AI: built into the tools, reviewed by humans
I only bring AI into a project after running it in my own work. This site is the test bench: audio narration of the articles, a training course sent by email, tool-based diagnostics, all built and maintained by me. What matters is not the tool you pick, it is the task you target, human control and adoption by the team.
Why AI divides as much as it promises
Generative AI creates contradictory pressure: adopt it fast for fear of falling behind, or distrust it after a disappointment, made-up content, a mediocre result passed on without review. Both reactions lead to the same problem: tools deployed without method.
The right question is never "should we use AI" but "which specific task, with what measurable gain, under what supervision". Mapping tasks before choosing a tool, piloting on a limited scope and setting a clear rule on who validates what the machine produces prevent most disappointments, in either direction.
On top of this comes a real regulatory layer with the EU AI Act. Prohibited practices have been banned since February 2025. The Digital Omnibus on AI, published in July 2026, pushed the obligations for high-risk systems back to 2 December 2027 for Annex III systems and to 2 August 2028 for those built into regulated products; the requirements themselves are unchanged. Scoping an AI project means knowing which category it falls into before deploying it.
Sources: Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), OJ L of 24 July 2026 ; Regulation (EU) 2024/1689 on artificial intelligence
My method
If a tool does not save me time on my own work, I will not recommend it for a team's. Once past that filter, integration follows four steps.
Map the tasks
Volume, repetition, format: estimate the expected gain before choosing a tool.
Assess the risk
Data involved, category under the AI Act, and the person who validates the output.
Pilot with a small team
One specific task, measured before and after, on a scope that can still be corrected.
Govern and train
A written review rule, training in critical review, then the rollout.
In a web agency: AI at every stage, reviewed every time
On a web project run in an agency, I put AI tools to work at every stage: SEO, lead generation, working prototypes, UX and help with functional specifications. For each one, the same rule: one specific task given to the tool, and someone from that role who reviews the result before it moves on.
That review is what makes the whole thing usable. A specification prepared with AI is reviewed by the person who will defend it in front of developers, a generated prototype is tested by the person who designs the UX, an SEO text is approved by the person accountable for it.
Tools I built to scope an AI project
- AI Automation ROI CalculatorBefore recommending an AI tool to a client, the first question is always the same: is it actually worth the setup cost? This calculator starts from the real parameters of the task, subtracts the solution costs you enter and gives the net saving, the first-year ROI and the payback period.
- EU AI Act Compliance CheckerThe EU AI Act has been progressively applying since August 2024. Provider, deployer, importer: your obligations depend on your role and your system's application domain. This 6-step questionnaire identifies your risk level and the applicable articles.
- Spec Decomposer"I want a contact form" is a brief, not a specification. This tool shows a client what really needs to be specified for a developer to start, without guessing, without going back, without scope creep.
Frequently asked questions
How do you identify tasks to automate with AI?
By first mapping repetitive, high-volume tasks with low human added value and a standardised format. My rule of thumb: if a task can be described as a short, stable procedure, it is probably automatable. The deciding factor remains the gain measured over twelve months, not the novelty of the technology.
Is generative AI reliable for professional use?
With clear governance, yes. Errors occur when AI is given tasks without systematic human review. The framework I apply: AI produces, a human validates before anything goes out. This greatly reduces the risks while keeping the real productivity gains.
Is my AI project covered by the AI Act?
It depends on its use. Prohibited practices have been banned since February 2025; high-risk systems (recruitment, credit, education, among others) have obligations pushed back to 2 December 2027 by Regulation (EU) 2026/1744. My free compliance checker places a project in these categories in a few questions.