Skip to content

Shifting up a gear with AI
9 things to know to push your processes further

By Updated at 12 min read

You probably use ChatGPT, Claude or Gemini.

Like most people, you were impressed at first by the quality of the answers. Then came the disappointment, fast, as soon as you brought your real work topics: vague answers, missing context, mistakes, an inability to grasp what is specific about your company.

That is a normal stage. Between using a chatbot now and then and running real AI-driven processes, there is a gear to shift. A general-purpose AI knows nothing about your trade, your processes or your data. Without that context, it hits its limits quickly.

To get something solid out of it, you need to understand how it works, where it excels, where it fails, and how to fit it into your processes.

These 9 points will give you the basics to move from a simple conversational assistant to an AI that works for you. Count on 10 minutes of reading to gain some maturity on the subject.

  1. It invents when it has nothing to go on
  2. It will always take your side
  3. Models have hit a plateau
  4. The big models are all much the same
  5. What makes the difference: the harness
  6. Subscription or API: the real gap
  7. Your data: what leaves, what stays
  8. AI amplifies expertise, it does not replace it
  9. Who validates, who is accountable

1. It invents when it has nothing to go on

A language model does not go and look up an answer somewhere. It produces the most plausible continuation of what you just wrote, word after word. It is nothing but statistics.

When it has the material, that plausible continuation is correct (most of the time). When it does not, it produces something anyway: it has no “I don’t know” reflex. This is what we call a hallucination, and the real problem is not that it exists, it is that the tone gives nothing away. The false comes out with the same confidence as the true.

The counter is twofold. Your critical thinking, sharper on the subjects you master than on the ones you are discovering. And your data: your documents, your figures, your files, your business context, structured and reachable by an AI agent. A model plugged into your data invents far less, simply because it no longer has a gap to fill.

The reflex to keep is simple: anything verifiable gets verified. A figure, a date, a proper name, a legal reference, a link. The rest (structuring, rephrasing, summarizing, suggesting angles) calls for far less vigilance.

2. It will always take your side

This is the least known point, and probably the most expensive one in a company.

These models are trained to produce answers that users rate as good. And on average, we rate as good the answers that tell us we are right. The result is a deeply obliging tool.

Show it an idea: it will explain why that idea is good. Show it the opposite option 10 minutes later, in the same tone: it will explain why that is a good idea too. This is not dishonesty, it is a manufacturing bias.

As long as you use it to produce, you get away with it. The day you use it to decide, you no longer collect an opinion, you collect your own opinion rephrased with confidence. “Is my strategy sound?” is a question that can almost never receive a negative answer.

3 habits limit the damage:

  • Do not sign the question. Present the option without saying it is yours, or present 2 of them on equal footing.
  • Ask explicitly for the counter-argument. “Give me the 5 reasons this plan is going to fail” produces completely different material than “what do you think”.
  • Be wary of the compliment. When the answer opens by praising the quality of your question, it is talking about you, not about the subject.

For a call that matters, have a second model review it, in a fresh conversation, without telling it the history.

And in the end, do not delegate a risk you will have to carry yourself: AI does not make the decision for you, it gives you what you need to make it with some perspective.

3. Models have hit a plateau

Since spring 2026, the step between 2 versions is barely noticeable on everyday work: writing, summarizing, analyzing, answering, translating. The differences play out on edge cases that most professional uses never meet.

Progress continues, but along narrower axes: long autonomous tasks (an agent chaining 30 actions without losing the thread), speed, cost, image and video.

What this changes for you: stop waiting. The next version will not unblock your project, because the model is no longer what limits you. What limits you is what you feed it and the way you fit it into a process.

That is rather good news. 2 years ago, whatever you built was obsolete by the next release. Today, a well-built process holds over time, and it will automatically benefit from the coming versions.

4. The big models are all much the same

Claude, ChatGPT, Gemini and the others sit very close together. On the vast majority of what you will ask them, the gap between 2 models is smaller than the gap between 2 ways of asking the same question.

What sets them apart is temperament. One is chattier, another follows instructions more literally, a third holds up better over a long task. You learn these reactions through use, the way you learn a vehicle’s: it is not on the spec sheet, you feel it after a few hundred kilometres.

The practical consequence goes against what you read everywhere: the initial choice matters little, habit matters a lot. Pick one, use it every day for several months, learn what it does well and where it goes off the rails. Switching tools every month puts you back to zero each time, and it is the skill that is expensive, not the subscription.

5. What makes the difference: the harness

If the model is no longer the limiting factor, what is? Everything you attach to it. The industry has a word for that: the harness.

A model on its own is a head with no hands and no eyes. It only sees what you paste into the chat window, and it can do nothing but answer. The harness is what gives it access to your reality and the means to act on it. The same model, in 2 different harnesses, does not give the same results at all.

It is built around 3 pillars:

The frame and the permanent rules. A set of fundamental instructions the agent re-reads at the start of every session: the company’s editorial line and voice, the phrasings that are off limits, product naming conventions, validation rules. Context no longer has to be re-explained at every exchange, it is set once and for all.

