AI customer service escalation an agent is also judged by how it steps aside

I

What customers really fear

In July 2024, Gartner published a survey conducted in December 2023 among 5,728 customers: 64% of them would prefer that companies not use AI in their customer service, and 53% would consider switching to a competitor if they learned a company was going to do so¹. The top concern cited is neither jobs nor errors. It is the increased difficulty of reaching a person¹.

So customers don’t reject the technology first: they dread hitting a wall. Gartner’s Keith McIntosh recommends that the AI agent tell the customer it will connect them with an advisor if it can’t find a solution, then that the exchange turn seamlessly into a conversation with that advisor, who picks up where the agent left off¹.

In other words, trust in an AI agent is built largely on the certainty that there is a way out. An agent that knows when to step aside makes everything it handles alone the rest of the time acceptable.

II

The legal framework: a growing right to human intervention

Several European and French texts, already in force or about to apply, require human intervention in specific situations. They don’t create a general right to speak to an advisor in every customer service setting. They do, however, point in a clear direction.

GDPR: fully automated decisions

Article 22 of the GDPR establishes the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects the person². Where such a decision is permitted, because it is necessary for a contract or based on explicit consent, the controller must at a minimum guarantee the right to obtain human intervention, to express their point of view and to contest the decision².

Most answers from an AI agent are not decisions within the meaning of this text. A refund refusal or an account closure issued without human review, however, can come close.

Distance financial services: a right in force since June 19, 2026

Directive (EU) 2023/2673 requires Member States to ensure that, when the trader uses online tools for a financial services contract concluded at a distance, the consumer has the right to request and obtain human intervention at the pre-contractual stage and, in justified cases, after the contract is concluded³. Its recitals expressly cite chatbots and automated advice³. Member States have applied these rules since June 19, 2026³.

In France, Ordinance No. 2026-2 of January 5, 2026 created Article L. 222-5-1 of the French Consumer Code, in force since June 19, 2026. It applies to financial services contracts concluded at a distance: when the trader uses online tools, the consumer must be able to contact a human being and talk with them, in the language of the pre-contractual information, before the contract is concluded and afterward where necessary for its proper understanding or performance⁴⁵. The article does not apply to products subject to specific rules on the explanations to be given to the consumer⁴.

Consumer credit: from November 20, 2026

Under Article 18 of Directive (EU) 2023/2225, when the creditworthiness assessment relies on automated processing, the consumer can obtain human intervention from the lender: a clear explanation, the opportunity to express their point of view and a review of the decision⁶. Member States must apply these rules from November 20, 2026⁶.

AI Act: telling customers they’re talking to an AI

Since August 2, 2026, Article 50(1) of Regulation (EU) 2024/1689 requires providers to ensure that AI systems intended to interact directly with people are designed so that those people are informed they are interacting with an AI system, unless this is obvious from the circumstances and context of use⁷. Regulation (EU) 2026/1744 of July 8, 2026, known as the digital omnibus on AI, postponed several deadlines in the text, but the only transitional period it adds to Article 50 concerns the marking of generated content provided for in paragraph 2⁸. This transparency is the prerequisite for handoff: a customer can only ask for a human if they know they don’t have one yet.

What has been announced, and what hasn’t yet

The European Commission is preparing a Digital Fairness Act. According to the European Parliament’s research service (note of October 5, 2026), this proposal is expected in the fourth quarter of 2026⁹. Its final content is therefore not known. Gartner, for its part, predicted as early as February 2024 that the European Union would mandate a “right to talk to a human” in customer service by 2028¹⁰. That is a forecast, not a legal text.

One point of French law is worth recalling: the phone number provided to obtain proper performance of a contract or the handling of a complaint cannot carry a surcharge (Article L. 121-16 of the French Consumer Code)¹¹. This rule therefore concerns the number offered when the handoff to an advisor goes through the phone for one of these reasons.

III

The signals that should trigger a handoff

The cases that call for a human are largely predictable. They should be written down, tested and maintained rather than left to the model to improvise. They fall into four families.

The explicit request

When a customer asks for an advisor, the agent doesn’t negotiate. It can offer one last time to handle the request if it is certain it can, but it doesn’t insist. Difficulty reaching a person is precisely the first fear cited by customers surveyed by Gartner¹. For a financial services contract concluded at a distance through online tools, it is also a right⁴.

Frustration

The tone changes, messages get shorter, capital letters appear, the customer writes that they’ve “already explained.” These markers can be described in the agent’s rules. The right reflex is not to pile on spoken empathy, but to offer the handoff before the customer asks for it.

