AI strategy: from opportunity to portfolio

You don't choose the most striking demo, but the use cases that match strategy, capacity, and risk appetite.

After this chapterYou can prioritize a business-oriented AI portfolio and record value, evidence, owner, and stop condition for each choice.

Your route through this level

build one pilot dossier with Atelier Noor and complete the corresponding row of the same worksheet after each chapter. The assignments offer fictional solo routes where needed. You practise design and assessment; you do not need to own a business or wait 90 calendar days. Keep supplied exercise figures, your own chat observations and still-unknown business results separate. Real interviews, other reviewers and working connections are extension activities that require their own arrangements. The final dossier is complete when every rubric criterion is satisfactory with suitable evidence; a learning assignment does not prove that a production environment works.

Your progress0 of 48 lessons

1.1

Building an AI opportunity map

Start from business objectives and process friction, not from a list of product features.

Map revenue growth, customer value, lead time, quality, risk, and workload as separate objectives. Then look for processes where information is sought, assessed, rewritten, or transferred.

For each opportunity, record the current process, people involved, available data and the decision that needs to improve or happen faster. An opportunity without a process owner or a defined measure does not enter the pilot list.

Starting level: you can design a controlled workflow and test set as taught in Advanced. Expert teaches you to design, assess and manage a business pilot; it is not a complete programming course. First work through the assignments using the fictional Atelier Noor case or invented data. You do not need to upload real customer files, connect services or perform external actions to practise the method. After this lesson, start the workshop “One pilot worksheet for all eight chapters”. Build the same dossier step by step; each chapter refers to its row.

Business objective
Process friction
Available data
Decision point
Owner

Terms in plain language

Use case: A specific application of AI to a clearly defined task.
Pilot: A limited trial with measurement and stopping rules agreed in advance.
Process owner: The person responsible for the day-to-day way of working.

How you can use this

A service provider suspects that slow quote follow-up leads to missed opportunities and first investigates the handover and available data. The suspicion is not yet a demonstrated cause.

Try this prompt

Create an AI opportunity map for [company] under revenue, customer, operations, quality, and risk. Provide for each opportunity the process, user, data, decision point, and owner. Do not invent company data.

Knowledge check

Customer enquiries remain unanswered for a long time at the fictional small business Atelier Noor. The team immediately suggests a chatbot but does not know where the delay occurs. What do you document first?

The current waiting times and handovers, the desired improvement and who is responsible for it.The number of questions a chatbot can answer, with a target automation percentage as the initial goal.The price and capabilities of three chatbots as the basis for choosing the pilot.

Your practical exercise

Identify five opportunities for a fictional company. Start with Noor and add only opportunities labelled as your own hypotheses, such as internal scheduling or a summary of instructions. For each opportunity, record the process, role, supplied evidence and missing information. No real process owner available? Write five interview questions and leave unanswered business questions open. Completion check: each opportunity has an owner assigned as a design choice; invented interview answers do not count as business evidence.

Guided practice · then try it yourself

One pilot worksheet for all eight chapters

Work with Atelier Noor; you do not need your own company, a paid service or a real connection. The worksheet contains a fictional completed starting version and a blank version. Here, Expert means managing business use: choices, evidence, controls and ongoing management.

Open the pilot worksheet with completed example and blank version

Copy the blank version into a text file or notebook. Choose Noor’s draft replies, or an entirely fictional task of your own. Write down the scope and prohibited actions.
After each chapter, complete the corresponding row. Distinguish assumptions, planned tests, your own observations and supplied practice results. Use evidence names that let you retrieve the relevant work product.
Take turns playing the fictional roles of Noor (decision-maker), Alex (user) and Sam (critical reviewer). Record one objection per role. Do a separate second reading against the rubric; state that this remains self-assessment.
Update the decision row on days 15, 30, 60 and 90. These are pilot days, not a required study duration. A lesson exercise can simulate the same decision process on paper.

First check: can Alex use your task description to determine what information is needed, what to check and when to involve Noor? Add missing arrangements. Keep version 1; date later changes and record their reasons.

Before the real start: legal decision checklist

Complete this checklist gradually at H2, H6 and H8. In the first lesson, simply choose an owner and note the open questions. The questions help determine what assessment is needed; ticking them off is not legal clearance. For this course, your execution remains entirely fictional.

When	Question for your pilot dossier	Record
H2 – application	What does AI do, and with what consequences for people?	Application, AI Act role, possible prohibitions/high-risk classification and transparency; source and assessor.
H2 – data	Is personal data needed, and has the need for a DPIA been assessed?	Purpose, legal basis, minimum fields, roles; DPIA decision with a reason or an open action.
H6 – incident	Who acts in the event of a possible data breach?	Contact route, time of awareness, harm limitation, notification assessment and safely held evidence.
H8 – supplier	Which data processing takes place under whose instructions?	Roles and required processing arrangements; no real processing while a requirement is missing.
H8 – staff	What training and what information or participation are needed?	Role exercise, work instruction, workload impact and applicable consultation with a date.
Before practical use starts	Which open question could affect lawfulness or people?	Owner, necessary assessment and decision: resolve, narrow the scope or wait.

Short self-check at the corresponding chapter

Write down your first action and the reason each time. These are separate fictional scenarios; they do not change Noor’s existing pilot data. Then read the model answers.

After H2: Noor wants to replace her internally reviewed draft reply with a bot that talks directly to customers. Can the existing scope decision still suffice?
After H2: a service provider wants to use all personnel files to rank employees automatically for promotion. The supplier calls it an ordinary summary. What assessment is missing?
After H6: a fictional processor discovers that a shared folder containing customer data is accessible to unauthorised people. It wants to wait two days until all technical details are established. What do you have it do first?
After H8: a supplier processes customer addresses solely according to your instructions. There are no binding processing arrangements yet. Does a paid subscription make the processing ready to start?
After H8: a small employer without a CPBW or trade union delegation adds extra AI review work to existing daily duties. Why is the statement ‘we are too small for consultation’ insufficient?
After H8: everyone has ticked this course as complete. One reviewer still does not know when to block an incorrect draft. What training evidence is missing?

View the answer and assessment criteria

A usable starting version specifies: a draft reply about LAMP-2 from the approved fact sheet; no customer files or permission to send; Noor as owner; human review of every result; first test the hypothesis of moving from 12 to 7 minutes including review. The expected 25 hours per month represent capacity, not proven cash savings.

H4 contributes source assessment to the dossier; H5 provides practice in checking claims and measuring results. They do not automatically turn the reply pilot into a delivery or sales project. New tasks require a separate scope decision.

Model answers for the legal decision checklist

A new scope decision: direct AI interaction requires a different transparency assessment. Record the provider, customer notification and allocation of review responsibilities before that change. A staff member reviewing drafts is a different route from an autonomous customer bot.
Assess the actual function of ranking employees for promotion, the personal data involved and the consequences. Investigate AI Act classification and the DPIA obligation before using real files; the supplier’s marketing name is not classification evidence.
Restrict access, preserve necessary evidence and notify the controller without undue delay. Record the time of awareness and uncertainties. The controller then assesses notification duties; incomplete technical details are no reason to stay silent.
No. In the controller–processor relationship described, the required arrangements must first be in place. Also record the data and instructions that establish the relationship; payment does not answer that question.
Workload and well-being do not disappear because the company is small. Establish the direct participation route with the prevention adviser and adjust the allocation of tasks. CAO 39 and 81 have their own conditions; do not claim that every small employer must follow both complete procedures.
Task-specific exercise evidence: have the reviewer recognise a fictional error, stop release and escalate correctly. Record guidance and a further exercise round. A tick does not replace that support.

1.2

Selecting use cases based on impact, speed, and complexity

Prioritise value and the opportunity to learn; let high risk and limited verifiability constrain the score.

Assess impact, time to initial evidence, technical and organisational complexity, data quality and the consequences of errors. Apply firm exclusion rules for prohibited data, lack of an owner or unacceptable consequences.

A total score should not hide a critical weakness. Test whether the ranking changes when the impact is lower or the implementation time is higher.

Impact
Time until evidence
Complexity
Data quality
Error impact
Disqualifying condition

Terms in plain language

Disqualifying condition: A firm exclusion rule that a high overall score cannot outweigh.

How you can use this

An internal summary can be more pilot-worthy than automatic pricing, even when the latter promises more theoretical revenue.

