Free resource in French · Word + Excel Kit
AI Customer Service Kit: 24 prompts to progress through 4 levels
The best everyday use cases for customer service and support, from handling sensitive issues to managing problems, with ready-to-copy prompts, seven Excel tools, and 300 practice tickets. A step-by-step, self-directed learning path using tools your organization has already validated.
Conceived by Flavien Besset, founder of NextStart.AI · updated November 9, 2026 · 5,350 professionals trained · satisfaction 9.4/10
The AI Customer Service Kit, in brief
- What it is
- A self-sufficient skills development kit for customer service and after-sales service teams: 24 everyday use cases, each with its prompt, organized into four levels of mastery, and a workbook of seven tools.
- For who
- Advisors, after-sales service technicians, supervisors, quality managers, knowledge base managers and customer relationship managers of large private or public organizations, and managers who want to lead it as a team.
- Format
- A 58-page Word document: user guide, 2026 figures, fundamentals, usage framework, 24 fact sheets, 4 progress challenges, 4 current affairs files, a section for facilitating the kit in teams, and 5 appendices. An Excel workbook with 7 tools, including a set of 300 mock tickets.
- What it contains
- Copy prompts and follow-up prompts, instructions from six assistants (Projects, Gems, Copilot agents), missions of six agents and what they never do, checkpoints, realistic gains, the three fires for customer data, escalation matrix, quality grid, catalog of standard responses, and twelve questions to ask before putting an AI in front of customers.
- Tools
- Copilot Chat and Microsoft 365 Copilot (often called "Copilot Premium"), ChatGPT, Claude, Gemini and Mistral, in their enterprise versions, or your organization's internal assistant.
- What is it based on?
- Gartner's 2025 and 2026 customer service surveys, Ipsos bva's 2025 Customer Service Observatory, the AI Act and the European Commission's Article 50 guidelines, CNIL recommendations, publisher prompting guides, and AI training sessions conducted with customer relationship teams from large organizations.
- Price
- Free, in exchange for a professional email address. No unsolicited emails.
- Update
- Version 1 of November 9, 2026.
- What it doesn't replace
- Your organization's security policy and AI charter, your business policy, the advice of your lawyers and your DPO, nor the practice supported on your own cases.
- Go Further
- AI training for customer service and after-sales support, GDPR kit, AI charter template.
Customer service teams have often had access to Copilot or ChatGPT for months. They use it to rephrase emails, rarely to respond based on terms and conditions, analyze reasons for contact, or maintain the knowledge base. This isn't due to a lack of willingness: no one has told them where to start, what customer data they are allowed to share with the tool, or how to track their progress.
This page presents what recent studies say about AI in customer service, the four-level progression we use in training, the 24 selected use cases, and a sample worksheet. The complete kit, in Word format with its accompanying Excel workbook, can be downloaded for free further down this page.
1. AI in customer service: what the 2026 figures show
According to a Gartner survey conducted in October 2025 among 321 customer service leaders, and published on February 18, 2026, 91% of them are under pressure from their management to implement AI in 2026. Budgets are following suit: another Gartner survey, conducted from April to May 2026, measures a 38% increase in customer service AI spending, while the overall budget for the function is increasing by only 2%.
Customers, for their part, set their own conditions. In a Gartner survey conducted in February and March 2026 with 3,566 customers, 87% considered it essential to be able to contact a human advisor when the company uses generative AI, and only 27% would try a chatbot again after a bad experience. In France, the Ipsos bva 2025 Customer Service Observatory found that 72% of respondents would feel misled if the company did not clearly state from the outset that they were communicating with an AI, and that the chatbot remains the least satisfactory channel.
Three consequences for a customer service team: start by assisting advisors before automating customer interactions; learn from your own cases rather than generic examples; and set a clear framework from the outset, because customer data is among the most sensitive in the company and Article 50 of the AI Act requires, since August 2, 2026, that customers be informed that they are interacting with an AI.
