AI in your business: where to start — concretely (and without the gimmicks)
The classic mistake: doing AI for AI's sake
For the past two years, the pressure has been everywhere. Boards of directors are demanding an "AI strategy." Competitors announce they are "leveraging artificial intelligence." Vendors showcase impressive demos.
The result: many companies launch an AI project without really knowing what problem they are trying to solve. They deploy a chatbot that misses the mark, buy an "AI-powered" tool that changes nothing about existing habits, and fund a proof of concept that never makes it to production.
This is not a technology problem. It is a methodology problem.
Useful AI is invisible. You don't notice it because it is part of the workflow — and you measure it by time saved, errors reduced, or better-informed decisions.
Start with the problem, not the technology
The right question is not: "How could we use AI?" It is: "Where are we losing time, money, or quality — and could data fix that?"
This is a fundamental shift in mindset. Across the 45+ projects I have delivered over my 11 years in IT — in France, the UAE, Senegal, and internationally — the projects that stood the test of time always started from a real, documented, quantifiable pain point.
AI is a tool. A powerful tool, but a tool. It does not create value on its own; it amplifies what already exists in your processes and your data.
Identify 1 or 2 high-ROI use cases
You do not need to transform your entire organization at once. You need a first quick win — measurable, and credible enough to build internal momentum.
The most common high-ROI use cases:
-
Automating low-value repetitive tasks — document processing, data entry, email triage, recurring report generation. These tasks consume human hours without requiring complex judgment. This is where intelligent automation delivers the best returns.
-
Accelerating an existing process — contract review, lead qualification, customer feedback analysis, meeting summary generation. AI does not replace the expert; it prepares the ground so the expert can decide faster and better.
-
Unlocking dormant data — in many organizations, years of customer, operational, or financial data have never been properly analyzed. A well-calibrated model can surface signals that no one had time to look for.
The golden rule: choose a use case where you can measure before and after. Without a baseline indicator, there is no steering — and no conviction to move forward.
The data question: quality, access, confidentiality
AI is only as good as the data you feed it. This is the reality that vendors rarely mention in their demos.
Before discussing models, ask yourself three questions:
1. Quality — Is your data reliable, complete, and consistent? Poorly structured data will produce unusable — or even dangerous — results.
2. Access — Can your teams query their own data easily? I have seen organizations where the data existed but was inaccessible to the people who needed it. At MGEN, we built a data hub and APIs to solve exactly this problem; at Air Liquide, a health data platform on AWS with rigorous master data management. This foundation is often what is missing.
3. Confidentiality — Which data can you send to an external cloud service? Patient data, HR data, strategic data: some information cannot leave your security perimeter. This is a non-negotiable point, not an option.
Build vs Buy: use what exists or build custom?
For most use cases, existing models (GPT, Claude, Gemini…) are sufficient — provided they are properly integrated into your processes and given clear, scoped instructions.
But there are sectors where data sovereignty is non-negotiable.
Healthcare is the most obvious example. This is precisely why at Saana, where I serve as CTO, we built a sovereign vertical LLM: health data cannot flow through uncertified third-party infrastructure. Our partner hospitals in Montpellier use an application that makes clinic educational content accessible to patients — hosting and data 100% sovereign, with no compromise.
The same principle applies to other sectors: defense, finance, legal, sensitive HR. Whenever confidentiality is critical, a custom solution or private hosting is not a luxury — it is a necessity.
Risks to frame from day one
AI is not infallible. Three risks deserve to be addressed explicitly from the outset:
-
Hallucinations — Language models can produce incorrect answers with apparent confidence. Any AI application that touches important decisions must include a human verification mechanism or source validation step.
-
Confidentiality — Sending internal data to a cloud model without prior legal review exposes the organization to real regulatory risks (GDPR, trade secrets…). This must be addressed before launch, not after.
-
Dependency — Relying 100% on a single AI vendor without a continuity plan is repeating the cloud dependency mistakes of the 2010s. Diversify, document, and maintain ownership of your data.
Training teams: the most underestimated investment
The best technology fails if the people who are supposed to use it do not understand what it does — or its limitations.
Having trained 350 people over my career, including 50 specifically on AI, has taught me one thing: resistance to change is not bad faith. It is uncertainty. When teams concretely understand what the tool does, why it sometimes gets things wrong, and how to verify its outputs, adoption follows naturally.
Building a training session into every AI deployment is not optional. It is what transforms a tool into a working habit.
A 4-step framework to move forward without getting lost
Here is the methodology I apply systematically, whether for a project lasting a few weeks or a 12-month transformation:
1. Identify — Map painful processes and select 1 to 2 use cases with an estimable ROI. No more.
2. Test small — Launch a fast pilot (4 to 8 weeks) on a limited scope, with real data and real constraints. The goal is not perfection — it is learning.
3. Measure — Rigorously compare before and after against the indicators defined upfront (time saved, errors reduced, user satisfaction…). Across my projects, this rigor has delivered an average of -30% on delivery timelines.
4. Scale — Once the pilot is validated, deploy at larger scale, train the teams, document, and monitor. This is where lasting value takes root.
Useful AI is invisible — and measurable
AI that works, you do not see. You see its effects: a report that takes 10 minutes instead of 3 hours, a customer service team that responds faster, a salesperson who prepares meetings with insights they would never have had time to find on their own.
This is not science fiction. These are achievable results with a solid methodology, well-organized data, and trained teams.
The hype will pass. The value created stays.
Let's talk about your project
Do you have a process that costs too much in time or errors? Data you don't know how to leverage? An automation idea you haven't dared to test yet?
Twenty helps executives and teams move from idea to implementation — with the expertise of a large organization and the closeness of a trusted partner.
Reach out to Dibrilou directly:
- WhatsApp: +33 6 34 42 50 56
- Email: contact@twentyconsultancy.com
A first conversation is often enough to know whether AI can create value for you — and where to start.