Free resource · Word Kit
AI Roadmap Kit: Deploying Generative AI in 12 Months
All the models to frame, prioritize, roll out, train and manage the deployment of generative AI in a large organization, quarter by quarter.
The AI Roadmap Kit, in brief
- What it is
- A complete kit to frame, prioritize, roll out, train and manage the deployment of generative AI in a large organization, quarter by quarter, over 12 months.
- For who
- Transformation directors, HR directors, IT directors, innovation directors, training managers and project managers mandated by a management committee.
- Format
- 28-page Word document, 5 parts and 8 appendices, all templates ready to fill in.
- What it contains
- Scoping note, minimal governance, express diagnosis, prioritization matrix, four-quarter roadmap, training plan by population, dashboard, month-by-month calendar, use case and pilot sheets, AI committee agenda, launch checklist, FAQ, self-assessment.
- What is it based on?
- Our deployments in large French organizations, the obligations of Article 4 of the AI Act on AI governance, and the 12-month deployment frameworks published by specialist firms.
- Price
- Free, in exchange for a professional email address. No unsolicited emails.
- Update
- Version 2, updated on October 26, 2026: new section on mandatory steps before opening licenses (consultation with the CSE, including pilots, review of access rights, register of AI systems), article 4 of the AI Act updated in the AI Omnibus, Copilot Chat and Microsoft 365 Copilot clearly distinguished, ninth capability "Delegate and supervise".
- What it doesn't replace
- The legal analysis of your contracts, the risk analysis of your CISO and your DPO, nor the social dialogue specific to your organization.
- Go Further
- Adoption Support, AI charter template, AI Ambassadors Kit, Manager's toolkit for dealing with AI, Measurement and ROI kit, AI Act kit from the sponsor.
Most large organizations no longer need to "launch AI": it's already there, in their employees' browsers, often via personal accounts, rarely with training, and almost never with measurement. The issue has changed fundamentally. It's no longer about convincing, but about organizing: defining a scope, selecting a limited number of use cases, proving value through measured pilot projects, training each user group on what applies to them, and then scaling up and industrializing. This is a twelve-month undertaking, and it requires planning.
This page explains why a written roadmap makes a difference, details the five steps and four trimesters we recommend, describes the mistakes that cause first-year failures, and gives you a complete Word-format kit to build your own.
1. Why a roadmap, and not just another launch?
A failed AI deployment rarely fails for technical reasons. It fails because no one has documented why the organization is committing to it, how far it will go, and who decides. Licenses are opened, an awareness campaign is launched, then usage stagnates in two directions, and the executive committee asks what the return is. Without a measurable starting point and written milestones, no one can answer.
A roadmap addresses this issue proactively. It sets the ambition in a single sentence that managers can repeat, defines the scope (target groups, tools, data, exclusions), establishes a streamlined governance structure, and provides the management committee with a milestone to validate each quarter. It transforms an intention into a sustained commitment. And it has a political advantage: an unwritten roadmap is reinterpreted at every meeting; a written and dated roadmap protects the project from shifting moods.
Our deployments in large organizations confirm this: the organizations that survive the first year are those that agreed to start small but noticeably, with three to five use cases in operation from the first quarter, and that measured a pilot before scaling up. Those that tried to train "everyone" on a generic AI in a large lecture hall have, six months later, a disappointing usage rate and a licensing budget that is difficult to justify.
2. The five steps of the roadmap
The kit follows a five-step process. Each step consists of a few pages of methodology, templates to fill out, and a decision area to be signed by the sponsor.