2. Four levels, saving time by delegating
The lists of customer service prompts published online are almost all catalogs, categorized by request type, without any progression. They're used for browsing, rarely for learning. The kit organizes use cases into four levels, each with its own learning objectives and a challenge to verify mastery of the level.
| Level | What you do | What you learn | Example of a record |
|---|---|---|---|
| 1. Save time | Short prompts, without identifying customer data, a result in five minutes. | Formulate a request in four parts, anonymize, impose a format, identify an invention. | Finding the right words to say to an angry customer |
| 2. Structure | Comprehensive prompts, on your documents: general terms and conditions, complaints, ticket exports. | Quote the article, separate the fact from the hypothesis, write "unknown", check a calculation. | Analyze an export of 300 tickets: reasons, pain points, causes |
| 3. Customize and improve reliability | Permanent assistants: Project, Gem, Copilot agent, with your tone, your standard answers and your escalation matrix. | To organize the context once and for all, to test an assistant on tricky cases. | The knowledge base assistant that cites its sources |
| 4. Delegate | Planned agents and missions Cowork, with safeguards and review. | Assign recurring tasks, define what the agent never does, review outputs. | The high-risk claims monitoring agent |
Level 3 is the one most teams skip, going directly from ad-hoc use to agents. Yet it's this level that makes AI reliable: an assistant who has your terms and conditions doesn't invent them, and an assistant who has your escalation matrix doesn't promise what shouldn't be promised.
3. The 24 use cases, family by family
Each row represents a family of situations, each column a level. You progress through the columns; you select from the rows, depending on the week's difficulty. Each level covers the entire profession: no need to wait until the end of the kit to tackle the real issues.
The details of each case follow: the situation, the realistic payoff, the point to check before using the result, and what you obtain. The complete prompt, or the assistant's instructions and the agent's mission, is in the kit.
Win time
Rephrase an answer that is too technical or too dry
- You've drafted a response that's correct in substance, but it's riddled with internal jargon ("your case is in RTI status," "the package is experiencing a hub anomaly") or it sounds curt. The customer is already annoyed; a poorly worded response will only reignite the conversation. You want a clear, polite, and friendly version, without changing the content.
- Savings: 5 to 10 minutes per tricky answer.
- Please note: Every fact, timeframe, and amount is identical to your original version.
Deliverable: Your response rewritten in simple French, in the tone of your company, with technical terms explained or replaced, and a list of what has been changed.
Finding the right words to say to an angry customer
- A client sends you a very aggressive message: all caps, accusations, threats to leave the company or post a review. You know that your first response will determine what happens next. You want to take a step back, understand what they really expect, and find the right opening lines, without apologizing for a mistake that hasn't been made.
- Savings: 10 to 15 minutes per difficult message.
- To be verified: The opening acknowledges the inconvenience without acknowledging a responsibility that your company has not established.
Deliverable: A three-point reading of the message (what happened according to the client, the dominant emotion, what he really wants), then two proposed ways to open a response, one more formal, the other more personal.
Transforming an internal procedure into a step-by-step guide for the client
- Your internal procedure for resetting equipment, activating a warranty, or returning a product is written for advisors: abbreviations, screen references, internal steps. Customers call for procedures they could perform themselves if they had clear instructions. You want a customer version to send by email or publish in the FAQ.
- Savings: 15 to 20 minutes per procedure.
- To verify: Do the step-by-step yourself, or have it tested by a colleague who is not familiar with the procedure.
Deliverable: A numbered step-by-step guide for the client, with a clear title, prerequisites, steps and what to do if it doesn't work, without any internal information.
Classify about twenty verbatim comments by theme
- You have about twenty recent customer comments in front of you: responses to a satisfaction survey, online reviews, end-of-call notes. Your manager asks you which ones come up most often. Reading and sorting them manually takes an hour. You want an initial ranking to review.
- Profit: 20 to 30 minutes per batch.
- To check: Randomly check five rankings: the pattern matches the text.
Deliverable: A table that sorts each verbatim into a pattern, with the sentiment (positive, neutral, negative), then a count per pattern and the three main lessons.
Prepare a supportive debriefing with an advisor
- You are a supervisor or quality assurance expert. You listened to a call or reviewed a written exchange that went badly: the advisor interrupted, provided incomplete information, or lost patience. You want to prepare feedback that fosters improvement without demotivating the team.
- Gain: 10 minutes per debriefing.
- To verify: The areas for improvement describe facts that you have actually observed.
