AI transformation roadmap: what a good one contains
A good AI transformation roadmap is a decision document, not a presentation. It holds the business levers, the use cases in ranked order, a card for each one with preconditions, its AI Act classification and its success criteria, what not to build, and the first project, ready to start.
Most AI roadmaps we see in companies are presentations: a vision, a list of tools and a diagram with three phases. They look good in the board meeting, and nobody can execute them the next morning. A good AI transformation roadmap is something else: a decision document that the first project comes straight out of.
What is an AI transformation roadmap?
A written document that says which AI use cases are worth doing in your company, in what order, what has to be true before each one, how success is measured, and what is not worth building. By "transformation" we mean something practical: one process at a time, taken into production, measured, then the next.
The companies that deploy AI at customers write just as concretely. Anthropic describes the role that builds the plan as someone who goes from "signed SOW to a sequenced roadmap with clear milestones, dependencies, success criteria, and value hypotheses" (Anthropic posting). Scale AI says: "We work with them to understand the biggest levers for their business" (Scale AI posting).
What should an AI roadmap for a company contain?
- The business levers. The cost or revenue lines you want to move: hours lost in a process, response time to customers, errors that cost money. Not "we want a chatbot".
- The processes as they really run. Who does the work, on which systems, with which data, and where the time goes, learned from the people who do it.
- The list of use cases, ranked, with the reason for each position.
- A card for each use case: the problem, the users, the systems and data it touches, the baseline, the production milestone, the adoption metric, the preconditions, the AI Act classification, a rough budget, and the expected return, written as a number you check in your own pilot.
- What not to build, and why.
- The order and the dependencies: what has to happen before what, from data to access.
- Who owns what: the sponsor, the process owner and the internal engineer for each use case.
- The first project, ready to start: its scope, its success criteria and what the company provides.
How are use cases ranked?
With a few simple criteria, scored 1 to 5, the same for every use case. The order comes from the scores, not from one department's enthusiasm.
| Criterion | The question | What a 5 means |
|---|---|---|
| Value | Which cost or revenue line does it move? | A large, measurable process with an owner who wants the change |
| Feasibility on your data | Does the data exist, and can it be reached? | Real data, checked on an extract, reachable through existing systems |
| Volume | How often does it happen? | Every day, in high volume |
| Integration effort | How many systems does it write into? | One, with documented access |
| Risk | What do GDPR and the AI Act say? | Minimal risk, no automated decisions about people |
| Adoption | Will people use it? | The users asked for help exactly there |
The first use case is not necessarily the most valuable one. It is the one with clear value and the fewest preconditions, because the first project has to reach production and be used. The criteria extend the score in how to choose your first AI use case with integration effort and adoption, which matter once you compare use cases across the company.
Which preconditions belong in it?
- Data: where it lives, who owns it, how complete it is, checked on a real extract rather than a description.
- Integration: through which API or export the system reaches the data, and where it writes its result.
- Process: is it stable, or about to be reorganised?
- People: who uses it, who checks it, what training is needed.
- Legal: a GDPR impact assessment where needed, and the AI Act classification.
Preconditions are why projects stall. In BCG's survey of 1,250 executives and senior AI decision-makers, 72% named integrating AI with existing systems, tools and APIs as a challenge (BCG, September 2025; BCG sells AI consulting). And among EU companies that considered AI in 2025 and did not go ahead, 70.3% named a lack of expertise (Eurostat, KS-01-26-009). A good roadmap puts them on paper before they cost money.
How do you classify each use case under the AI Act?
| Category | Examples | What it means for the roadmap |
|---|---|---|
| Prohibited (Art. 5) | Social scoring, manipulation | Not in the roadmap. Prohibitions apply since 2 February 2025 |
| High-risk (Annex III) | CV screening in recruitment, creditworthiness of individuals, life and health insurance pricing | Deployer duties from 2 December 2027; rarely the first use case. Sapio does not build credit scoring or insurance pricing systems |
| Transparency (Art. 50) | A chatbot or agent that talks to customers | People must know they are dealing with an AI system, from 2 August 2026 |
| Minimal risk | Extracting data from invoices, searching internal documents | No specific duties beyond GDPR and AI literacy (Art. 4) |
The classification is made on the specific use case, not on the technology: the same model can be minimal risk in one process and high-risk in another. The dates are those of Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744. The details are in how to prepare for the EU AI Act. This is general information, not legal advice.
What should not be in it?
- Tools as goals. "Roll out Copilot" is not a use case.
- ROI promises. The return is estimated, then measured in your own pilot, on your data.
- Twenty use cases started at once. One, then the next.
- Use cases with no owner in the company. Without an owner, there is no adoption.
How does the roadmap turn into the first engagement?
The first use case card becomes the engagement charter: the production milestone, the adoption metric and the baseline are copied from the roadmap and signed at kickoff. Then a senior engineer works inside your team about three days a week, on that use case, until it runs in production and is used, with one of your engineers alongside. How those 90 days run is in AI implementation engagement model, and the role in what is an embedded AI engineer.
How does Sapio build an AI transformation roadmap?
We start from the Tech Audit: four hours on your site and a short report with 3–5 opportunities. For companies that want a plan across several use cases, the Deep Dive follows: two visits on site, research on your sector, a written roadmap for each use case, with the preconditions, the AI Act classification, a rough budget and the success criteria, and a one-hour presentation to your leadership.
Every step, from the contact form to the engineer in your team, is on our AI consulting page. If you want to know where your roadmap would start, tell us about the process in the contact form; after a short call we send you an offer that fits: AI consulting or a Tech Audit.
Sources
- Anthropic, Technical Deployment Lead posting, accessed 10 October 2026
- Scale AI, Frontier Agents Engineer posting, accessed 10 October 2026
- BCG, The Widening AI Value Gap, September 2025
- Eurostat, statistical report KS-01-26-009 on AI use in enterprises, 2026 edition, published 26 March 2026
- Regulation (EU) 2024/1689, the AI Act, and Regulation (EU) 2026/1744
72% of the executives and senior AI decision-makers surveyed by BCG named integrating AI with existing systems, tools and APIs as a challenge (BCG, The Widening AI Value Gap, September 2025).
Frequently asked questions
What does an AI transformation roadmap contain?
The business levers you want to move, the processes as they run, the use cases in ranked order, a card for each use case with preconditions, AI Act classification, a rough budget and success criteria, what not to build, and the first project, ready to start.
How do you rank AI use cases?
Score each use case from 1 to 5 on the same criteria: value, feasibility on your data, volume, integration effort, risk and adoption. The first use case is the one with clear value and the fewest preconditions, because it has to reach production and be used.
How long should an AI roadmap be?
As long as it takes to be executable. A few pages of context and a one-page card for each use case are usually enough. If the plan does not say what starts on Monday, who owns it and how success is measured, length does not help.
Should the roadmap include the AI Act classification?
Yes, for every use case. A high-risk use case under Annex III brings duties for the company using it from 2 December 2027 and is rarely a good first project. An agent that talks to customers has transparency duties from 2 August 2026.
What comes after the roadmap?
The first use case in the plan becomes a project: the production milestone, the adoption metric and the baseline are signed at kickoff, and a senior engineer works inside your team about three days a week until the system is in production and in use.
Want to discuss a project?
Book a free discovery call with the Sapio team.