| Stage | What you decide | What the kit gives you |
|---|---|---|
| 1. Frame | The ambition in one sentence, the scope (populations, tools, data, exclusions), five principles, a governance with three roles and two rituals. | Three formulations of ambition, the scope table, the roles table, the ten-question diagnostic questionnaire. |
| 2. Prioritize | The three to five quick wins of the first quarter and the one or two structuring projects to be framed. | The four levels of use cases (prompt, assistant, agent, cowork), the value x ease matrix, the go / no-go criteria, the benchmarks by function. |
| 3. Unroll | The transition from one quarter to the next, validated by the management committee on a written milestone. | The four-quarter roadmap with objectives, actions, deliverables, milestones and common mistakes. |
| 4. Train | Who is being trained, in what, in what format, at what time of year. | The nine capacities, the plan by population, the comparison of formats, the training register expected by article 4 of the AI Act. |
| 5. Pilot and measure | The indicators monitored, the rituals observed, the year-end decisions made. | The dashboard in four families, three methods of measuring value, the agenda of the AI committee, warning signs and their responses. |
Two choices structure everything else. The first is the four levels of use cases: a prompt expert, a configured assistant, an agent connected to the organization's data, and an entire case delegated to AI with human review. In year 1, the majority of widespread use cases fall under the first two levels; agents and delegated cases are introduced in the second half of the year, once the skills, charter, and licenses are in place. The second choice is minimal governance: a sponsor on the steering committee, an AI lead who dedicates one-third to one-half of their time to it, a circle of experts consulted on specific topics (security, personal data, legal, social dialogue), and a network of ambassadors. A one-hour AI committee meeting per month for the first six months, then every six weeks, and a thirty-minute quarterly review are sufficient.
3. The four quarters
The typical roadmap outlines a year for an organization with several hundred to several thousand employees. A regional office or subsidiary can compress the schedule; a multi-country group will expand it. What remains constant is the order: you don't generalize before you've measured, you don't deploy agents before the standard practice is established.
| Quarter | The goal | End-of-quarter milestone |
|---|---|---|
| T1 | Acculturating and framing: diagnosis, framing note, published charter, trained ambassadors, licenses open to volunteers, three to five quick wins in circulation. | Framework note and charter approved, active ambassador network, 3-month barometer. |
| T2 | Pilot and prove: one or two structured projects, measured before and after over six to eight weeks, training of managers. | Numerical assessment of the pilots and written decision to generalize. |
| T3 | Generalize: licenses across the entire scope, mandatory entry path, training in waves of profession, library of 20 to 30 use cases, measurement by department. | Target usage rate achieved over the year. |
| T4 | Industrialize: first agents and files delegated into production, review of the charter, annual review, decisions to stop / maintain / invest. | Annual report presented and roadmap for year 2 approved with its budget. |
The kit details the actions week by week for each quarter, the expected deliverables, and common mistakes, while Appendix A provides a month-by-month calendar with, for each month, the objective, actions, target population, and milestone. This is the page the management committee will follow.
4. The five mistakes that cause first-year failures
Wait until the charter is perfect before opening up the licenses. Employees are already using personal tools; each week of delay prolongs this unregulated use. A short charter published in month 2 and revised in month 10 is better than an exhaustive one published in month 6.
Address all use case ideas. A startup organization quickly receives dozens of ideas. The toolkit asks users to rank them according to their value and ease of implementation, to select three to five quick wins and one or two structuring projects, and to write down why the others are discarded. A discarded idea without explanation reappears three months later in a different form.
Let the pilots linger. A pilot program lasts six to eight weeks, involves a lead, a measured initial indicator, and a quantified target, and concludes with a written decision: to roll it out more broadly, extend it once, or stop. A pilot program without an end date is a project that dares not speak its name.
Train "everyone" in generic AI. The most impactful training programs are short, tailored to specific roles, based on real-world use cases for participants, and conclude with an individual commitment. The kit offers a plan for each target group: management committee, designated contacts, ambassadors, managers, employees by role, agent owners, IT and security personnel—each with its own objective, format, and timing within the year.
Measure fifteen approximate indicators. Three well-measured figures convince a management committee: a usage indicator, a value indicator measured before and after implementation, and a quality or risk indicator. The dashboard of the toolkit is divided into four categories and is populated with data already provided by the tool's administration console and a quarterly barometer.