Deliverable: A fifteen-minute debriefing plan: one strength to highlight, two areas for improvement formulated as observable behaviors, three open questions to make the advisor think, and a commitment to set together.
Write the weekly update for the team
- Every Monday, you send a message to the team about the past week: volumes, deadlines, satisfaction, key achievements. You want it to be read, to explain the figures simply, to recognize the team's contributions, and to set a clear priority for the week ahead.
- Savings: 20 minutes per week.
- To verify: Each figure is identical to your dashboard.
Deliverable: A short message, structured in three parts (what happened, what we remember, the priority of the week), with the indicators explained in one sentence each.
Structure
Responding to a complex request based on your conditions
- A customer asks a specific question about their rights: warranty coverage, return shipping costs, the cancellation period, termination, and exchange conditions. The answer lies in your general terms and conditions or warranty policy, but you need to find the correct article, apply it to their situation, and explain it without legal jargon.
- Savings: 15 to 25 minutes per complex request.
- To verify: The cited article exists and says exactly what the note claims: open the document.
Deliverable: A proposed response that cites the applicable article, explains what it means for this client, lists the missing information to conclude, and an internal note that separates what is certain from what remains to be verified.
Handling a claim from start to finish
- A written complaint arrives, detailing several exchanges: defective product, failed intervention, disputed billing. The customer is exasperated and has already contacted the company twice. You must reconstruct the events, identify what was promised, decide on the next steps with your manager, and provide a comprehensive response.
- Savings: 30 to 45 minutes per claim.
- To verify: The chronology is correct: check the dates in the ticket tool.
Deliverable: A timeline of events, a list of commitments already made by the company, the gaps between what was promised and what was done, the treatment options provided for by your policy, and a draft response to be completed after a decision.
Write a knowledge base article based on a solved case
- You've just solved a case that no one else knew how to handle: a rare error message, an unusual administrative procedure, an incompatibility between two products. If nothing is documented, the next advisor will search for it again. You want to transform this exchange into a knowledge base article, reusable by the team and, later, by an AI assistant.
- Savings: 20 to 30 minutes per article.
- To be verified: The solution actually worked, in this order.
Deliverable: A structured article in the format of your database (symptom, cause, solution, special cases, search keywords), with a client version if the content can be published.
Analyze an export of 300 tickets: reasons, pain points, causes
- Each month, it's necessary to explain why customers contact the service: which reasons are increasing, which pain points are recurring, and what could be prevented. You have access to a ticket export with the reason entered by the advisors, the channel, the response time, and sometimes a free-text comment. The reasons entered are imperfect.
- Savings: 2 to 3 hours per monthly analysis.
- To verify: The total number of tickets analyzed corresponds to your export.
Deliverable: A four-part analysis: distribution of contacts by reason and by channel, recurring irritants in comments, avoidable contacts (missing information, follow-up, internal error) and five proposed actions, each supported by figures.
Build a quality assessment grid tailored to your channels
- Your quality assessment grid is from the era when everything was done by phone. It doesn't address chat, social media, or AI-generated responses. You want a grid for each channel, with observable and weighted criteria, that supervisors apply consistently.
- Savings: 2 hours to build the grid, then 5 minutes per evaluation.
- To check: Test the grid with two supervisors on the same five exchanges: the scores should converge.
Deliverable: A grid of 8 to 12 criteria per channel, grouped into three sections (understand, respond, engage), with each criterion including a definition, observable elements, a three-level rating scale, and a weighting. This grid is located in the "Quality Grid" tab of the workbook.
Comment on the monthly dashboard and prepare the committee
- Each month, you present customer service results to management: volumes, response times, first-contact resolution, satisfaction, complaints, and costs. The figures are there, but you need to explain them, link variations to causes, and propose solutions—all in one page.
- Savings: 1 hour per committee.
- To verify: Each figure in the score appears in the dashboard.
Deliverable: A one-page note: the three key messages, the commented indicators (evolution, probable cause, level of certainty), the risks, and two or three decisions to ask the committee.
Customize and make reliable
The home tone response assistant
- Each advisor has to explain the company's tone, phrases to avoid, terms and conditions, and standard responses to the AI. The result: inconsistent answers from one advisor to another, and the same amount of time wasted on every conversation. You want a team assistant who already knows all of this.
