Every automated phone process will hear the sentence “Can I speak to a person?”, and how it answers says more about the operator than any demo. The route to a human is not an escape hatch for when the AI fails; it is a designed part of the service, with its own rules, capacity, timing and measurement.
We have built human routes into our processes since our first large deployments, and we have run them at scale across collections, sales, appointments, surveys and customer service. This piece sets out what the research says customers expect, why the phone makes the handover harder than in a chat window, and how we design the route out: when to stop, how to transfer and what to do when nobody is available.
What customers say, and what the numbers do not say#
Two Gartner releases from 2026, each reporting on a survey of 3,566 B2B and B2C customers, put the expectation plainly. In one, 87% said that companies using generative AI for customer service must provide access to a human agent. In the other, only 27% said they would try a chatbot again after a negative experience.
| Finding | Source and sample | How to read it |
|---|---|---|
| 87% say GenAI service must provide access to a human agent | Gartner, 4 Aug 2026; 3,566 customers | A stated preference, not a legal duty |
| 27% would try a chatbot again after a negative experience | Gartner, 2 Sep 2026; same sample | About chatbots, not only phone channels |
| 85% of service leaders are expanding human agents’ responsibilities | Gartner, 28 Apr 2026; 321 leaders | Leaders’ plans, surveyed Sep–Oct 2025 |
| 95% of consumers expect their information to carry across channels | Genesys, State of Customer Experience 2026 | Vendor research (contact-centre software) |
Figures as summarised from the publishers’ releases; see sources for full methodology.
It is worth being precise about what these figures show. They are survey answers about expectations, mostly concerning chatbots rather than telephone conversations, and they do not tell you how often a well-designed voice process actually needs a human. But they point in one direction. A customer who feels trapped in an automated channel does not come back to it willingly, and the cost of a bad automated experience is paid not in that call but in every later contact that customer refuses.
Why the phone makes it harder#
In a chat window, a “talk to an agent” button can sit on the screen for the whole conversation. On the phone there is no button. The customer has to say it, the system has to understand it, and the handover has to happen in real time, while the customer waits on the line, with a team on the other side that may be busy, closed or not set up to receive the call.
Outbound adds a further asymmetry. When a customer calls you, they have chosen the moment and they want something from you. When you call them, you have interrupted their day. A customer who asks for a person in an outbound call may want help, may want to complain about being called, or may simply be testing whether anyone is there. Each needs a different answer, and none of them should hear “I didn’t understand”.
We have treated the human route as part of the design for a long time. One of our earliest public deployments, the passenger information line at Gdańsk Lech Wałęsa Airport, launched in 2017, was built with handover in mind; in 2018 roughly 70% of its calls ended with the automated assistant and roughly 30% were transferred to a consultant. The ratio matters less than the principle it reflects: the machine takes what is repetitive, and people take what needs judgement, empathy or authority. We said in 2016 that this was how automation should divide the work, and the research above suggests customers agree.
Knowing when to stop#
The first design decision is the list of situations in which the automated conversation stops being the right tool. In regulated collections that list is long and specific, and it is agreed with the client before launch, not discovered in production.
- The customer asks for a person. The plainest trigger, and the one that must never be ignored or argued with.
- Verification fails. The customer cannot or will not confirm their identity, so nothing about the debt can be discussed.
- Recording consent is withdrawn. The conversation cannot continue on the recorded channel.
- A death or a legal dispute is mentioned. These need people with authority and care, not a script.
- A data erasure request or a processing objection. These have legal consequences and must be routed and recorded correctly.
- Financial hardship is declared. The customer needs the bank’s hardship handling, not a negotiation over a payment date.
Each trigger maps to a defined outcome: a live transfer to the client’s team, a hard stop in which the call ends politely and the case is flagged in the daily output so the client’s systems can act on it, or a referral to the right channel or service. Which outcome applies is configured per deployment and per market, because banks’ obligations and their local teams differ between countries.