5. Train each population and remain compliant
Training is the lever that transforms access to tools into real-world use. Rather than training "on ChatGPT" or "on Copilot," the toolkit recommends training in nine transferable skills applicable across tools: searching, writing, summarizing, analyzing, designing, brainstorming, coaching, coding, and delegating and supervising—skills that have become essential with agents. Each group works on a subset of these skills, in a format tailored to their needs: a conference or workshop for the management committee, a full day followed by monthly meetings for ambassadors, a half-day for managers, practical sessions or webinars specific to each role for employees, and a two- to three-day sprint on a real-world case for agent owners.
This approach also fulfills a legal obligation. Since February 2, 2025, Article 4 of the AI Act requires companies that provide or use AI systems to take steps to develop their employees' AI skills, taking into account their role and the context of use. Omnibus Regulation (EU) 2026/1744, in force since July 27, 2026, has made this an obligation of means: it no longer requires guaranteeing a "sufficient level" for each individual, and the Commission and Member States must support these efforts, particularly those of SMEs. Monitoring has been the responsibility of national authorities since August 2, 2026. The regulation does not mandate certification or a standard program. However, it is advisable to be able to answer the question "who has been trained, in what, when, and by whom?" in just a few minutes: the toolkit provides a checklist for the training register to be kept from year 1 onwards, and refers to the AI charter template regarding the rules of use.
6. What the kit contains
The kit is a 28-page, editable Word document with a decision area at the end of each step. It contains, in the order of the process: the user guide and the three rules before starting; the Framing section with the three ambition formulations, the scope table, the five principles, the role table, and the diagnostic questionnaire; the Prioritizing section with the four levels, the value x ease matrix, the go/no-go criteria, and the use case benchmarks by function; the Implementing section with the four-quarter roadmap; the Training section with the nine capabilities, the plan by population, the format comparison, and the register checklist; and the Managing section with the dashboard, measurement methods, the two rituals, and the warning signs. and eight annexes: 12-month calendar, use case sheet, pilot scoping sheet, AI committee agenda, end-of-quarter 1 launch checklist, sponsor's frequently asked questions, self-assessment of maturity rated out of 40, resources to go further.
The kit, part by part: what each part contains and what it is used for. The templates to fill in are in the Word document.
Defining the framework: ambition, scope and governance
State your ambition in one sentence
Ambition is not a slogan: it's a phrase that the executive committee can repeat and that managers can relay. It indicates the desired benefit, the target population, and the timeframe. The toolkit offers three sample formulations that can be adapted (productivity ambition, business ambition, cultural ambition).
Define the perimeter
The scope addresses four questions: which populations, which tools, which data, and which exclusions. Exclusions are as important as inclusions: stating from the outset that personally identifiable customer data or classified documents remain outside the scope of public-facing tools avoids months of debate.
Establish five principles
The principles define the methodology, fit on a single page, and are included in the organization's AI charter. Five principles are common to most successful deployments: human decision-making, data protection, training for all users, measurement of deployments, and sharing of best practices. No deliverable is sent without review.
Establish minimal governance
The governance structure for year one should be streamlined: a sponsor who is a member of the executive committee, an AI lead who dedicates 30 to 50% of their time to it, a circle of experts consulted on specific topics, and one ambassador for every 30 to 50 employees. Two rituals are sufficient: a one-hour AI committee meeting, held monthly for the first six months and then every six weeks, to prioritize the backlog and monitor key performance indicators, and a thirty-minute quarterly review by the management committee to approve the transition to the next quarter.
Establish a rapid diagnosis
The rapid assessment takes one to two weeks: a ten-question questionnaire sent to a sample of employees, supplemented by five to eight thirty-minute interviews with senior managers. The questions focus on actual usage, account type, training received, obstacles encountered, and time saved. Quantifying the gap between usage and training is the most effective argument for securing the mandate and resources for the first year. For a shared assessment within the management committee, the toolkit refers to the AI Maturity Grid, whose profiles correspond to its quarterly phases.