- Reward: 10 minutes saved per written response.
- To check: Test the assistant on five tricky cases before sharing it: a late refund request, a customer demanding a gesture, an off-source question, an aggressive message, a message containing an IBAN.
Deliverable: A shared assistant that prepares written responses in accordance with your tone guidelines, based on your validated standard responses and conditions, and that signals what it does not know.
The complaints and escalation assistant
- Sensitive complaints are handled differently by each advisor: some are escalated too late, others too early. The escalation matrix exists, but no one reviews it during peak periods. You want an assistant who categorizes each complaint according to your rules and prepares the processing.
- Gain: 20 minutes per claim, and fewer missed climbs.
- To be verified: The rule mentioned exists in the matrix and applies correctly.
Deliverable: For each attached claim: the qualification (reason, severity, risk signals), the escalation level provided by your matrix with the rule applied, the documents to be gathered and a draft interim response.
The knowledge base assistant that cites its sources
- The answer exists somewhere: in a database article, a memo, a customer service procedure, or a product meeting summary. Advisors, especially new ones, take time to find it, or they call a colleague. You want an assistant who searches your validated content and answers by citing the article, without ever making things up.
- Savings: 5 to 10 minutes per difficult search.
- To verify: Open the cited article the first few times, until the assistant has proven itself.
Deliverable: An internal assistant who answers advisors' questions from the knowledge base, cites the article and its date, says "I can't find it" when that is the case, and notes unanswered questions to enrich the base.
The verbatim analysis assistant with your reasons
- Each analysis of verbatim comments starts from scratch: the AI invents different categories each time, and the months are not comparable. You want an assistant that always applies your official taxonomy of patterns and your methodology, to track trends over time.
- Earnings: Half a day per month.
- To verify: Check ten rankings at random each month.
Deliverable: An assistant that categorizes verbatim comments into your official motives and sub-motives, measures sentiment, identifies new irritants that do not fit into any category, and always produces the same summary format.
The difficult customer simulator for training advisors
- New advisors learn about difficult situations firsthand, with real clients. Role-playing exercises between colleagues are time-consuming and lack variety. You want a simulator where each advisor practices individually, as much as they like, and then receives feedback based on your quality assessment criteria.
- Benefit: 1 hour of training per new arrival per week.
- To check: Have the scenarios run by two supervisors before opening them up to the team.
Deliverable: An assistant who plays a customer (dissatisfied, confused, in a hurry, insisting on a gesture) on realistic scenarios of your business, adapts to the advisor's responses, then gives structured feedback according to your grid.
The indicators assistant: your definitions, your thresholds, your alerts
- Each manager interprets the metrics differently: response times aren't calculated the same way for email and chat, and alert thresholds vary depending on who's presenting. You want an assistant who understands your official definitions and thresholds, and who always interprets the dashboard in the same way.
- Savings: 2 hours per month.
- To be verified: The thresholds applied are those of the dictionary.
Deliverable: An assistant that comments on a dashboard according to your definitions, signals indicators out of threshold, suggests possible explanations classified by level of certainty and prepares questions to ask the teams.
Delegate
The agent drafts answers in the shared inbox
- The shared inbox receives hundreds of simple, repetitive requests every day: order tracking, address changes, invoice requests, and service appointment scheduling. The advisors almost always type the same response. You want an agent to prepare a draft for each simple request, which the advisor can then review, adjust, and send.
- Savings: 1 to 2 hours per advisor per day on simple requests.
- Please note: Each draft is reviewed and sent by an advisor.
Deliverable: For each incoming email classified as simple: a draft response is generated in the tool, based on the corresponding standard response, along with the detected reason and a confidence level. Complex or sensitive requests do not receive a draft, but a label.
The high-risk claims monitoring agent
- The riskiest complaints (formal notices, customers announcing legal action, claims of damages, mentions of media outlets or consumer associations, vulnerable customers) sometimes get lost in the shuffle and are only noticed after the third follow-up. You want an agent who monitors incoming messages and alerts the appropriate supervisor as soon as a red flag appears.
- Benefit: Climbs spotted on the same day instead of several days later.
