Generation AI - by Vincent Barat

Episode 4: “AI doesn't matter”

Faced with the financing difficulties they had to face from 2023 onwards and after the golden post-COVID period, many start-ups made “pivots” towards GenAI, which they chuck on everything. Without really understanding why or what purpose it served.

We could have done the same. Satya Nadella has also recently stated that SaaS will probably disappear, replaced by general-purpose AIs capable of interacting directly with data. This may happen one day, but we have remained cautious for a variety of reasons.

First of all, we had reservations about the very nature of these technologies:

• The inconsistency: the same causes must produce the same effects. However, two identical prompts do not guarantee the same outputs, due to the very nature of these “probabilistic” models.
• The approximation: when it comes to language, using synonyms is not very important. But at Albert, we provide a product that supports decision-making, based on numerical data. The slightest approximation can have dramatic social consequences
• The lack of a model dedicated to data: LLM allows you to make text, code. Computer vision allows you to understand images, sound... You can generate photos, videos... But there are still few models dedicated to data.

We also had a lot of questions about security:

• Sovereignty: many of our customers prohibit certain technologies or models. This is why we host our infrastructure with OVH and not AWS: so as not to close any doors. Too deep an integration of a particular technology, such as OpenAI (US risk) or Deepseek (China risk), could disqualify us from many large groups.
• Regulation: the framework is very fluid. The EU has enacted the AI Act, there is of course the GDPR, and betting everything on one technology could have proved risky. But the raison d'être of payroll software or HRIS, for example, is not just functional. As Anita Lettink often reminds us, it is also linked to legal requirements!

Finally, solving the GenAI economic equation is not necessarily trivial:

• The existing situation: companies already pay for their own models, and do not necessarily consider an additional cost for something they already use
• The cost price: do we need to rent server time, develop our own infrastructure, based on what use, what volume? Who will be willing to buy and how much, is it really worth it, knowing that only 15% of companies practice planning? Before going to Mars, wouldn't we first try to go to the Moon?

Our conviction from the start has been to bet on commodification. In 2003, Nicholas Carr wrote in an article that became famous in the IT world that “IT doesn't matter”. In the same way, all Gen AI services will revolve around pre-existing supplier actors, to which we must be agnostic. Because nothing is stable in this universe: there will be meteoric rises and bankruptcies, consolidations, partnerships. Having a long-term vision therefore means not imprisoning yourself with one supplier, and seeing the AI brick as a commodity.

This is why we have developed expertise in the technologies themselves, and in prompting tailored to our solution. After identifying the minor irritants for our users, we observed use cases where Gen AI removes these obstacles.

• Relevant use cases:

✅ Automate tedious tasks (e.g. extract a job architecture, clean up repositories).

✅ Inspire and facilitate brainstorming (e.g. create scenarios, identify business drivers).

✅ Structure and clarify ideas or concepts (e.g. generate reports and make data “speak”).

✅ Search for and synthesize internal and external data (e.g. skills benchmarking).

• Where do we exclude AI today?

❌ Specialized expertise: AI is a medium, not the expert himself.

❌ Decision-making: Decisions require multiple parameters and an understanding of the context that AI does not have.

The advantage is that this choice to develop expertise rather than a product module allows us to implement ourselves in any client ecosystem, and only if they so wish, using the technologies of their choice and in a secure manner. Thus, we have developed 6 agents:

• GPT Skill Builder: Easily build a skills repository based on the job architecture you provide.
• GPT Job Builder: Simply design accurate and structured job descriptions for your Strategic Workforce Planning process.
• GPT Job Archi Extractor: Automatically analyze your employee data to extract a clear and hierarchical job architecture.
• GPT Job Archi Cleaner: Clarify, consolidate and optimize your job architecture in just a few clicks.
• GPT Driver Builder: Quickly identify your key business performance drivers and transform them into rules that can be imported directly into Albert.
• GPT SWP Report Generator: Easily interpret your SWP results generated by Albert with clear and impactful reports.

These agents can be used “alongside” your own tools, and we will help you configure them. And you can access them for free right now!

They can also take a more integrated form, as a browser extension, to make Albert a true everyday companion for your planning tasks. And this while maintaining our initial vision: your safety, your conditions, your technology.

To read or reread the previous episodes of this series of articles devoted to the “Generation AI”, click here.