Essential steps before opening licenses
Four points are addressed during the scoping phase, not after: consult the CSE before any deployment, including pilots, and update the single document (several courts have suspended deployments launched without consultation in 2025 and 2026); review access rights to document spaces before opening a connected assistant like Microsoft 365 Copilot; open the register of AI systems; supervise the agents who act, following the ANSSI alert of April 13, 2026.
Prioritize: choose the first use cases
Inventory by distinguishing four levels
The toolkit categorizes use cases into four levels: the expert prompt (a reusable instruction), the configured assistant shared within the team, the agent connected to live sources, and cowork, an entire file entrusted to AI with human review. In year 1, the majority of generalized use cases fall under levels 1 and 2; levels 3 and 4 are introduced in the second semester. A prompt well distributed to 500 people is often more valuable than an agent used by only three.
Sort by value and ease
Each use case receives two scores from 1 to 5. Value combines time saved, the number of people involved, and the impact on quality or risk; ease of use combines data availability, technical simplicity, the absence of legal or social obstacles, and the existence of a motivated champion. The matrix distinguishes between quick wins to be launched in Q1, structuring projects to be defined in Q2, and supplementary and discontinued use cases for Q1.
Choose few and write why
The first quarter yields three to five quick wins and one or two key projects to be defined; rejected cases remain in the backlog with the reasons for the decision. The go/no-go criteria for a pilot project include: an identified business stakeholder, a six- to eight-week testing scope, accessible data without a prior integration project, feedback from the expert panel, a measured initial indicator with a quantifiable target, and a written exit strategy.
Benchmarks of use cases by major function
To begin the inventory process, the kit lists the most frequent use case families in year 1 for seven functions: all business lines, human resources, finance, sales and marketing, purchasing and legal, IT and data, and management. For example: management commentary and variance analysis in finance, responses to calls for tenders in sales, each with its usual starting level.
Unfold: the four-quarter roadmap
Term 1: Acculturation and structuring
Objective: to provide a framework, generate enthusiasm, and achieve initial visible results. The actions unfold over twelve weeks: needs assessment and scoping document, publication of the AI charter, consultation with the Works Council and opening of the AI systems registry, awareness-raising among the management committee, training of ambassadors, review of access rights followed by the issuance of licenses, first wave of volunteer training, and a three-month progress report. A common mistake: waiting until the charter is perfect before issuing licenses.
Term 2: Piloting and proving
Objective: to demonstrate the value of one or two measured, structuring projects and extend the training to managers in the relevant departments. The six- to eight-week pilot projects are structured with a framework, presented to a panel of experts, and conclude with a quantified assessment and a recommendation: to generalize, extend, or discontinue. A pilot project is evaluated using a maximum of three metrics: a usage indicator, a value indicator measured before and after, and a quality or risk indicator.
Term 3: Generalize
Objective: To extend usage to the entire scope of Year 1 and make it standard practice in each department. The quarter combines license deployment with a mandatory onboarding process, training in waves facilitated by industry ambassadors, rollout of successful pilot projects, and ongoing support. Deliverables: a library of 20 to 30 documented use cases and an adoption dashboard for each department.
Quarter 4: Industrializing and preparing for year 2
Objective: to move from individual to collective, tool-based use. Two to four connected agents or coworking spaces are deployed, the AI systems charter and register are revised, tools related to recruitment or employee evaluation receive a compliance plan for the high-risk deadline of December 2, 2027, and the annual report quantifies usage, value, quality, risk, costs, and return on investment. Three year-end decisions are made: discontinue unused services, maintain services that are useful at a constant cost, and invest where value is proven.
Training: the training plan by population
The nine skills to acquire
Rather than training specifically on "ChatGPT" or "Copilot," the toolkit recommends training in nine transferable skills: searching, writing, summarizing, analyzing, designing, brainstorming, coaching, coding, and delegating and supervising (assigning a task or file to an agent, monitoring the results, and knowing when to stop). For each skill, it describes what the employee will be able to do upon completion of the training.