- To be verified: The supervisor confirms or denies each alert in the ticket.
Deliverable: An alert sent to the on-call supervisor as soon as a message contains a signal from the escalation matrix, with the citation, the rule concerned and the expected processing time, and a daily summary of alerts.
The agent who maintains the knowledge base
- The knowledge base is aging: outdated articles, frequently asked questions without corresponding articles, contradictions between two versions. No one has the time to review it. Yet it's what feeds the advisors and, tomorrow, the AI assistants.
- Benefit: A database that grows richer every week without additional meetings.
- To verify: The database owner validates, corrects, or rejects each project.
Deliverable: Every Monday, a report for the database owner: questions asked the previous week without a corresponding article, articles that have not been reviewed since the deadline, contradictions identified, and draft articles to be validated.
The weekly Voice of the Customer summary agent
- Verbatim feedback accumulates from multiple sources: satisfaction surveys, online reviews, and end-of-ticket comments. Monthly analysis arrives too late to react to new issues, such as a defect in a batch of products or a bug in the application. You want an automated summary sent weekly to the relevant teams.
- Benefit: Effortless weekly summaries, irritants identified in a few days.
- To be verified: The person in charge reviews each summary before distribution and checks at least two figures.
Deliverable: Every Friday, a one-page summary: volume of verbatim comments, breakdown by pattern according to your taxonomy, irritants on the rise compared to the previous four weeks, new irritants, five representative verbatim comments, and the signals to be transmitted to the product, logistics or digital teams.
Our mission Cowork : preparing the integration process for a new advisor
- A new group of advisors arrives in three weeks. The onboarding process needs to be prepared: a schedule for the first two weeks, quizzes on products and procedures, practical exercises, simulation scenarios, and fact sheets. The in-house trainer is doing all this manually, using scattered documents.
- Gain: 2 to 3 days of preparation saved per promotion of new advisors.
- To be verified: The internal trainer and a supervisor review everything, test the quizzes and play two scenarios before the first session.
Deliverable: A complete integration report produced by a project assigned to Claude Cowork or Copilot Cowork : day-by-day planning, a twenty-question quiz per module with answers, ten fictional case studies, five scenarios for the simulator in sheet 3.5, and a memo sheet per key procedure.
The monthly reporting and alert agent for indicators
- The monthly report takes half a day to compile: extracting data from three tools, updating charts, and writing commentary. And between reports, a trend (a sharp increase in response time on a channel, a rise in follow-up calls) can go unnoticed for weeks. You want an agent to compile the report and monitor the thresholds weekly.
- Benefit: Half a day per month and timely alerts.
- To be checked: The person in charge reviews the project, checks three figures in the sources, adjusts the comments and distributes it.
Deliverable: On the first working day of the month, a draft of a complete monthly report, annotated according to your indicator dictionary (assistant in sheet 3.6). Every Monday, an alert if an indicator crosses its threshold.
4. An example of a worksheet: finding the right words to say to an angry customer
All the forms follow the same structure: the situation, what you get, the prompt to copy, a follow-up prompt, the checkpoints, the realistic gain, and the safeguard. Here is the prompt for form 1.2, to be pasted as is into your assistant, replacing the bracketed passage with the client's anonymized message.
You are in charge of a customer service department, trained to handle dissatisfied customers. Here is a message from a customer, anonymized:
[PASTE THE MESSAGE]
1. Summarize in three lines: the facts as the client describes them, the dominant emotion, what he is really asking for (explicitly or implicitly).
2. Offer two possible responses of no more than four sentences each: one formal, one more personal. They should acknowledge the emotion and discomfort, rephrase the request, and announce the next step.
Rules: Do not acknowledge any fault on the part of the company, do not promise any compensation, do not contradict the customer. No stock phrases such as "we understand your frustration".
Before using the outcome: the opening acknowledges the inconvenience without assigning responsibility that your company hasn't established; no wording sounds scripted; the announced next step is realistic and time-bound. If the message mentions a lawyer, mediator, or danger, the escalation matrix takes precedence over the response.
5. The three non-negotiable rules
AI prepares, advisor responds. No AI-generated response is sent to a client without human review. No goodwill gesture, refund, extension, or commitment is included unless it is stipulated in your policy or approved by an authorized individual.
