# Human handoff: when an AI conversation should pass to a person

- URL: https://plandora.com/blog/human-handoff-for-ai-conversations/
- Authors: Emily Carter (Editorial)
- Published: 2026-10-22
- Updated: 2026-10-22
- Facts checked: 2026-10-22
- Category: Trust

> Human handoff is the point where an AI conversation passes to a person. Candidate triggers are a visitor request, a sensitive or regulated topic, no approved answer, distress, a sale that needs a person, and repeated failure. In Plandora a request does not stop the Mind; a person taking over does. Say who replies and when, and measure the wait.

## Key takeaways

- Gartner's survey of 5,728 customers (December 2023) found only 14% of issues fully resolved in self-service, so design the way out before the answers.
- Seven candidate triggers to configure and test: explicit request, sensitive or regulated topic, low confidence, complaint or distress, buying that needs a person, repeated failure and VIP.
- In Plandora a request is not a transfer: the Mind keeps replying until a person takes over, so Boundaries must say what it tells the visitor while it waits.
- A good handoff message names who is picking up, what happens next and when, and says the team can read what was already said.
- Measure handoff volume, waiting time, reason and outcome, and read the conversations before judging the numbers.

An AI conversation should pass to a person when the visitor asks for one, when the topic is sensitive or regulated, or when replies keep failing. In Plandora a request is not yet a transfer: the Mind keeps replying until a person takes over. Write the triggers, and what the Mind says while it waits, before launch.

## Is asking for a person the same as handing over?

No. In Plandora a conversation is in one of three states, and the Inbox shows which:

| State | What you see in the Inbox | Who replies |
|---|---|---|
| Mind replying | No tag on the conversation | The Mind |
| Requested | A Requested tag and a "Human assistance requested" banner, with Take over and Dismiss | The Mind, still |
| Human handling | A Human handling tag | The person; the Mind is paused on that conversation |

When the Mind asks for a person, your team is notified and the request is recorded, but the Mind keeps the conversation. By default it tells the visitor it has asked the named person and can keep helping meanwhile. Only a person taking over, or sending a reply, pauses it; Release to AI hands it back.

So read Boundaries carefully. A rule such as do not answer pricing questions once a person has been requested is an instruction the Mind follows in its replies. It can be imperfect, and it is not the pause. A request also needs a real tag from the mention picker: without one the Mind has no one to ask, and its instructions tell it to say live handoff is not set up.

## Why does human handoff matter?

Self-service resolves fewer problems than its usage suggests, and customers say they want a person within reach. Three surveys point the same way, but each counted something different:

| Source | Who was asked | What was counted | Result |
|---|---|---|---|
| [Gartner survey, published August 2024](https://www.digit.fyi/only-14-of-customer-service-issues-are-resolved-in-self-service/) (as reported by DIGIT) | 5,728 customers, surveyed December 2023 | Customer service and support issues fully resolved in self-service | 14% overall; 36% for issues customers called "very simple" |
| [SurveyMonkey, December 2025](https://www.surveymonkey.com/curiosity/customer-service-statistics/) | 2,017 US adults, non-probability online panel, modelled error estimate of 2.5 points | Share who strongly prefer a human over an AI agent | 79% (SurveyMonkey's headline) |
| [Twilio, 2025](https://www.twilio.com/en-us/press/releases/twilio-s-latest-report-highlights-conversational-ai-s-rapid-adop) (vendor-reported) | 4,800 consumers in 15 countries, fielded August to October 2025 | Share who call it important to be able to move from an AI agent to a human one; share who reported a smooth handoff | 78%; 15% |

Mind the limits. Gartner's figure is about self-service in general, not AI chat. Twilio sells communications software, hence vendor-reported. SurveyMonkey gives 79% overall but 70% for women and 60% for men on the same preference, which cannot be reconciled for one measure, and prints no question wording, so treat 79% as its headline, not a checked figure. Its 89% figure has no sample or date and is not used. The three do not combine into one statistic. They suggest a direction: customers want a route to a person, and few say they have had a good one.

The US Consumer Financial Protection Bureau's June 2023 report on [chatbots in consumer finance](https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/chatbots-in-consumer-finance/) warns of "doom loops": loops of "repetitive, unhelpful jargon or legalese without an offramp to a human customer service representative". It concerns US financial institutions, but a visitor who cannot leave the bot has little reason to trust the business behind it. A Digital Mind is not the human expert and should not be presented as one, and a person should stay involved in financial, legal, medical and other consequential decisions. Handoff is how that holds in practice.

