Before choosing technology
The question becomes concrete when it meets an activity.
A model can summarise, classify, write, compare, find relationships, produce code or use tools. These capabilities remain generic until they enter a specific situation.
Choose an activity and ask what result it should leave, which information it uses, who checks it, what happens if the AI is wrong and which part can be tested without changing the whole process at once.
“Use AI in marketing” is still a topic. “Prepare a first product-page draft from the catalogue, specifications and brand voice, then have the marketing lead approve it” already connects activity, sources, result and review.
Practical possibilities
Five ways AI can enter a project.
Understand and prepare information
Organise documents, connect related content, surface questions and prepare a synthesis for a person.
Transform documents and data
Extract fields, classify requests, compare versions or adapt content to a form that can be checked.
Support a decision
Collect evidence, show differences and propose options while responsibility and the final choice remain assigned.
Coordinate a workflow
Recognise the type of request, prepare the next steps and preserve state within defined effects and stopping conditions.
Build a digital product
Help design a system, write code, analyse defects, update documentation and prepare tests.
Choose the form
AI assistant, automation or agentic system?
A person directs the work.
The AI analyses, proposes or prepares an object. This form suits work that needs frequent human judgement or changes often.
The route is predefined.
The model classifies or generates part of the result while inputs, controls and destinations are established in advance.
The route can adapt.
The system preserves an objective and state, chooses among steps or tools and requires explicit sources, expertise, permissions and recovery.
The more complex form is not automatically the better one. The choice depends on how much the work changes, which effects the system can produce and how much interpretation is required between steps.
Practical object
First AI Project Map.
Fill in these eight fields for an activity you know. Answers can be provisional: the gaps show what must be understood before anything is built.
- 01
Activity
Which task, problem or stage of work do you want to improve?
- 02
Result
What should exist, change or become easier?
- 03
People
Who uses the result, who checks it and who is accountable for it?
- 04
Sources
Which documents, data, applications or people contain the necessary information?
- 05
Form
Do you need an assistant, automation or a system that chooses among several steps?
- 06
Effects
What may it read or prepare? What may it modify, send or publish?
- 07
Evidence
How will you recognise a useful result and detect an error?
- 08
First test
What is the smallest test that still represents the real work?
We want to improve [activity] to obtain [result]. The system uses [sources] and prepares [object] for [people]. It may perform [permitted effects]; [accountable person] reviews [decisions or risks]. The first test succeeds when [observable evidence].
Illustrative examples
Three scenarios that reveal the structure.
A professional firm compares recurring documents
An assistant prepares differences, missing data and questions. A professional verifies the result before it enters client work.
A team receives requests from several channels
An automation classifies the requests and prepares a common record. Uncertain cases remain visible to a responsible person.
A developer needs continuity
An assistant reads architecture, decisions, state and tests. Writing code, committing changes and publishing remain distinct actions.
These scenarios illustrate possible structures; they are not real cases or promises of results.
First trial
A first test should be able to fail usefully.
An initial test does not prove that AI works in general. It shows whether the project has sufficient information, criteria and controls.
Use a small but representative sample, name an accountable person, preserve inputs and outputs and decide when to stop. If the result is weak, you should be able to distinguish missing information, ambiguous instructions, an unsuitable capability or insufficient verification.
Frequently asked questions
Questions about getting started with AI.
Do I need to know how to code to start using AI?
No. You can begin with an AI assistant and a supervised activity. Code becomes necessary when you need integrations, controlled data access or repeatable behaviour that available tools do not already provide.
Do I need a large amount of data?
It depends on the use case. Some projects use a small set of well-chosen documents; others require structured data, permissions and quality controls. First identify the sources needed for the result.
When do I need an agentic system?
An agentic system becomes relevant when work must preserve state, adapt its route to what it encounters and use different tools. If the route is stable, a simpler automation may be easier to understand and control.
How do I evaluate whether an AI project is successful?
Compare the result with observable criteria such as time actually saved, errors detected, completeness, accepted quality or the ability to continue the work across sessions.