Customer data remains in the validated tool. Names, contact details, contract numbers, banking or health data never go into a consumer-grade tool. The kit categorizes customer data into three groups and specifies what can be entrusted to which tool, after replacing identifying data with labels.
The customer knows, and can contact a human. When an AI communicates directly with a customer, the customer must be aware of it: this is mandated by Section 50 of the AI Act. And the customer must be able to contact a human being when the situation requires it. The kit does not create any system that isolates the customer from a machine.
6. What the kit contains
Version 1, published on November 9, 2026. The kit is a 58-page Word document. It opens with a user guide (the matrix of 24 fact sheets, the one-hour-per-week pace, where to begin), the 2026 figures, the fundamentals (the four-building block method, sources, the five typical mistakes, tone), and the usage framework (the three fires for customer data, the GDPR and the CNIL's advice on chatbots, Articles 4 and 50 of the AI Act). Next come the four levels, each with its own introduction, six fact sheets, and a completion challenge. Four current affairs dossiers for 2026-2027 address the obligation to disclose one's AI, the right to speak to a human, customers arriving with their own AI, and what autonomous agents truly change for advisors. One section is for the manager who wants to facilitate the kit as a team, and five appendices complete the package. The Excel workbook includes seven tools: self-assessment, a library of 24 prompts, a set of 300 mock tickets with a calculated summary, a quality assessment grid, an escalation matrix, a catalog of standard responses, and a time-save measurement tool.
Download the AI Customer Service Kit (Word and Excel, free)
58 pages, 24 fact sheets across 4 levels, 4 progress challenges, 4 current affairs dossiers, and a workbook of 7 tools with 300 practice tickets. Written by NextStart.AI based on Gartner and Ipsos bva surveys, the AI Act, CNIL recommendations, and AI training sessions conducted with customer service teams from large organizations.
- Fundamentals of prompt customer service and a three-pronged usage framework
- 24 fact sheets: responding, sensitive situations, knowledge base, voice of the customer, quality, management
- Six assistants and six agents ready to configure, with their safeguards.
- Escalation matrix, quality grid, standard responses and 300 tickets in Excel
7. Frequently Asked Questions
Do we already need to use AI to operate the kit?
No. The self-assessment grid indicates where to start, and level 1 begins with short prompts, without identifying customer data. An advisor who is already comfortable starts at level 2 or 3, after completing the challenge of passing the previous level to verify their progress.
Can we paste messages or customer tickets into the tool?
Only in the tool validated by your organization for this purpose, in the enterprise version, after replacing identifying data with labels. Banking data, health information, or litigation records are never included. The kit framework categorizes customer data into three sections, and the workbook provides 300 practice tickets.
Can AI answer customer questions on its own?
Not in this kit. The AI prepares, the advisor reviews, decides, and sends. No document promises a gesture, a refund, or a delay that your policy doesn't stipulate. Putting AI directly in front of customers is a different project, with transparency obligations (AI Act, section 50) and guaranteed access to a human: Appendix D provides the checklist.
What tools does the kit work with?
Copilot Chat and Microsoft 365 Copilot (often called "Copilot Premium"), ChatGPT, Claude, Gemini, and Mistral, in their enterprise versions, or your organization's internal assistant. The prompts are copied as is; the level 3 and 4 guides describe the specific setup for each tool.
Is the kit suitable for after-sales service teams and outsourced contact centers?
Yes. The guidelines cover technical and warranty requests, and the usage framework applies to anyone who handles your customers. When using a service provider, verify that their tools and contract adhere to the same data rules.
Can the kit be distributed to the entire team?
Yes, freely within your organization, in its original form; the prompts and the workbook can be adapted without restriction.
8. Going further
The kit is sufficient on its own for a motivated advisor. It's faster and goes further when the team practices it together on their own situations: that's the purpose of the AI training for customer service and after-sales support from NextStart.AI, in 1.5 hours or half a day, prepared with your customer relationship manager and conducted using tools validated by your organization. To define the uses of customer data, the GDPR kit extends the chapter on the framework, and the AI and Social Dialogue Kit helps to involve employee representatives when the organization of work changes.