## When should an AI conversation hand off to a person?

Seven candidate triggers make a starting set to configure and test, not a standard or a complete list. The third column is what your Boundaries should tell the Mind to do; the Mind follows those words, but only a person taking over pauses it.

| Trigger | What it looks like | What the Mind should do | Who picks it up |
|---|---|---|---|
| Explicit request | "Talk to someone", "human", "agent" | Confirm and request a person; never ask the visitor to justify it | The tagged person or team |
| Sensitive or regulated topic | Symptoms, medication, legal or tax questions, investment suitability, privacy requests | Share only approved general information, say a qualified person will review, request one | A qualified person |
| Low confidence | No approved answer, or only a partial one | Say so, request a person, leave the gap visible | Support, plus the Knowledge owner |
| Complaint or distress | Anger, a refund dispute, bereavement | Acknowledge briefly, request a person, give an urgent route if one exists | A trained person, quickly |
| Buying that needs a person | A bespoke quote, negotiated terms, an exception, or readiness to commit | Collect what the person needs, then request one | Sales, or the expert |
| Repeated failure | The same question rephrased, or two unhelpful answers in a row | Stop the loop and offer a person | Support |
| VIP | A known high-value customer, partner or journalist | Ask for the account owner; this needs an identified contact | The account owner |

### Sensitive or regulated topics

Medical, legal, tax, investment and safety questions belong with a qualified person. This article is not legal advice, and rules differ by country and sector, so have a qualified adviser review your triggers.

An Inbox handoff is not an emergency service: while a request is open the Mind keeps replying, and nobody may see it for hours. Where a message could mean immediate danger, write a separately approved reply that points the visitor to an immediate route, such as the local emergency number, and says not to wait for the Inbox. Plandora's [telehealth and digital health page](/solutions/telehealth-digital-health) says urgent questions go to a person; that holds only if the urgent route is in the Mind's own reply.

One UK comparison: the Information Commissioner's Office, in [draft guidance on automated decision-making](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/automated-decision-making/what-are-the-adm-safeguards/) last updated on 31 March 2026, says Article 22C of the UK GDPR, as changed by the Data (Use and Access) Act 2025, requires safeguards for solely automated decisions with significant effects, including the right to "obtain human intervention on the part of the controller". That is a review after a decision, which the ICO separates from human involvement during the process, so it is related to a chat handoff but not the same. A chat reply is not automatically such a decision; take advice where a Mind approves, refuses or prices something for a customer.

### Low confidence

A Mind that guesses is the failure described in [why AI replies go wrong](/blog/why-ai-replies-go-wrong). Better to say no reliable answer is available, ask for a person, and leave the gap visible. In Plandora, Insights marks answers as, for example, No answer or Partial answer, and "Answered" is a system assessment, not proof the visitor was happy.

### Buying that needs a person

Routine availability and booking questions differ from negotiated terms: the Mind can answer them from approved Knowledge, and Booking can offer your booking page when your conditions are met. A bespoke quote, an exception or a commitment belongs with a person. Your Mind can move a contact into an Automatic stage when a reply gives enough evidence, but the won stage is always manual, and a booking link click shows interest without confirming a booking.

### Repeated failure

Two failed attempts on the same question is a sensible starting rule, a design choice rather than a research finding. Tighten it for regulated topics; loosen it where a third attempt costs the visitor little.

### VIP

A VIP is defined by your team's data, not the conversation, and the Mind can only act on it once the conversation is tied to an identified contact: a known WhatsApp number or email address, or a signed-in visitor through the website widget's identify option. Mark the contact with a tag or property in Contacts. Whether a VIP rule can also route conversations automatically depends on your Playbook and Automations, so test it before relying on it.

## How does handoff work in Plandora?

A handoff runs from the Playbook to the Inbox. Audience features such as the Inbox can be available through guided rollout or self-serve, depending on your plan.

