What follows is my personal reading of what’s changing with artificial intelligence, with references to the essays I’ve written. It isn’t a balance sheet or a product roadmap — it’s the view of someone who designs software architectures that have to ship and stick.
Last revised: 27 May 2026. I update this page when a new essay extends or changes the picture.
Thesis, in one sentence
AI is a political fact dressed as a technical one: it reallocates power, work, and attention long before it hits accuracy targets. Productivity claims without a theory of friction are marketing.
Everything that follows is the reasoning behind that sentence.
Three things I keep coming back to
- The bottleneck isn’t AI, it’s the body. Agents do more work; we don’t therefore do less. Anyone measuring productivity in output alone ignores the scarcest variable: human attention, with sleep, limits, and finite time.
- Integrating AI is a supervision problem, not an inference problem. The real cost isn’t the token: it’s the human verification loop required to trust the output. Skipping the loop makes implicit assumptions about risk that someone else ends up paying.
- “Use AI” is not a method. It’s a slogan. The method is redesigning processes so that AI becomes useful — otherwise you’re buying a cockpit and leaving the same inexperienced pilot.
Essays on this topic
Work with me
My job here is to help you separate technical choices from political decisions dressed up as technical. No keynote deck: operational documents, with the risks named out loud.
Who it's for
CTOs or Heads of Product weighing ‘where to put AI’ and wanting a second head that isn’t selling anything
European SMEs that have received an ‘AI transformation’ proposal and don’t know how to evaluate it
Product teams that already integrated AI into a critical flow and now realise that human oversight was an assumption, not a plan
Boards and committees signing off on an AI budget that want to understand what they’re signing before they sign
How I work
- Use-case assessment (2–3 weeks)
I take a process or product where you’re putting AI, take it apart, and tell you what holds up and what doesn’t. Output: a document exposing assumptions, hidden costs (supervision, correction, fallback) and metrics for an honest ROI.
- Governance design (3–4 weeks)
Who decides what, who supervises, who intervenes when the output is wrong, who answers the end customer. AI Act and deployer responsibility handled with practical operations, not as two separate disciplines.
- Decision coaching (ongoing)
A couple of calls per month for the heaviest AI choices: build vs. buy, vendor selection, incident response, internal and public communication. A second head, not delivery.
Engagement FAQ
- Do you also do model training?
No. My scope is strategy, governance and integration — not MLOps, fine-tuning or data science. If that’s what you need, I can point you to people with better skills.
- How long does a typical engagement last?
Two to four weeks for an assessment, a couple of months for governance design, ongoing for coaching. No open-ended engagements.
- How is it billed?
Per output, not per day. The price is tied to a deliverable agreed up front.
- Do you work with AI vendors or only with end clients?
Only with buyers and integrators. To avoid conflicts of interest, I don’t take engagements from AI platform or model vendors.
Email me at hello@margiovanni.it with a couple of lines of context. I reply within a few business days with a concrete proposal, or a polite no if it's not my scope.
Questions & answers
What do you mean by 'AI as a political fact'?
Adoption redistributes power, time, and attention long before any accuracy milestone. Treating it as a purely technical matter lets you lose sight of who is paying the cost of the transition — typically the people working next to the system.
Are you against AI?
No. I’m against rigged balance sheets. When the ROI of an AI project ignores integration, human supervision, error correction and cognitive debt, the project isn’t an investment: it’s a disguised cost transfer.