The entire series of free kits. To drive adoption: AI charter template, AI training plan template, AI Ambassadors Kit, Manager's toolkit for dealing with AI, AI-powered time-saving kit, Adoption Roadmap Kit, AI maturity grid, AI Measurement and ROI Kit, AI Governance Kit, Cowork kit, 50 Copilot M365 prompts, Copilot Studio kit, Deploy Microsoft 365 Copilot kit, Copilot Chat Kit, Copilot Excel kit, Copilot Teams and Outlook kit, Copilot Word and PowerPoint kit, Configure Claude kit, Deploying Claude in the Workplace Kit, Claude Agents Kit, Claude Skills kit, Claude's Marketing Skills, Skills Claude HR, Claude Plugins Kit, Claude Code kit, Claude Design kit et Vibe coding kitFor the AI project sponsor: AI Act kit, GDPR kit, Security and data kit, Shadow AI kit, AI and Social Dialogue Kit, frugal AI kit, AI Sovereignty Kit et responsible AI kitBy profession: AI HR kit, AI marketing kit, AI finance kit, AI kit for lawyers, AI kit for salespeople, AI toolkit for project managers, AI innovator kit, communicator's AI kit, AI Procurement Kit, AI Kit for Researchers et AI kit Customer service.
9. sources
The figures and references cited on this page and in the kit are based on the following sources, verified on November 9, 2026.
- Gartner, Gartner Survey Finds 91% of Customer Service Leaders Under Pressure to Implement AI in 2026, February 18, 2026.
- Gartner, Gartner Survey Finds 85% of Service and Support Leaders are Expanding Human Agent Responsibilities Despite Expectations of Mass AI Layoffs, April 28, 2026.
- Gartner, Gartner Survey Finds Customers Are 3x More Likely to Use Third-Party GenAI Than Company-Provided Chatbots for Customer Service, July 8, 2026.
- Gartner, Gartner Survey Finds 87% of Customers Say Companies Using GenAI for Customer Service Must Provide Access to a Human Agent, 4 2026 August.
- Gartner, Gartner Survey Finds AI Spending by Customer Service Leaders Has Surged by 38%, Despite Overall Service and Support Function Budgets Rising by Just 2%, 26 2026 August.
- Gartner, Gartner Survey Finds Only 27% of Customers Would Try a Chatbot Again After a Negative Experience, September 2, 2026.
- Gartner, Gartner Survey Finds 21% of Service and Support Leaders Have Reduced Human Agent Headcount Because of AI, November 5 2026.
- Gartner, Gartner Predicts Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues Without Human Intervention by 2029, March 5, 2025.
- Ipsos bva and Voted Customer Service of the Year, Customer Service Observatory 2025: Humans (still) at the heart of customer relations in the age of AI, November 21 2025.
- Salesforce, State of Service: AI agents and the future of customer service, 2025-2026.
- European Commission, Commission publishes guidelines on transparency obligations for providers and deployers of certain AI systems (IP/26/1653), July 20, 2026.
- Stibbe, The AI Act's Transparency Obligations: Rules, Scope and Timeline, 2026.
- Hunton Andrews Kurth, EU Digital Omnibus on AI enters into force (Regulation (EU) 2026/1744), July 2026.
- KI-Campus, EU AI Act: AI Omnibus recalibrates Article 4, AI literacy remains mandatory, July 28, 2026.
- National Consumer Institute, Financial services subscribed to remotely: what protection for the consumer?, 2026.
- France Savings, Insurance, credit, online savings: what changes for subscribers on June 19, 2026, 2026.
- Selectra, What Spanish law requires of customer service departments by the end of December 2026, 2026.
- CMS, Inside the EU's planned Digital Fairness Act: digital design, subscription traps, and influencer marketing, October 2026.
- CNIL, Chatbots: CNIL's advice on respecting people's rights, February 19, 2021.
- CNIL, How to deploy generative AI: the CNIL provides initial clarifications, July 18, 2024.
- Microsoft, Enterprise data protection in Microsoft 365 Copilot and Microsoft 365 Copilot Chat, 2026.
- Anthropic, Prompt engineering overview, online.
- Google Gemini for Google Workspace Prompting Guide 101, 2024.