Outside collections the triggers change but the logic is the same. In lead qualification, the trigger is positive: a customer who is interested, consents and is ready to talk to an adviser now. In a satisfaction survey, it is a respondent who wants to complain rather than rate. In appointment management, it may be a patient whose question is clinical rather than administrative. In every case the decision about when to stop belongs in the specification, approved by the client, not in the improvisation of a model.
Never make the exit hard to find#
There is a temptation, wherever automation is judged by the share of calls it keeps, to make the route to a person slow or obscure. We think that is the same mistake as per-minute billing: a metric that rewards the operator for something the client does not actually want. A customer kept in an automated conversation against their will is not a contained contact; it is a complaint that has not been filed yet.
The transfer is a product#
Once the decision to hand over has been made, the transfer itself has to work. That sounds obvious and is where a handover most often goes wrong, because the transfer depends on telephony, on the client’s systems and on a team of people, none of which the conversation designer controls.
Where the call goes#
We run our own telephony, so the mechanics of a transfer are part of our specification rather than a dependency on someone else’s. A call can be transferred to a phone number, to a queue or to a specific skill group in the client’s contact centre, and it can travel over the public network, over a VPN or over a SIP connection where the client has one. The choice affects cost, quality, what information can travel with the call and how failures are detected.
When the call goes#
A transfer only makes sense when someone can answer. Transfers are therefore limited to the receiving team’s working hours, and in sales they are paced to that team’s capacity. A contact centre always has a finite number of people; transferring more interested customers than it can take means those customers wait, lose interest or hang up, and the effort of qualifying them is wasted. We would rather schedule a callback than put a warm lead into a queue nobody will answer in time.
The pacing works in the client’s favour in a second way. When the bank’s or insurer’s advisers receive only customers who have already said they are interested and agreed to talk, they spend their day on real sales conversations rather than on persuading people that a conversation is worth having.
What travels with the call#
The single most irritating thing about a transfer, for the customer, is having to start again. Genesys’s 2026 research, which is vendor-produced, reports that 48% of companies still do not pass data from virtual agents to human agents, while 95% of consumers expect their information to carry across channels so they do not have to repeat themselves. Whatever the exact figures in your market, the experience is familiar to anyone who has been asked for their date of birth three times in one call.
What can travel depends on the client’s systems and on the integration we agree, which is why we treat it as part of the specification rather than a nice-to-have. At a minimum the receiving team should know who the customer is, whether identity has been verified, why the call was made, what has been said so far and why it was handed over. The case’s dialogue and business statuses, recorded for every conversation, are the raw material for this; the integration decides whether it arrives on the adviser’s screen in time or in the next morning’s file.
Every day, in every project, our people listen to selected calls, especially unusual ones, to hear with their own ears what technology still cannot. – Dawid Wójcicki, Founder & CEO of Vee
When nobody is there#
Many requests for a person arrive when no person is available: in the evening, at the weekend, or during a peak when the client’s team is full. An automated process that can only transfer has nothing to offer at those moments. One that is designed for them can still end the call well.
- Callbacks. Whether we accept callback requests, and until what hour, is agreed before launch. A callback request is its own outcome with its own retry clock, different from a missed call or a busy line.
- Scheduled transfer. For sales, a call back at a time the customer chooses, inside the advisers’ hours, often converts better than a transfer into a busy queue.
- Text follow-up. A short message confirming what was agreed, or giving the number to call, closes the loop when the conversation cannot.
- Honest closure. If no human route exists for this process at this hour, the customer should be told when and how they can reach someone, not left to try again blindly.
Some processes run with a live fallback to the client’s contact centre; others run without one, for example where the client has no team for that process at all. Both are legitimate choices. What is not legitimate is to leave the choice unmade and let customers discover it.
What the route out looks like by sector#
The principles are the same everywhere, but the route out takes a different shape in each kind of process, because the reason for the handover and the people at the other end are different.
Early collections#
Here the route out is mostly about protection. The exception paths above exist because a customer in arrears may be in a difficult situation, may be the wrong person, or may have rights the conversation must respect. The handover goes to the bank’s own team or to its hardship process, and a hard stop is often the correct outcome rather than a transfer. A good collections dialogue recognises these moments early, before it has asked for a payment date that the customer cannot honestly give.