Try this prompt

Rank [use-cases] by impact, time to proof, complexity, data quality, and error impact. Show assumptions, knock-outs, and sensitivity. Do not give a winner if key data is missing.

Knowledge check

A fictional use case has the highest overall score for expected value and speed. Nobody can take responsibility as process owner, although ownership is a mandatory requirement agreed in advance. What do you do?

Start with technical supervision and appoint a process owner later.Postpone the pilot until an authorised process owner is available.Start if the total score remains high after deducting points for missing ownership.

Your practical exercise

Score six candidate tasks; use your fictional opportunities from lesson 1.1 and add one more hypothesis. Set weights and exclusion criteria in advance. Ask a second decision-maker to respond if one is available. Solo: reassess first from the perspective of workload, then from the impact of errors, and change one weight. Completion check: the ranking’s sensitivity is visible; unknown essential data prevents a definite winner. The second round is not an independent assessment.

1.3

From assumption to business case

An AI business case is a testable hypothesis with a baseline measurement, full costs, and uncertainty.

Formulate which result changes, for whom, and against which baseline. Include implementation, control, training, management, corrective work, and license or API costs.

Use low, base and high scenarios. Identify the assumption with the greatest influence and the least expensive experiment that can test it.

Keep released capacity separate from actual cash savings. Fewer minutes per task may free up time for other work; this does not automatically reduce payroll costs. Count those hours as realisable financial value only when you can explain how they avoid costs or generate additional margin. Do not count the same hours again as additional profit from sales.

Baseline measurement
Value formula
Total costs
Scenarios
Decisive test

Terms in plain language

Business case: An evidence-based comparison of expected value, costs, risks and alternatives.
Baseline measurement: Performance before the change, measured using the same definitions as during the trial.

How you can use this

A content workflow counts writing time, review, publication, and correction of errors.

Try this prompt

Build a business case for [use-case] with baseline measurement, value formula, one-time and monthly costs, three scenarios, critical assumption, and cheapest valid test.

Knowledge check

The fictional Atelier Noor handles 300 enquiries a month. The old process takes 12 minutes per enquiry; the AI process, including review, takes 7 minutes. How do you include this difference in the business case?

As 60 hours of monthly capacity gained, based on the 300 enquiries and original handling time.As 25 hours of monthly financial savings, valued at the hourly rate of the employees involved.As 25 hours of potential capacity gained; financial value requires a justified use of that time.

Your practical exercise

Create a scenario analysis and choose which assumption the pilot should measure first.

1.4

Configure, connect, build or buy

Choose the simplest solution that demonstrably meets the requirements.

An instruction or skill may be sufficient for a defined task. A plugin can suit approved external data or actions; an API or supplier solution can suit scale, integration, logging or a custom user experience.

Compare data flows, permissions, vendor lock-in, maintenance work, recoverability and total costs. Do not build a custom solution without a demonstrated need, or buy a black box that prevents essential checks.

Requirements
Data flow
Permissions
Management
Exit

Terms in plain language

Permissions: Rights to access information and carry out actions.
Vendor lock-in: Dependency that makes switching providers difficult or expensive.
API: A technical interface between software applications.

How you can use this

A team first tests a controlled skill before the same logic is built into the CRM.

Try this prompt

Compare configuration, a plugin, custom API development and a supplier solution for [use case] on fit, data, control, maintenance, costs, scale and exit. Recommend the simplest valid option.

Knowledge check

A small business wants to structure ten internal status notes each week. A tested reusable instruction meets the requirements; no automated connection is needed. Which choice is best justified now?

Use the tested instruction; reconsider integration if requirements change.Build an API connection to accommodate the expected growth in weekly volume in advance.Choose a managed platform to incorporate the instruction directly into a wider business environment.

Your practical exercise

Set out the solution options and add a specific condition for switching or leaving the chosen solution.

Source for this lesson

OpenAI – Skills and plugins
Instruction packages and connections; check availability and permissions separately.
Checked: 2026-09-07

1.5

Creating a balanced AI roadmap

Combine fast learning pilots with foundations for data, people, and governance.

Alongside the pilot, plan how measurement will work, data agreements, training and decision-making. Order the dependencies: integration follows process and access design.

Use phases with criteria for moving forward, and reserve capacity to close discontinued experiments and remove solutions that deliver no value.