1. **Write the boundary.** In Playbook › Boundaries, say when a question needs a person, tag who should handle it with the mention picker (typed text creates no tag), and say what the Mind may tell the visitor while it waits.
2. **Test it.** In the chat preview, try a question that should be handed off, then send one more message while the request is open. Check Audience › Inbox for the request.
3. **Watch the queue.** Unreplied lists contacts whose latest message came from the contact, so a conversation the Mind has already answered while a request is open may not appear there. Look for the Requested tag in the list too. Mine (assigned to you) and AI handling (conversations your Mind has, open requests included) are other quick filters.
4. **Reply or take over.** Type a reply, or select Draft with Mind and edit it; drafting uses AI credits and the button shows the cost. If the visitor escalated because earlier replies kept missing, read a draft as critically as a colleague's, or write your own. Sending a reply pauses your Mind on that conversation and assigns it to you. The handler menu also offers Take over, Assign to and Release to AI, plus Dismiss request when the Mind asked but can carry on.
5. **Keep the team informed.** Add comment saves an internal note the contact never sees. Summarize writes a short summary you can reuse as the close summary.
6. **Finish.** Close moves the conversation to Closed, with an optional summary. Reopen restores it. Archive hides the contact from active views and outreach but keeps the history. If you turn on auto-closing of idle conversations in Stages, it skips conversations held by a person, so close or release what you finish.

### Does the context travel with the conversation?

Largely, yes. The conversation stays in the Inbox thread, and Contact details shows the contact's stage, tags, properties, collected answers and private notes, so a person can read what was said, though they may still need to confirm details. In [Twilio's 2025 survey](https://www.twilio.com/en-us/press/releases/twilio-s-latest-report-highlights-conversational-ai-s-rapid-adop) of 4,800 consumers (vendor-reported), 54% believed AI agents rarely or never have context about them as customers.

### What do channels change?

Replies follow the channel. WhatsApp, Messenger and Instagram replies fall within a 24-hour window; on WhatsApp, once it closes you need an approved template. You cannot reply on a disconnected channel, or when the contact has no address for it. Website chat is the sharp edge: a reply goes into the chat thread, and Plandora's Inbox documentation describes no other route to a visitor who has left. If a website visitor may leave before a person answers, have the Mind collect an email address or WhatsApp number first. Email is outbound only.

## How do you write the handoff message?

By default the Mind tells the visitor it has asked the person to step in and can keep helping, and it is told not to promise a time, so test any window you add to Boundaries. A good message has four parts:

- **Who.** Name the person or team, so the visitor knows a human has been asked.
- **What they can read.** Say the team can read the conversation, so they start from what has been said.
- **When.** State a window you can keep, or when someone is next on shift.
- **An urgent route.** If waiting is unsafe, point to a phone number or emergency services.

A template with placeholders:

> This needs a person, so [name or team] has been asked to pick it up. They can read this conversation and will pick up from here. Expect a reply [within the window you can keep]. If it cannot wait, [urgent route].

An example, not a recommendation of any hours: a visitor disputes a charge late in the evening and the team answers on working days only. The Mind says the accounts team has been asked, that they reply on working days and can read the conversation, and where to go if it cannot wait. The first human reply then says who is writing and what happens next. Avoid long apologies, asking the visitor to rephrase, and promising an outcome, because a refund or exception is the person's decision.

## What should you disclose when a conversation changes hands?

Tell the visitor when a person joins and when the Mind returns, or they may credit the expert with the AI's answers, or blame the AI for a person's reply. Disclosure rules vary by jurisdiction; [AI disclosure and the EU AI Act](/blog/ai-disclosure-and-the-eu-ai-act) covers the European rule. People who take over can see transcripts and contact details, so handoff widens who handles personal data; Plandora's FAQ puts responsibility for notices, permissions and lawful use of collected data on the deploying business. See also [data protection for AI chat in Southeast Asia](/blog/data-protection-for-ai-chat-in-southeast-asia).

## How do you staff handoff and set response-time expectations?

No response-time benchmark is cited here, because none could be verified. These practices follow from how the product behaves:

- **Name one owner per trigger.** A sensitive-topic handoff should reach someone qualified, not whoever is online.
- **Promise a window you can keep.** A long one that is met beats a short one that is missed.
- **Decide what happens out of hours.** Because the Mind keeps replying until someone takes over, write what it should say when nobody is on shift.
- **Check for open requests on a schedule.** Look for Requested conversations as well as Unreplied. Plandora's documentation gives an Automation that emails a list of people waiting for a reply as a starter example; check what that list includes before relying on it for requests. Deciding who deserves a person is judgement, the argument in [automation should follow judgment](/blog/automation-should-follow-judgment).
- **Watch the channel clock.** A WhatsApp handoff that outlasts the window can only be restarted with a template.
- **Agree ownership across teams** before rollout; [rolling out Digital Minds across an organisation](/blog/roll-out-digital-minds-across-an-organisation) covers the structure.