The failure loop

The agent has rephrased twice, the customer has asked the same question again, or the same action has failed several times. Beyond a threshold you set, continuing costs more than handing over. That threshold should be low: a customer doesn’t count the agent’s attempts, they count their wasted time.

High stakes

Some topics warrant a human even when the agent could answer: formal complaint, cancellation, health question, dispute, vulnerable situation, amount above a defined ceiling. The cost of an error there exceeds the benefit of an automated answer, and some of these cases may fall under the texts cited above²⁶.

IV

Decision matrix

The table below offers a working basis. The thresholds and delays are examples to adapt to your business, not standards.

V

The warm handoff: never make the customer repeat themselves

The rule is simple to state and demanding to uphold: the customer shouldn’t have to repeat what they just wrote. That is the meaning of Gartner’s recommendation, with an advisor who picks up where the agent left off¹.

A warm handoff passes along three things. First, a two- or three-sentence summary of the request, written for the advisor and not for the customer. Then the verified elements: confirmed identity, order or case number, actions already taken and their outcome. Finally, the reason for the handoff, so the advisor knows whether they’re getting a complex request, an unhappy customer or a sensitive case. The advisor keeps the full history at hand, but shouldn’t need to read it to get started.

On the customer side, the handoff must be announced precisely: what will happen, within what time, and what has already been passed along. Here is an example message.

I’m connecting you with an advisor from our customer service team. I’m passing along a summary of our conversation: your order No. 48213, the parcel shown as delivered but not received, and the check already made with the carrier. You won’t have to explain everything again. The estimated wait is four minutes. You can stay on this page, the conversation continues here.

This message summarizes the facts, promises not to make the customer repeat anything, gives a delay and says where the conversation continues. It promises no outcome that the advisor hasn’t yet approved.

VI

Continuity across channels

A customer often starts on one channel and finishes on another: a question on the website, a clarification on WhatsApp, a call to wrap up. Human handoff must not reset the case at every channel change.

In practice, this requires a conversation ID shared across channels, a single case file on the advisor side and a clear rule on the reply channel: by default, the one the customer is using at the time of the handoff. If the advisor needs to call back by phone, the customer must have agreed to it and know which number will appear.

VII

Outside business hours: deferred handoff and callback commitment

An AI agent answers at any hour, your teams don’t. At night and on weekends, handoff therefore takes another form: deferred handoff.

The agent announces it honestly. It states that advisors are available from a given time, offers the customer a choice of callback slot or reply channel, creates the case with the full context and confirms the commitment in writing. The European text on distance financial services also holds, in recital 40, that pre-contractual human intervention should be obtainable free of charge during the trader’s office hours³. Beyond finance, the same logic usefully applies: a dated, kept commitment is better than advertised availability that is impossible to deliver.

A health or distress situation is the exception: it can’t wait until the next day, and the agent must immediately direct the person to the appropriate emergency numbers, following instructions written in advance.

VIII

Feedback loop: every handoff teaches something

A human handoff is also data. Every time an advisor takes over a conversation, they know why the agent wasn’t enough: missing information in the knowledge base, a rule that was too strict, a case never seen before, or on the contrary a situation that rightly belonged with a human.

This information needs to flow back. The advisor records the real reason at the end of handling, in a few clicks. A regular review of handoffs then makes it possible to tell three categories apart: legitimate handoffs, to keep; avoidable handoffs, which call for a correction to the agent; and late handoffs, where the agent should have passed the conversation on sooner. The handoff rate is not an indicator to push down at any cost. A low rate achieved by holding on to customers who wanted out is a bad result.

IX

Humans: a scarce, and therefore premium, resource

What follows is analysis, not observation. As AI agents take on repetitive requests, advisors’ time concentrates on what remains: complex cases, emotionally charged moments, binding decisions. The advisor stops being the first tier of service and becomes its most valuable tier.

A well-designed handoff saves human time on every case, since the advisor starts with the context instead of reconstructing it. And access to a human can become a quality marker, deployed where it creates the most value: retaining an unhappy customer, supporting an important decision, protecting a vulnerable person. European texts target precisely these high-stakes moments³⁶. Companies that have organized this handoff before being forced to will start with a head start.

X

What this means for your organization

Start by writing your handoff rules. List the signals from the matrix that apply to your business, set the thresholds (number of failures, amounts, sensitive topics) and write them in plain language, so that your business teams can review and edit them.