Sales and lead qualification#
Here the route out is the goal. The automated call exists to find the customers who want to talk to an adviser now and to hand them over while their interest is warm. The design questions are capacity, hours and timing: how many live transfers the sales team can take each hour, what happens to the interested customer at six in the evening, and how quickly a scheduled callback is made. A fresh lead cools by the hour, so the handover has to be faster than the cooling.
Appointments and healthcare#
In appointment confirmation most calls end without a human, because the information needed is simple: will the patient attend, or should the slot be offered to someone else? The route out matters for the minority of calls that are not simple, such as a question about preparation for a procedure or a request that only the clinic’s staff can handle. The handover there is often a message to the facility or a callback from its reception rather than a live transfer, because clinic phone lines are often busy enough already.
Customer service and surveys#
In inbound customer service, calls are answered without a queue and answers come from content approved in advance; the route out handles what falls outside that content. In satisfaction surveys it handles the respondent who wants to complain. In both, the worst outcome is a customer who explained their problem to the automated conversation and then has to explain it again to a person, which is why what travels with the call matters as much as the transfer itself.
Deciding the division of roles before launch#
Before any process goes live, we agree its operating rules with the client. The human route is a central part of that conversation, alongside the schedule and the number of attempts.
- When we call, in which windows, how often and how many times.
- Whether we take callbacks, and until what hour.
- In which situations we transfer to the client’s team, and to which number, queue or skill.
- Whether there is a live fallback to the contact centre, or none for this process.
- What information accompanies a transfer, and in which system the adviser sees it.
- How many transfers per hour the receiving team can absorb, and what happens above that level.
These rules are tested before launch, often in a “friends and family” round in which the client’s own employees call the process and try to break it, and then on a sample of the base rather than the whole database. A transfer that rings in an empty queue during a pilot is cheap to fix; the same fault discovered in full production is a stream of complaints.
The research on service leaders points the same way. Gartner found that 85% of the leaders it surveyed were expanding their human agents’ responsibilities, and a separate Gartner survey found that only 21% had reduced headcount because of AI, while 7% had already rehired. The people who receive transferred calls are not a leftover from the automation project. They are part of its design, and their capacity and skills should be planned with it.
Measuring the route out#
If the human route is a product, it needs product measures. A single “transfer rate” tells you very little, and treating it purely as a failure rate pushes the design in the wrong direction.
| Measure | What it tells you |
|---|---|
| Requests for a person, by point in the dialogue | Where the conversation stops meeting the customer’s needs |
| Transfers completed versus attempted | Whether the telephony and the receiving team are actually there |
| Wait and abandonment after transfer | Whether pacing matches the team’s capacity |
| Outcome after transfer | Whether the handover produced the result it was meant to |
| Callbacks requested versus completed | Whether the out-of-hours promise is kept |
| Hard stops by reason | Whether exception paths are triggered correctly and recorded |
Measures are agreed per process; outcome after transfer usually depends on data from the client’s systems.
Measures show where to look; listening shows what is happening. Requests for a person concentrated at one point in the dialogue usually mean that the conversation is asking something the customer cannot answer, or offering something they do not want. Because Vee employs no call agents, a correction made there applies to every live conversation at once, and the next day’s numbers show whether it worked.
What to ask a supplier#
- What exactly happens, second by second, when a customer says “Can I speak to a person?” Can you play us a recording?
- Which situations trigger a transfer, a hard stop or a referral, and who approved that list?
- How do you pace transfers to our team’s capacity and working hours?
- What information reaches our adviser with the call, and on which screen?
- What does the customer hear when nobody is available, and how are callbacks scheduled and kept?
- How do you report requests for a person, and what did you change last month because of them?
Customers will forgive an automated call that is short, clear and honest about what it is. They are much less forgiving of one that will not let them go. Designing the route out is how you earn the second call, and in a regulated business, it is also how you show that the automation serves the customer rather than the other way round. It is a large part of how we run every process, and the first thing we ask to see in anyone else’s.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS