |
An AI chatbot preparing patients before a GP call cut phone hold times by 29% in a Sussex pilot — explicitly built to supplement staff, not replace them. |
Problem by problem, the honest version
Some of this, Anserve was built around. Some of it isn't a chatbot problem at all — and we'd rather say so than let the claim ride.
"The bot repeats the same standard answer once a question leaves the script"
Retrieval-Augmented Generation retrieves the relevant passage from an organisation's own documentation instead of matching against a fixed script. That's a real difference from scripted bots — but it isn't a guarantee. An agent is only as good as the knowledge connected to it, and it will still meet questions it can't answer. The honest claim isn't "we always know it" — it's "when we don't, we hand off" (next).
"Chatbots make it harder to reach an actual person"
Below a confidence threshold, or the moment someone asks for a human, the full conversation escalates — with context — to the Anserve Console, on by default with no setup, or to Slack. A person replies, and the visitor sees it in the same chat window. Escalation isn't a fallback bolted on afterwards; it's the second half of how the product is built.
"The chatbot sounds confident and just makes things up"
Grounding answers in retrieved passages, instead of generating freely, is exactly what cuts this down. "Reduces" is the honest word, not "eliminates" — no AI system can promise zero hallucination. The escalation path above is the answer to the residual risk, not a bigger accuracy number.
"You can't tell you're talking to a bot until it fails you"
Every Anserve chat can show a visible AI icon, so visitors know from the first message they're talking to an assistant, not a person.
"You can't save or take the conversation with you"
A "Download transcript" button is available when a chat ends, so visitors can keep a copy without asking anyone for it.
"AI repeats outdated prices and specs about a brand"
This one is about general-purpose AI — ChatGPT, Gemini, Perplexity — answering from whatever's indexed online, including stale product pages and affiliate content. Anserve's own agent answers from an organisation's live, connected knowledge sources rather than scraped web pages, so it doesn't have this problem internally. But no vendor, us included, controls what a general AI engine says about a business elsewhere on the web.
"Providers automate support and move phone lines to a cheaper call centre, at the same time"
That's a sourcing decision, not something any chatbot vendor changes. What it does confirm is the pattern worth naming: customers can tell when "automation" means cheaper, versus when it means better.
Four questions worth asking any AI answer platform
- Does it ground answers in our documentation, or generate from general knowledge?
- What happens automatically when it isn't confident — does a person get looped in, or does the visitor hit a wall?
- Can a visitor tell, from the first message, that they're talking to an AI?
- Can they get a copy of the conversation afterwards?
Sources reviewed August 2026. We'll update this page as new research appears.