The plan by population
The plan provides for each population an objective, a format, the skills worked on and the time of year: management committee (1.5-hour to half-day workshop, T1), AI referent (2-3 day course, T1 and T2), ambassadors (one day then a monthly meeting), managers (half-day, T2), employees by profession (2-hour to one-day session, T1 for volunteers and T3 for all), agents and coworks bearers (2-3 day sprint, T4), IT, security and data (T1 and T4), and supervisors of sensitive systems (half-day with the DPO and legal, T4, before December 2027).
Choose the formats
The toolkit compares six formats and their limitations: a short session of 1 to 2 hours, a practical webinar of 1.5 to 3 hours (maximum 20 participants, cameras and exercises required), a practical workshop of half a day to a full day (maximum 12 participants), a sprint of 2 to 3 days, an asynchronous learning path, and monthly meetings. The asynchronous learning path provides a common foundation on a large scale, but its completion rate is low if it is not linked to a group session: it serves as a prerequisite, not as the main training.
Documenting for compliance
A simple training register is sufficient in year 1: it should allow you to answer the question "who was trained, in what, when, and by whom?" in just a few minutes. The checklist includes the names of the participants, the content of each session, certificates of completion of the initial training, the date the charter was read, and the year's training plan, revised in quarter 4.
Pilot and measure
The dashboard in four families
The dashboard groups indicators into four categories: adoption (licenses, active users, trained employees), usage (documented cases, shared assistants and agents), value (reported time saved, measured gains on pilot projects, satisfaction), and quality and risk (incidents, deliverable recovery rate). Active users are counted across the entire scope of the project, using the administration consoles (including the Copilot dashboard), and shadow AI, such as unintended agent actions, fall under the risk category. The 12-month benchmark is to start with the values measured during the initial assessment and aim for the next level of the AI Maturity Grid (10%, 30%, then 60% of employees using a provided tool weekly), with a target for time saved per use case measured before and after implementation.
Measure the value without making a mistake
Three complementary methods, from the simplest to the most rigorous: self-reporting within the quarterly barometer, before/after measurement on a sample of tasks from a pilot project (the most convincing for an executive committee), and monitoring a process indicator already tracked by the business. The toolkit advises against converting saved hours into euros too early: the return on investment calculation is included in the annual report, along with the actual costs of licenses, training, and internal time.
The two rituals
The AI committee meets for one hour each month for the first six months, then every six weeks. It comprises the sponsor, the lead contact, two or three relevant managers, and a rotating invited ambassador. The meeting covers period indicators, pilot progress, backlog prioritization, incidents, and decisions. The quarterly management committee review lasts thirty minutes and includes a summary of the four indicator categories, the end-of-quarter milestone, the decision to move to the next quarter, and the resources allocated.
Recognizing warning signs
The toolkit associates seven warning signs with a recommended response: an active user rate that stagnates below 20% after six months, use cases that all originate from the same department, pilot projects that drag on without a decision, persistent unauthorized uses, a sponsor who is no longer present on committees, underutilized paid licenses (which need to be reallocated), and agents that proliferate without an owner. Regarding personal accounts, it primarily recommends verifying that the primary tool meets the needs driving external users.
Annexes
Eight annexes bring together ready-to-use tools: 12-month calendar with objective, actions, trained population and milestone of each month, use case sheet, pilot scoping sheet, AI committee agenda, end-of-quarter 1 launch checklist, sponsor's frequently asked questions, self-assessment of maturity and resources to go further.
Frequently Asked Questions from the Sponsor
Nine answers to questions posed by a client, including two new ones: should Microsoft 365 Copilot be purchased for everyone, and should the Works Council be consulted for a simple pilot program? Regarding the budget, the toolkit identifies three areas: licenses (the most visible), training (the most cost-effective), and internal time (the most overlooked), and observes that successful deployments allocate at least as much to training and support as to licenses in year one.
Self-assessment of maturity
Ten statements, scored from 0 to 4, give a total out of 40, to be repeated at 6 and 12 months. Four readings: from 0 to 10, scattered uses; from 11 to 20, framework established; from 21 to 30, adoption in progress; from 31 to 40, organization equipped, each referring to the quarter of the kit to be prioritized.