## What should you measure?

Plandora's documentation does not describe a handoff-wait report, so the waiting measures below are proposals you track yourself.

| Measure | The question it answers | Where to look |
|---|---|---|
| Handoff requests | Are triggers firing too often or too rarely? | Plandora's guidance: when requests for human help rise, read those conversations in the Inbox |
| Wait for a person | How long does a visitor wait? | Proposed: time from the request (the banner shows when the person was asked) to the first substantive reply from a person, not the Mind's holding message. Track it apart from the time until someone takes over, and watch the slowest waits as well as the typical one |
| Open requests | How many are still waiting, and for how long? | The Requested tag and banner, counted at set times |
| Reason | Which trigger fired? | Hover over the banner on an open request to read the reason recorded with it; sample these monthly and use Add comment for your own notes |
| Gaps | Which questions had no answer? | Insights, Conversations: No answer and Partial answer, then Add an answer |
| Outcome | What happened next? | The close summary, and the contact's Stage |

A falling handoff rate is not automatically good news: visitors may not be finding the exit, the doom loop described above. Read the conversations before you judge the numbers.

## What is a 7-point handoff checklist?

1. **Triggers are written down and tested**, each with a named owner.
2. **Boundaries tag the right person or team** and say what the Mind tells the visitor while it waits.
3. **An explicit request is honoured on the first ask**, with no questions about why.
4. **The handoff message is written and tested**, and emergencies have their own approved reply.
5. **A reply route exists** for each channel, including the WhatsApp window and a way to reach web visitors.
6. **Someone owns the queue**, checks Requested and Unreplied on a schedule, and the out-of-hours message is accurate.
7. **Every handoff ends**: Release to AI or Close with a summary, then review reasons monthly and fix Knowledge.

## Frequently asked questions

### What is human handoff in an AI conversation?

Human handoff is the moment an AI conversation passes to a person, either because the visitor asked or because a rule says the AI should not continue. A good handoff carries the full conversation and any collected details, tells the visitor who is picking up and when, and lets the person start from what was already said.

### When should a chatbot transfer to a human?

Candidate triggers are an explicit request, a sensitive or regulated topic, no approved answer, a complaint or distress, buying that needs a person, repeated failed answers and VIP contacts. Treat the list as a starting configuration, set a named owner for each trigger, and test the handoff before launch rather than after the first complaint.

### Should an AI chatbot always offer a human option?

Rules differ by country and sector, so take qualified advice for yours. Customers say they want the option: in Twilio's 2025 survey of 4,800 consumers (vendor-reported), 78% said it is important to be able to switch from an AI agent to a human agent. A clear route to a person also avoids the doom loops the US CFPB described in 2023.

### Does Plandora pause the Digital Mind when a person replies?

Yes, but not when the Mind merely asks. A request from the Mind notifies your team and the Mind keeps replying. Sending a reply or choosing Take over in the Inbox pauses it on that conversation and assigns it to you, and Release to AI hands it back. Write what the Mind may say while the visitor waits.

### What should you measure after launching human handoff?

Track how often handoffs happen, how long visitors wait for a person (from the request to the first substantive human reply), which trigger fired, which questions had no answer, and what happened afterwards. Read the conversations behind the figures. In Plandora, "Answered" in Insights is a system assessment, not proof that the visitor was satisfied.

## Sources

- [Customer service trends & statistics for 2026: Why consumers still trust humans over AI](https://www.surveymonkey.com/curiosity/customer-service-statistics/), SurveyMonkey, accessed 2026-10-22
- [Gartner survey: only 14% of customer service issues fully resolved in self-service (report of the August 2024 Gartner release)](https://www.digit.fyi/only-14-of-customer-service-issues-are-resolved-in-self-service/), DIGIT, accessed 2026-10-22
- [Twilio's latest report highlights conversational AI's rapid adoption, challenges, and perception gaps](https://www.twilio.com/en-us/press/releases/twilio-s-latest-report-highlights-conversational-ai-s-rapid-adop), Twilio, accessed 2026-10-22
- [Chatbots in consumer finance](https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/chatbots-in-consumer-finance/), Consumer Financial Protection Bureau, accessed 2026-10-22
- [What are the ADM safeguards? (draft guidance)](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/automated-decision-making/what-are-the-adm-safeguards/), Information Commissioner's Office, accessed 2026-10-22