The AI landscape has moved extraordinarily fast these past few years, and these past few months. Vendors promise the sky; in practice it turns out harder than expected, every time. What makes an implementation succeed is a real grasp of what AI can and cannot do, combined with the right business decisions and genuine technical knowledge of the technology itself. Missing any one of the three is enough. Over the past year I have gone through every facet of AI intensively and actually implemented it, rather than read about it — what follows comes out of that, out of what worked and out of what did not.
01
Operational: the repetitive middle
The wins that hold up are unglamorous. Drafting the same kind of email, summarising a call, turning a form into a structured record, first-pass triage of a shared inbox. Small, frequent, and cheap when wrong. Start there, not with the thing that would impress a board — the board demo is where pilots go to die.
02
Integrations: one system, not five
AI becomes useful when it can see the data you already have and write back into the tool people already open. That is also where the cost sits: authentication, permissions, and someone owning it when the API changes underneath. One integration done properly beats five half-wired, and the half-wired ones are the ones that quietly stop working.
03
Agents: acting, not answering
An agent does things: sends, books, files, pays. The question is not whether it can, it is what happens when it is wrong at three in the morning with nobody watching. Narrow remit, a log of everything it did, and a human approval step wherever a mistake costs money or is public.
- What can it do without asking?
- Where is the record of what it did?
- Who notices if it goes quiet — or goes wrong quietly?
04
AI doing the development work
This genuinely changes the shape of building software: a prototype in an afternoon, an unfamiliar language within reach, boilerplate gone. It does not remove the need to understand what came out. Review it the way you would review the work of a capable new hire — because that is what it is, confidence included, and a new hire who is never uncertain is the one you check hardest.
Plan de mise en œuvre
Week 0
Write down the before
Pick one repetitive task and measure it as it is today: how long, how often, how many errors. Skip this and you will never be able to show whether anything improved, and you will end up quoting someone else's percentage.
Week 1–2
Do it by hand, with AI
Same task, same people, AI in the loop, nothing automated yet. Two weeks is enough to learn whether the task suits it at all, and stopping costs nothing.
Week 3–4
Wire it into one system
The tool people already have open. Resist the platform purchase until you know which single integration you actually need.
Before it goes wide
Decide where a human signs
List the actions where being wrong costs money, reputation, or a regulator's attention. Those keep an approval step permanently — not until you trust it, permanently.
Month 2
Measure the same thing again
The metric from week 0, unchanged. If it did not move, say so and stop. That result is worth more than a pilot that quietly continues because nobody wants to be the one to kill it.
Ongoing
Name an owner
Someone whose job includes it when the model changes, the API changes, or it starts being confidently wrong. Without a name on it, it degrades and nobody notices.
Pièges
Buying a platform before having a use case
The demo always works — it was built to. Start from a task you can name and a number you already have.
Automating a bad process, faster
If the process is wrong, AI makes it wrong at scale and with more confidence. Fix the process first; it is usually the cheaper of the two anyway.
Agents with write access and no audit trail
The failure is not dramatic, it is quiet: a hundred small actions nobody reviewed. Logging stops being optional the moment it can act rather than answer — and a log nobody reads is not a log. Monitor it for errors and keep figures on it: how often, what kind, and whether it is trending up.
Confidence read as accuracy
It sounds equally certain when it is right and when it is not. Any process that depends on a human noticing hesitation will fail, because there is none to notice.
No baseline, so nothing to prove
The most common outcome I see: something clearly helped, nobody can show it, and the budget conversation goes badly.
Gains
Time back on repetitive work
The unglamorous tasks are both where it is most reliable and where the hours actually are.
First drafts stop costing time
The most expensive minutes of a task are usually the first ones: the blank document, the mail you restart three times. A usable first version in seconds takes that away. What comes after it — checking, cutting, rewriting — still takes exactly as long as it did.
Institutional knowledge becomes searchable
Documents, calls and threads become answerable rather than merely stored. This is retrieval doing the work, not generation — and it is the part that tends to keep paying.
Small teams reach further
A company of one to ten people can now do things that used to need a department, provided somebody owns the output rather than forwarding it.
Risques
Data leaving the company
Know which service processes what, where it runs, and what the contract says about training on your inputs. Check it per tool — nothing carries over from one vendor to the next.
Regulatory exposure
GDPR already applies, and the EU AI Act adds obligations that depend on what the system does and in which sector. I am not the person to tell you which category you fall into: get that assessed rather than assumed. I built a readiness check for exactly that question.
Vendor dependence
Pricing, models and terms change on their timetable, not yours. Keep the prompts and the data on your side of the line so that switching is a project rather than a rebuild.
Skills quietly atrophying
If nobody on the team can do the task unaided in two years, the tool has become infrastructure. That may be an acceptable trade — but decide it deliberately rather than discovering it.
Enthusiasm outrunning evidence
Mine included. Anyone selling you this should be able to say what they measured and on how many cases. If the answer is a slide, it is not an answer.