Then check your regulatory exposure. If you sell financial services at a distance through online tools, Article L. 222-5-1 of the French Consumer Code already applies, unless product-specific rules say otherwise⁴. If you grant consumer credit, prepare for the November 20, 2026 deadline⁶. If your AI agent makes decisions with significant effect, review them against Article 22 of the GDPR². In every case, make sure your customers know they’re talking to an AI⁷.

Define the format of the file passed along and test it with your advisors: they are the ones who know what they’re missing when they take over.

Plan for off-peak hours: an honest unavailability message, callback slots, emergency instructions for health or distress situations.

Finally, set up the feedback loop: reason qualification by the advisor, periodic review, agent corrections. Your teams stay in control of what the agent handles alone and what it hands over to them.

XI

Where QLOA stands

XII

Frequently asked questions

Does a customer have the right to talk to a human rather than an AI?

Not systematically. European and French texts provide for it in specific situations, notably: fully automated decisions with significant effect (Article 22 of the GDPR), financial services concluded at a distance via online tools (Article L. 222-5-1 of the French Consumer Code, since June 19, 2026) and, from November 20, 2026, automated creditworthiness assessment for consumer credit. Elsewhere, human handoff remains a customer service quality choice.

When should an AI agent transfer the conversation to an advisor?

In four families of situations: when the customer explicitly asks, when they show signs of frustration, when the agent fails several times on the same request, and when the stakes are high (complaint, cancellation, health, dispute, vulnerability, large amount). These rules should be written and tested in advance, rather than left to the model’s discretion.

How do you keep the customer from repeating their request after a transfer to an advisor?

By passing the advisor a summary of the exchange, the elements already verified (identity, order or case number, actions attempted) and the reason for the transfer. The full history remains accessible. On the customer side, the agent explicitly states what has been passed along, the estimated delay and the channel where the conversation continues.

What should an AI agent do outside customer service hours?

It offers a deferred handoff: it says when advisors will be available, lets the customer choose a callback slot or channel, creates the case with all the context and confirms the commitment in writing. Health or distress situations are the exception and must be directed immediately to the appropriate emergency numbers.

Should customers be told they’re talking to an artificial intelligence?

Yes. Since August 2, 2026, Article 50 of the European AI regulation requires providers to design systems intended to interact directly with people so that those people are informed they are talking to an AI, unless this is clear from the context. This information also conditions the handoff: a customer only asks for a human if they know they don’t have one yet.

XIII

Sources

1. 1

   Gartner, “Gartner Survey Finds 64% of Customers Would Prefer That Companies Didn’t Use AI For Customer Service,” press release, July 9, 2024. gartner.com

2. 2

   Regulation (EU) 2016/679 of April 27, 2016 (GDPR), Article 22, EUR-Lex. eur-lex.europa.eu

3. 3

   Directive (EU) 2023/2673 of November 22, 2023 amending Directive 2011/83/EU as regards financial services contracts concluded at a distance, and repealing Directive 2002/65/EC, recitals 15 and 40, Article 1 (Article 16e inserted into Directive 2011/83/EU) and Article 2, EUR-Lex. eur-lex.europa.eu

4. 4

   French Consumer Code, Article L. 222-5-1, created by Ordinance No. 2026-2 of January 5, 2026, in force since June 19, 2026, Légifrance. legifrance.gouv.fr

5. 5

   Report to the President of the Republic on Ordinance No. 2026-2 of January 5, 2026 on the distance marketing of financial services to consumers, JORF No. 0004 of January 6, 2026, Légifrance. legifrance.gouv.fr

6. 6

   Directive (EU) 2023/2225 of October 18, 2023 on credit agreements for consumers and repealing Directive 2008/48/EC, Article 18(8) and Article 48, EUR-Lex. eur-lex.europa.eu

7. 7

   Regulation (EU) 2024/1689 of June 13, 2024 laying down harmonised rules on artificial intelligence, Articles 50 and 113, EUR-Lex. eur-lex.europa.eu

8. 8

   Regulation (EU) 2026/1744 of July 8, 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (digital omnibus on AI package), Official Journal of the European Union, July 24, 2026, EUR-Lex. eur-lex.europa.eu

9. 9

   David Ashton and Maria Niestadt, “Towards a digital fairness act,” EPRS, European Parliament, October 5, 2026. eprs.europarl.europa.eu

10. 10

    Gartner, “Gartner Predicts that with Advances in GenAI 30% of Fortune 500 Companies Will Consolidate Postpurchase Tasks Into a Unified Customer-Facing Role by 2027,” press release, February 19, 2024. gartner.com

11. 11

    French Consumer Code, Article L. 121-16, Légifrance. legifrance.gouv.fr