Pilot
Measurement plan
Data
People
Governance
Capacity to stop and close down

Terms in plain language

Governance: Agreements about who decides, checks, adjusts and can stop the work.

How you can use this

Quarter one investigates two use cases; in quarter two, only an application with sufficient evidence on value and risk proceeds. Choosing to implement neither can also be a sound decision.

Try this prompt

Create a 90-day and two-quarter roadmap for [portfolio]. Show dependencies, owner, budget range, decision gates, and stopping points.

Knowledge check

An SME is planning an AI pilot. The integration date is fixed, but the baseline measurement, data agreements and review capacity are still missing. What change belongs in the roadmap?

Keep the integration date and use the first pilot week to establish the baseline and review needs.Complete the baseline, data agreements and review arrangements first; tie the integration date to those starting conditions.Have the technical connection delivered first and use its acceptance test as the pilot’s starting gate.

Your practical exercise

Plan one 90-day pilot with three formal decision gates.

1.6

Ownership and decision rights

Each workflow needs a business, process, data, and technical point of contact.

The business owner is responsible for value and risk appetite; the process owner for the way of working; the data owner for access and quality; and technical management for integration and continuity.

Record who designs, tests, approves, changes, pauses, and decommissions. One person can combine roles, but decisions must remain visible.

Business owner
Process
Data
Technology
Authority to stop

Terms in plain language

Business owner: The person ultimately accountable for business value and acceptable risks.

How you can use this

A sales assistant makes proposals; only the account manager sends price or contract information.

Try this prompt

Create a decision-rights matrix for [workflow], covering design, access, testing, publication, incidents and shutdown. Name one person with ultimate accountability.

Your practical exercise

Fill in the matrix with roles and simulate one incident. Then complete row H1 of the pilot worksheet; keep assumptions and chosen starting conditions clearly identifiable.

Worked example

Atelier Noor chooses its first pilot

Fictional practice material; incorrect answers have been created deliberately for this exercise.

All data and answers in this case are fictional. Atelier Noor is a small business that handles 300 customer enquiries a month. Noor wants to reduce the workload without making incorrect promises to customers.

Input

Compare three options: A drafts answers from an approved product information sheet, after which an employee checks each draft; B independently sends quotes with discounts; C summarises a public trade newsletter. Estimated preparation time is two weeks for A, six for B and one for C. The hypothesis for A is a reduction from 12 to 7 minutes per enquiry, including human review. There are no measurement results yet.

Deliberately flawed practice answer

“Choose B: automated quotes generate the most revenue. A guarantees €750 profit a month. Start both immediately.”

Check

For B, evidence is missing on revenue, pricing authority and recovery from an incorrect quote. That rules it out for this first pilot, even if a weighted score looks attractive. For A, the calculation is 300 × 5 minutes = 1,500 minutes = 25 hours. At €30 an hour, this represents at most €750 in potential capacity value. There is no evidence that all this time will become available or generate money. C appears feasible, but the potential time saving has not been measured.

Improved result

“Start A as a limited trial with fictional enquiries. Noor is the process owner; an employee checks every answer. First measure current handling time, then time, factual accuracy and corrective work. Provisional monthly costs are €250; one-time setup costs are €1,000. Stop if there is an unauthorised promise or a data leak. B stays outside this trial until pricing rules, authority and recovery are demonstrably in place. For C, measure the problem before assigning priority.”

Try it yourself

A fourth proposal, D, is estimated to save 40 hours a month, but the rights to process the necessary data are unknown. May D rank above A because it seems to save more time? Write down a decision and the first next step.

View the model answer

Do not start D on the basis of its high estimated value. First establish the permitted processing and data flow. Only once that condition is met can D be assessed on comparable value, feasibility and risk.

Chapter assignment

Bring everything together

Make a shortlist of three use cases with scorecard, business case hypothesis, solution choice, owner, and 90-day roadmap.

My notes

Maximum 10,000 characters per note.

Progress and notes are stored only in this browser on this device. Do not enter sensitive data. Download your notes regularly. This course sets no automatic expiry date. You can delete the data through your browser’s site-data settings; export anything you wish to keep first. Browser settings or cleanup may erase it earlier. These local notes are not sent to Finaudax.