Download the AI Roadmap toolkit (Word, free)
28 pages, 5 sections and 8 appendices, all templates ready to fill in. Written by NextStart.AI based on our deployments in large French organizations and the requirements of the AI Act regarding AI governance.
- Scoping note, minimal governance, express diagnosis
- Pilots' Prioritization Matrix and Go/No-Go Criteria
- Four-quarter roadmap and month-by-month calendar
- Training plan by population, dashboard, fact sheets and checklists
7. Frequently Asked Questions
Does it really take twelve months?
Many employees already use AI, but few use it effectively, and almost no one uses it collectively and in a measured way. The twelve months are dedicated to transitioning from scattered individual uses to shared, structured, and tool-based applications. The timeline is compressed into a limited scope, but it doesn't skip the measured pilot phase.
Where do you start if nothing exists?
The quick diagnostic toolkit involves ten questions sent to employees and a few interviews with management. Quantifying actual usage, personal accounts, and the lack of training is the most effective argument for securing a mandate and resources.
Does the kit require a tool?
No. It recommends a standard enterprise-licensed tool, chosen by your organization, that works with Microsoft 365 Copilot, ChatGPT, Claude, Gemini, Mistral, or an internal assistant. The nine capabilities it provides are transferable between tools. In a Microsoft environment, the standard tool is often Copilot Chat, included in Microsoft 365; the Microsoft 365 Copilot license (often called "Copilot Premium") is reserved for users with proven usage.
What does the AI Act say about training?
Since February 2, 2025, Article 4 of the AI Act requires companies that provide or use AI systems to take steps to develop their employees' AI skills, taking into account their role and the context of use. Omnibus Regulation (EU) 2026/1744, in force since July 27, 2026, has made this an obligation of means: it no longer requires guaranteeing a "sufficient level" for each individual, and the Commission and Member States must support these efforts, particularly those of SMEs. Monitoring has been the responsibility of national authorities since August 2, 2026. The regulation does not mandate certification or a standard program; the toolkit offers a training register to document who has been trained, on what, and when.
How can we measure the value without making a mistake?
The toolkit offers three methods, ranging from the simplest to the most rigorous: self-reporting within a quarterly barometer, before-and-after measurements on a pilot project, and monitoring a process indicator already used by the business. It advises against converting the hours saved into euros too soon.
Can the kit be used without assistance?
Yes. All the templates are in the Word document and can be used as is. We get involved when the organization wants an external assessment, a scoping workshop with its management committee, or the development of its training plan.
8. Going further
The kit structures the approach; it doesn't carry it out for you. NextStart.AI supports large organizations at each stage: diagnosis and adoption roadmap with the management committee, practical training by profession (Prompt Party), management track Managers Catalyst, training of AI ambassadors, and Agent Party, Cowork Party or Challenge for the second half of the year.
Read also : AI Maturity Grid: Assess your organization's readiness in 90 minutes, AI Charter in Business: The 2026 Model (Downloadable), AI Ambassador Network: The 2026 Toolkit, Managing a team in the face of AI: the 2026 toolkit et Measuring the uses, value and ROI of AI: the 2026 toolkit.
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, AI maturity grid, AI Measurement and ROI Kit, AI Governance Kit, Cowork kit, 50 Copilot M365 prompts, Copilot Studio kit, Deploying Claude in the Workplace Kit, Claude Agents Kit, Claude Skills kit, 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 et communicator's AI kit.
9. sources
The regulatory references cited on this page are based on the following sources.
- European Union, Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (AI Act), article 4, Official Journal of 12 July 2024.
- European Union, Regulation (EU) 2026/1744 (omnibus IA), adopted on 8 July and published on 24 July 2026amending in particular Regulation (EU) 2024/1689.
- European Commission, AI Literacy, Questions & Answers, updated on 27 July 2026 (application of article 4 since 2 February 2025, control by national authorities from 2 August 2026, no mandatory certification).
- CNIL, Entry into force of the European regulation on AI: the CNIL's first questions and answersUpdated on August 17, 2026.