Operational tooling (know-how and commands). This is where your business procedures, the agent’s skills and pre-programmed commands come together. Rather than letting every team member improvise their prompts, you encapsulate precise working methods: check and optimize a visual, review the quality level of a text, or trigger a complex sequence of actions through a single dedicated command. The work no longer depends on the way the question was asked: the team uses standardized procedures that guarantee a consistent result.

The connections to your systems. Direct access to the tools that hold the truth: your ERP, your CRM, your website, your databases. Thanks to standard protocols like MCP, wiring those sources up has become simple. A question asked in plain language (“what is revenue by product since the start of the month?”) then produces an analysis based on real sales, which avoids answers that are merely “plausible” but wrong (point 1).

One constraint runs through all 3 pillars: the harness serves as much to forbid as to allow. Connectors must only expose controlled access (read-only on sensitive data, secrets out of reach, mandatory human validation before publishing). Guardrails are not informal instructions, they are written into the structure of the tooling itself.

It is this layer, not the model, that sets you apart from your competitors. It is made of your trade, your data and your rules. It does not come with a subscription, and it stays valid when you change model.

6. Subscription or API: the real gap

There are 2 ways to pay for AI, for 2 different uses.

The subscription is a monthly flat fee: one human, one interface, a fixed price. The API is technical access to plug the model into your own tools, billed by the volume of text exchanged.

My own setup costs me about €200 a month in subscriptions. The same usage billed through the API would top $4,000. That is what changes everything day to day: with a flat fee, you try, you throw away, you start over without watching the meter.

In practice:

  • Flat fee when someone is at the controls. Trial and error, back and forth, a judgment call every time. The cost is known upfront, so nobody holds back.
  • API when nobody is sitting in front of it anymore. A task that is written down, tested, always identical, that has to run on its own: 500 product descriptions to rewrite, emails to sort, a synchronization to feed.

Order matters: you fine-tune on the flat fee, then you industrialize through the API, once the volume justifies it.

The mistake to avoid: keeping a repetitive, perfectly documented task in your own hands when it could have been running on its own for the last 6 months. You have to accept, measure and optimize the cost of deploying at scale rather than paying the human cost of it.

7. Your data: what leaves, what stays

There are 3 categories to tell apart, and one rule to write down.

What leaves for the AI provider: everything you paste into the interface. On a consumer account, those exchanges can be used to improve the models, depending on your account settings. On professional, enterprise and API plans, use for training is (in principle) disabled by default, with a limited retention period.

What stays with you: what you do not send, obviously, but also what you process on models hosted in Europe or on your own servers. That is a real option today, a little less capable and heavier to set up, to be kept for sensitive processing.

What never gets sent: passwords and access keys, a complete customer file with names, health data, anything covered by a non-disclosure agreement. When the need exists anyway, you anonymize upstream (names become identifiers, processing happens on neutral data) or you go through a self-hosted model.

8. AI amplifies expertise, it does not replace it

Software people have a phrase for this: garbage in, garbage out. What goes in determines what comes out.

Without domain knowledge, a model shuffles hot air around very elegantly. With your data, your business context and the eye of someone able to judge the result, it becomes a multiplier.

The consequence is counter-intuitive: it is your best people who gain the most. Someone who can already recognize good work goes 2 to 3 times faster. Someone who cannot mostly produces mediocrity, faster than before.

So you do not install “AI in the company”. You install it around a person who masters their subject, on a process they know by heart, and it is that combination that produces the gain. It is also why a competitor will not reproduce your advantage by copying your prompt: they are missing your data, your trade and your eye.

9. Who validates, who is accountable

10,000 files deleted in one second.

An AI agent was cleaning up what it took for a temporary folder. It was in fact the image library of a production website.

The backups did their job. They existed because 15 years in the trade had taught me to have them. That is the whole subject of this last point: what comes out commits your company, never the provider of the model.

2 regimes coexist, depending on whether an automatic guardrail exists or not.

When one exists (tests that check the code, a calculation you can re-run, a figure that has to come out right), you can let AI work more freely: the machine checks the machine. That is the airport security gate, it is not the pilot inspecting every passenger.

When there is none, a human validates before publishing. An article, a product page, a client email, a decision: no automated test will tell you whether a text rings true, whether it is faithful to your voice, whether it will resonate with your audience. A text that “smells” of AI, even subtly, is trust lost.

So for every process you set up, ask a single question: if the output is wrong, who notices, and when? If the answer is “nobody” or “the client”, a step is missing.


The model is not the subject. What makes the difference is the whole chain: your know-how, your data, a decent model, the tooling that connects them, guardrails, and someone who validates. Everyone has access to the model for €20 a month. The rest gets built, and that is precisely where the gap sits between using ChatGPT and shifting up a gear.

You can read the detail of that chain from a practitioner’s side in my experience report on 2 years of AI adoption. To hear more about real use cases already in production, subscribe below.

AI, SEO and Web strategy session

Which process should you accelerate first?

Evaluate the expected gain, available data, and required guardrails before committing to an automation project.

Choose my first use case

€350 excl. VAT, up to 90 minutes

Email • RSS • Twitter

Get notified of new posts

Keep reading about AI: