An AI project does not start with the choice of the model, but with the business value, the data, the trust, the integration into the IS and the ability to move from prototype to real impact.
Over the past two years, many organizations have been approaching AI through the wrong door.
The first question asked is often:
Which model should we use?
GPT? Claudius? Mistral? Llama? An open source model? A proprietary model? A specialized model? A model hosted in Azure, AWS or elsewhere?
This question is legitimate.
But she comes too soon.
Because when a company starts an AI project, the choice of model is almost never the real starting point.
The first question should be much simpler, but also much more demanding:
What business problem do we want to transform, with what data, within what trust framework, and with what industrialization capacity?
This is where an AI project really begins.
The model is visible, but it is not what creates the value
Generative AI models are fascinating because they are spectacular.
They write, summarize, classify, translate, reason, generate code, analyze documents, interact with users and sometimes orchestrate complex actions.
They give the impression that the main part of the project depends on their selection.
However, in a company, value does not come only from the model.
It comes from its ability to be connected to a business context, to reliable data, to governance rules, to existing processes and to concrete uses.
A very good model, poorly integrated, often produces little value.
A simply correct model, well contextualized, well governed and well integrated into a business process, can profoundly transform an activity.
This is an essential idea:
AI is not just a generation technology. It is an integration technology.
The first question: what business use?
Before talking about model, we need to talk about use.
Do we want to reduce the processing time of a file?
Improve the quality of customer support?
Accelerate documentary production?
Assist developers in the software life cycle?
Help sales teams prepare for their meetings?
Detect operational anomalies more quickly?
Automate part of a repetitive business process?
An AI project should not start with:
“We want to use AI.”
It should start with:
“We have a specific, measurable, recurring, costly or strategic problem, and AI may be able to help us solve it.”
The nuance is important.
In the first case, the company is looking for a testing ground for a technology.
In the second, it seeks a lever to improve a business result.
The second question: what data?
An AI without useful data remains a demonstrator.
In many projects, the obstacle is not the model.
It is the quality, accessibility, freshness, governance or structuring of data.
Where is the data?
Are they in files, databases, SaaS tools, emails, tickets, wikis, ERPs, CRMs?
Are they reliable?
Are they up to date?
Are they understandable?
Can they be exploited by an AI without exposing sensitive information?
This is often where the AI project becomes an architecture project.
Because for an AI assistant to provide a useful response, it must understand the context.
And this context does not only come from his general training.
It comes from internal data, business documents, organizational rules, histories, repositories, past decisions and constraints specific to the company.
The subject is therefore not only:
“Which model to choose?”
But rather:
“How do we give the model the right context, at the right time, with the right level of security?”
The third question: what level of confidence?
Generative AI introduces a new reality in information systems: probability.
Classic software executes a deterministic rule.
An AI model produces a likely response, influenced by its training, prompt, context and parameters.
This does not mean that it is unusable in business.
But that means it has to be supervised.
What level of error is acceptable?
Which answer needs to be verified by a human?
What actions can be automated?
What decisions should remain under human control?
How to trace the answers?
How to audit sources?
How to prevent sensitive data from leaking?
How to deal with bias, hallucinations or non-compliant responses?
Trust cannot be decreed.
It is built through architecture, governance, testing, observability and safeguards.
A serious AI project is therefore not just about connecting a model to an interface.
It consists of creating a reliable system around the model.
The fourth question: what integration into the IS?
An AI prototype can be impressive in just a few days.
A reliable AI product requires real industrialization capacity.
This is often where many projects fail.
They work in the laboratory, but not in production.
They impress in demonstration, but do not integrate into real processes.
They respond well on a few examples, but become unstable on a large scale.
They appeal to innovation teams, but remain difficult to adopt by business lines.
To move from prototype to product, you have to deal with very concrete subjects:
identity, access rights, confidentiality, supervision, monitoring, cost management, response quality, performance, latency, security, compliance, lifecycle management of prompts, models and data.
In other words, AI must become a component of the information system.
Not an isolated experiment.
The real subject: building an AI value chain
A successful AI project rarely relies on just one element.
It is based on a complete chain:
Business use case → data → context → model → security → integration → user experience → value measurement.
The model is an important piece.
But it’s just one room.
Value comes when the whole works as a coherent system.
This is why the most mature organizations are not just wondering which model to use.
They are building a platform, a method and governance capable of replicating AI uses at scale.
They move from demonstrator logic to product logic.
From a logic of experimentation to a logic of industrialization.
From a logic of technological fascination to a logic of business impact.
The model is chosen at the end, not at the beginning
The choice of model should be a consequence of needs, not a starting hypothesis.
If the use case requires complex reasoning, the choice will be different.
If it requires low latency, the choice will be different.
If it processes sensitive data, the choice will be different.
If it requires private deployment, the choice will be different.
If it must be highly specialized, the choice will be different.
If it has to optimize costs, the choice will be different.
The right model therefore depends on the context.
There is no absolute best model.
There is a model adapted to a use, a risk, a budget, an architecture and a business strategy.
Bottom line: don’t start with the model, start with the value
Starting an AI project by choosing a model is like starting a digital transformation project by choosing a technical framework.
It is sometimes necessary.
But that’s not where success lies.
The real question is not:
“What model are we going to use?”
The real question is:
“What capability do we want to create for the business?”
An ability to decide better.
To produce better.
To better serve customers.
To reduce complexity.
To accelerate the teams.
To make operations more reliable.
To transform the employee experience.
To create new services.
AI is not a model project.
It is a project of value, data, architecture, trust and adoption.
And this is precisely why the first question should never be:
“Which model to choose?”
But rather:
“What problem do we want to solve, with what level of confidence, and how will we move from experimentation to real impact?”
