5.85 million support conversations: where AI fits and when the customer needs a person

This week brought three studies on AI in customer service and a checklist from Twilio. They reach the same conclusion: AI is useful where it has documentation and access to data, and harmful where the customer sees no path to a human. A bot that answers cheaply and misses the point creates repeat contacts. At the PDR company in the US, AI doesn't answer the customer; it reviews the calls, and I'll explain why.

What came out

Wonderchat via CustomerThink, October 23, published a study of 5,848,078 support conversations at more than 1,000 companies from March 2023 to August 2026. The headline number: 687 companies on the same platform, the same models and the same release resolve between 46% and 88% of conversations, with a median of 69%. A 42-point gap on identical technology means the cause lies in how support is organized. The share of conversations the agent closed on its own grew from 56.7% in 2024 to 75.3% in 2026, and the gap stayed the same. Of 5,000 failed conversations, in 58.5% the answer existed nowhere the agent could reach; outdated content accounted for 1.2%. Existing customers with a problem make up 53% of volume and resolve at 60.8%; pre-sales questions are 23% and 73.6%. The authors count a handoff to a colleague as an unresolved conversation.

Seismic, October 20, also via CustomerThink, surveyed 200 C-level executives in North America and Northern Europe (NewtonX, July 20 to 30, 2026): 75% trust a person more when their advice differs from what the AI produced, and 55% go to the vendor's representative when negotiating price. That's about B2B purchases, but the conclusion for service is the same: the higher the stakes, the more a person is needed.

Five9, October 17 on CustomerThink, with the report itself out on October 14: a survey of 3,000 consumers in the US, the UK and Germany in April 2026. 47% find AI in service useful, but only 20% use it “most of the time” or “always.” Trust in AI rises from 26% to 55% when a human option is visible. For 71% it matters to know they're talking to AI. 92% give a bot less than 10 minutes before asking for a person, and about a third repeat everything from scratch after the handoff.

On October 21, Twilio gave five checks for a customer engagement platform, and two of them I would hang on the wall. First: start a conversation in chat, continue over SMS, then call and look at what the operator sees; by Twilio's data, 74% of consumers repeat themselves to an AI assistant, and 76% say the human after it knows almost nothing. Second: don't hand the platform your business logic and your system of record.

CustomerThink, October 22, added a roundup on voice AI: Market.us forecasts a voice agent market of $47.5 billion by 2034, and Gartner expects contact center staffing costs to fall by $80 billion in 2026. My copy of the text cuts off after that.

Why this matters to a business

Take a company that rents out construction equipment. A customer writes in the chat: “the compressor won't start, the crew is standing idle.” That's the existing customer with a problem, the 53% in Wonderchat's data who resolve worst of all. If the bot answers with an excerpt from the manual and no “get a person” button, in 10 minutes the customer will call and tell the whole thing over again. If the bot asks for the model and the contract number, pulls up the rental history and hands it to an operator, the operator starts with the solution.

The takeaway: first the bot needs documentation, which you most likely don't have yet, and access to records; the model is secondary. Then a visible path to a person and a handover of context. Without those, the bot saves an operator's minutes and loses the customer.

How we do it

At the PDR company in the US, which I've worked with since 2021, AI doesn't talk to the customer. The system records every call, and AI works out who called, a lead, an existing customer or an insurer, and how the employee answered. A call to the IVR creates the lead in the CRM on its own and attaches the recording and the breakdown. Then the salesperson takes over: sees the lead in the super app, calls and messages the customer from there, and it's the salesperson who talks to the customer.

The idea of a PBX with AI call analysis was mine, and the owner took it. I started the IVR in 2024 and by 2026 had brought it to an internal PBX with custom logic. The result for the company: it used to employ 20 people, today more than 30; the flow of leads multiplies from year to year, while the sales team has grown only a little and keeps up with every inquiry. Every real lead goes through the funnel to the end, and conversion to sale has grown by 30%.

I would repeat this split in any service business: AI prepares the context, the person decides and talks to the customer.

What I would do

Before launching a bot, count which questions come in and where the answers to them live. By Wonderchat, the biggest category of failures is questions nobody has written an answer for, and an experienced employee will have to write them.

Give the bot access to records: the contract, the order status, the rental history. Otherwise it answers in generalities.

Show the “get a person” button from the first message, and on handoff give the operator the whole conversation. Test it by hand: chat, then SMS, then a call.

Say that AI is answering: by Five9, that matters to 71% of people.

How to connect calls, the CRM and tasks without losing anything, I covered in the article How to reliably connect documents, CRM and tasks. Directions: chatbots, IP telephony and CRM integrations. Cases: technical partner to a PDR company and a service team raised conversion by 30%. Other reviews: how AI call analysis works for us and how I read the numbers in vendor case studies; earlier I wrote about conversation design for AI support and the economics of AI support by repeat contacts. If you're deciding what to trust to a bot and what to leave to people, tell me what inquiries come in and where they go: Discuss your task.