Regardless of the model as long as we have knowledge: GPT-5 is not the magic answer to all your AI projects in business, and that is good news.
For 18 months all customer relations actors have been agreeing to say that artificial intelligence is a very good ally to respond always faster on many contact channels, 24 hours a day, 7 days a week. Whether at the heart of the internal tools of advisers to help them on a daily basis, or in direct contact with customers, AI has many promises to increase operational efficiency and satisfaction of all stakeholders.
However, many projects started in proof of concept do not go as far as the real production, on a scale, of these solutions. Now is the time to grasp these technologies to integrate them where they make sense, in support and confidence.
Business knowledge, essential base of confidence
A successful AI project is an AI project that goes into production with all the guarantees necessary to understand the behavior of the algorithm before making it available to an end user. Rather than a race for the biggest language models, it is above all a data and governance project to break the information silos, pool efforts within the organization and organize knowledge management for the future
Bringing together and structuring all sources of information is not simple, I note it with large accounts for more than 10 years. This step is however essential to consider the success of a project involving generative AI. The construction, or optimization, of this source of truth, verified and validated by all internal experts, is the key to success for all IA projects of large companies.
It is necessary to transform the raw data available everywhere in more or less structured formats, into information actually usable for your teams as for AI. Our R&D work on information research and knowledge management subjects confirm the importance of complementarity between business experts and AI agents!
Take advantage of the conversational to solve problems hitherto inaccessible
I have been convinced for several years, this is not the number of parameters or the time spent to train models of language that will make the difference for our customers on their very concrete use cases. On the contrary, the more the models are able to perform ever more complex tasks, the more we must be able to supervise them to make them usable in a professional context in production.
OPENAI goes in this direction with GPT-5 by emphasizing the capabilities of the model not to respond or to explain why it can only partially respond to a user’s request. Stop hallucinations and dangerous responses for the end customer who exchanges with a chatbot using generative AI!
To pass the course of proof of concept, it is the ability of a product to deliver fair answers, with good tone, which makes the difference. And if we are still going a little further, the connection of a knowledge base to a complete ecosystem around CRM tools and complementary data channels is at least as important as the supposed intelligence of a generative model. Large languages of languages are very good at summing up, reformulating, adding context, and making the exchanges more conversational, let us use them to enrich experiences!
Master the specific opportunities (and risks) specific to customer relations
We cannot let an LLM respond to a customer without having control over this answer. The black box effect is not possible! Model suppliers (OPENAI in mind) announce to work on reducing hallucinations, but you have to go even further with solutions tested and proven specifically for the uses of each sector and of each profession.
The management of conversational toxicity is also central to the deployment of AI tools to an audience of customers. We have a duty to do better than the filters on shelf in Providers to thwart subjects at reputational risk, insults or attempts to “jailbreak” from the chatbot. This is what we do with our research team that published a paper on the subject at the beginning of summer during the Coria-Taln conference, an academic reference on information research and language processing subjects.
Seize the opportunity for the AI of trust made in France
Deploying AI agents is not an end in itself, but there are many opportunities, cases of relevant and efficient uses too!
At the beginning of the year Gartner predicted that 30% of the projects launched in proof of concept in 2025 would not go as far as production. This summer put it even further with a study which demonstrates that 95% of generative AI projects in business do not reach their objectives.
This can be against intuitive when Sam Altman announces GPT-5 as being as intelligent as a doctor in each field, yet business expertise and knowledge of customers by human professionals has never been so important. This expertise is essential to pass demonstrators to tools deployed in production which bring value and facilitate the daily lives of customers but also of their advisers. Let’s go further than the fashion effect to think in the long term with confidence tools on which we can count on a daily basis.
