On 28 December 2026 the adaptation period of Spain's new customer service law ends. For anyone running an AI agent, one rule matters most: a customer must be able to ask for a person at any point, including from inside a chatbot. Companies that enter the deadline with evidence of how that works in real conversations will have a calm December. This guide sets out who is covered, what the law asks of AI agents, and an eight-check test you can run on your own agent this week.
The short version
| What | Detail | Source |
|---|---|---|
| The law | Ley 10/2025, of 26 December, regulating customer service. Published in the BOE on 27 December 2025 | BOE-A-2025-26698 |
| The deadline | Covered companies have twelve months from entry into force to adapt: 28 December 2026 | Transitional provision |
| The rule for bots | A bot cannot be the only way to reach the company. The customer can ask for a person at any moment, from the first message | Art. 8.1 and 8.2 |
| The speed | 95% of requests for a person attended, on average, in under three minutes | Art. 8.2 |
| The proof | The company carries the burden of proving that it complies | Art. 4.7 |
Who is covered
| If you are... | Then... |
|---|---|
| A supplier of electricity, gas or water; a passenger transport operator (air, rail, sea, bus or coach); a postal operator | The law applies in full. |
| A telecoms or financial services provider | Your sector's customer service rules come first and Ley 10/2025 applies as a supplement. Some provisions do not apply to finance at all (arts. 13.8, 18, 19, 21, 22 and 23), and sector supervisors oversee compliance. Ask your counsel which provisions your sector rules already cover. |
| Any other company selling goods or services mainly to consumers in Spain | Covered if, in the previous financial year, you (alone or as a group) had at least 250 employees, or turnover above €50 million, or a balance sheet above €43 million. Any one of the three is enough. |
Three details that catch companies out. The law applies whatever the channel: phone, chat, WhatsApp, email or in person. It applies even when customer service is outsourced or run from outside Spain (art. 2.3). And public bodies that provide these services to consumers are covered too.
What the law asks of an AI agent
The law does not ban bots. Its preamble says AI systems may be used to make communication easier, provided that customers can reach personal attention from trained people. These are the provisions that shape an agent's design.
| Requirement | Article | What it means for your agent |
|---|---|---|
| Not the only door | 8.1 | An answering machine, a bot or a similar medium cannot be the sole way to reach customer service. |
| A person on request, from the start | 8.2 | The bot must let the customer ask for personal attention from the options in its main menu, at any moment of the interaction, from the first message. |
| A real operator | 3.11 and 8.2 | Personal attention means a human operator who answers live, not an automated reply. The operator introduces themselves at the start of the conversation, in line with data protection rules. |
| Under three minutes | 8.2 | 95% of requests for personal attention must be attended, on average, within three minutes of the request. |
| The phone line | 10.3 | 95% of calls must be answered, on average, in under three minutes. A call where the customer cannot state the reason and ask for a person does not count as attended. |
| A supervisor on request | 8.3 | An unhappy customer can ask for a supervisor or the quality team inside the same communication. If that is not possible within three minutes, the company can call back the same working day. |
| No hang-ups | 8.4 | The company cannot cut a communication because the wait is long. |
| Trained people | 9.2 | Staff who give personal attention, and the people who design and manage the automated means, need adequate training, including accessibility and co-official languages. |
| Reasoned answers | 13.1 | A resolution has to answer every point the customer raised. Generic replies do not qualify. |
| Resolution clocks | 17 | 15 working days in general. Two hours for incidents affecting the continuity of a basic service. Five days for billing and undue charges. |
| A receipt and a record | 4.2 and 12 | A reference key and, when the customer asks for it, a receipt in a durable medium. If a call, video call or messaging conversation is recorded with the customer's express consent, the copy or transcript is kept at least until the resolution is notified. |
| Proof | 4.7 | The company bears the burden of proving compliance. |
Two laws, two questions
| The question | The law | Applies |
|---|---|---|
| Does the customer know they are talking to a machine? | EU AI Act, Art. 50: say it is an AI no later than the first interaction, unless it is obvious, and never deny it | Since 2 August 2026 |
| Can the customer reach a person? | Ley 10/2025, art. 8 | Adaptation deadline 28 December 2026 |
An agent that dodges "are you a bot?" and then loops the customer through the same menu fails both questions in a single conversation.
Eight checks to run this week
Run these on agents you own or operate, with test accounts and no real customer data. Each check is one conversation. Repeat it on every channel you operate.
| # | Check | How to run it | A pass looks like |
|---|---|---|---|
| 1 | The menu | Open the agent cold. | A "talk to a person" option is visible among the main options from the first screen. |
| 2 | The first message | Type: "I want to talk to a person." | The handoff starts, or the agent says exactly how and when you will be connected. No loop, no new questionnaire first. |
| 3 | Deep in a flow | After four or five turns of normal use, ask again. | Same result. The option still works from inside any flow. |
| 4 | Wording and language | Try "agent", "operator", "human" and "persona", plus each co-official language you serve. | Every variant triggers the handoff, and the reply comes in the customer's language. |
| 5 | The clock | Time 20 requests across the week, including peak hours. | At least 95% reach a person within three minutes of the request. |
| 6 | The person | Let the handoff complete. | The operator introduces themselves at the start. |
| 7 | Supervisor and long waits | Ask for a supervisor. Stay in the queue longer than you would like. | The escalation works inside the same communication. The system never drops you for waiting. |
| 8 | The proof | Ask your team for the last 30 days: requests for a person, handoffs completed and wait time, per channel. | You can produce the numbers within a day, and each number traces back to a conversation. |
On check 5: the law sets the target as 95% of requests attended, on average, in under three minutes. How that average is calculated is a question for your counsel and for the auditor who will review your measurement system.
If check 8 fails, start there. Everything else depends on being able to show it.
Turning the deadline into evidence
The law asks covered companies to document an annual quality evaluation system and to contract an audit of it every year with an entity accredited by ENAC, Spain's national accreditation body (arts. 21 and 22). Smaller covered companies can justify doing both every two years. The audit checks that the company's measurements are reliable, with a maximum error of 5% per parameter, and the company must publish the documentation and the audit on its website (art. 21.4). Its subject is the measurement system, not the behaviour of your agent in each conversation.
Most measurement systems count calls and tickets. Few record what happens inside a bot conversation when a customer asks for a person: how many asked, how many were connected, how long it took and what the agent said in between. In the Lexic AI Agent Trust Index 2026, covering more than 50 production agents and 500 real conversations, around 8 in 10 agents promised an escalation to a person and did not complete it.
Lexic Compass is our independent, continuous audit of AI agents in production. For this law, it matches every request for a person against what happened next, shows the exact turn where a handoff failed, and re-checks when your model, prompt or flow changes. The result is a Trust Score across four pillars and a verdict: APTO, APTO CON CONDICIONES or NO APTO (fit, fit with conditions, not fit).
Compass does not replace the accredited audit in article 22, and Lexic.AI does not certify compliance with any law. It gives you independent evidence from real conversations, so that the numbers you report are numbers you can defend. Independence matters here for a plain reason: the team that builds and runs the agent can report its own handoff rate, and a regulator, a Board or a customer gives more weight to a check made by someone who did not build it.
Your countdown to 28 December
| By | Do | You end up with |
|---|---|---|
| 6 November | List every AI agent and channel in scope. Name an owner for each. | An inventory |
| 20 November | Run the eight checks. Pull a 30-day baseline. | A gap list |
| 4 December | Fix routing, staffing and logging. Train the people who design and run the agent. | Fixes shipped |
| 18 December | Run an independent check on real conversations and file the evidence. | An evidence pack |
| 28 December | Adaptation deadline. |
See how your agent behaves today
Request a Flash Audit on your own agent: a verdict on something already running, with the evidence behind each finding. Request a Flash Audit · See the Trust Index
Sources. Ley 10/2025, de 26 de diciembre, por la que se regulan los servicios de atención a la clientela, BOE-A-2025-26698. Regulation (EU) 2024/1689, Art. 50. Trust Index figure: Lexic AI Agent Trust Index 2026, n = 50+ agents, 500 real conversations, audits up to August 2026. Article numbers and thresholds verified against the BOE text on 23 October 2026.
This guide is general information, not legal advice. Lexic.AI is not a notified body and does not certify compliance with any law. The Trust Score is Lexic.AI's own methodology.