Third and final article in Lecko’s presentation of the state of the art in internal transformation . After hyperconnection (Hyperconnectivity in the workplace: digital becomes a burden) and responsible digitalization (Sustainable digital: no more hypocrisy), we’re going to talk about artificial intelligence in the work environment.
And of course, I can only recommend that you download this comprehensive study, based on an analysis of the practices of 16,000 users in real-life work situations.
2023 saw the explosion of generative AI, both for the general public and for professionals, and 2024 confirmed this trend with an acceleration in innovation and questions about its integration into the working environment. Hence the legitimate question: is generative AI a genuine revolution for the digital workplace, or just another technological brick?
The sudden irruption of generative AI
While AI is not a new concept, the history of generative AI is relatively recent. The “Transformer” concept was officially shared in 2017, and the launch of ChatGPT dates back to November 2022. The latter put generative AI under the spotlight, forcing other market players to launch their own solutions such as Google’s Gemini or Anthropic’s Claude.
Unlike traditional chatbots, which simply point to existing content, generative AI has the ability to create original content. These models, based on LLMs (Large Language Models), are trained on huge databases to generate text, images and voices that resemble human production, but are no more and no less than the product of a probabilistic model (the AI doesn’t create anything, it just suggests the most plausible text, and when it doesn’t have enough information on a subject it can fall victim to hallucinations and tell you anything).
Generative AI in the workplace: potential and limits
While using ChatGPT may seem like magic in a general context, with an undeniable “Wow effect”, enterprise requirements are different (Why enterprise AI can’t keep up with consumer AI: beyond ChatGPT, a more complex reality).
Reliability of responses and confidentiality of exchanges are essential, and a model trained on general data may not be sufficient for a company’s specific needs. For example, learning the French language via ChatGPT may be sufficient for everyday use, but there’s no guarantee that it won’t be for summarizing technical documents with specific jargon. Similarly, can we trust an LLM’s knowledge of French law to draft a contract?
There are several ways of overcoming these limitations.
First of all, contextualizing the prompt consists in giving very precise instructions to the generative AI to guide its response. This is particularly effective for simple tasks with a limited number of documents.
Next come RAG (Retrieval Augmented Generation) and RIG (Retrieval Interleaved Generation). These techniques integrate the company’s specific knowledge into the conversation, using internal or external data sources, to obtain more relevant answers. Above all, RAG makes it possible to take into account company-specific data without having to recreate an LLM.
However, creating and training your own AI model represents a major investment that few companies can afford. A learning cycle for a model such as GPT-4 is estimated at $12 million, according to the study (as far as I’m concerned, the figures I’ve found give this figure for GPT-3, but $63 million for GPT4 (How Much Did It Cost to Train GPT-4? Let’s Break It Down)), not to mention the environmental impact. At this price, we’re expecting a more-than-tangible ROI, even if we can also see it as a necessary intermediate step before reaping the rewards in the more-or-less distant future, and as a way of ensuring that we don’t fall off the learning curve (ROI Vs. RONI: why businesses should invest in AI despite uncertain ROI).
The adoption of generative AI in the enterprise: still a long way off
Despite the popularity of generative AI, and contrary to some optimistic speeches suggesting that it’s just a formality, its adoption in the workplace is far from widespread.
According to a study carried out by IFOP, 57% of AI users do so on a personal basis, compared with just 18% on a professional basis. 25% use it in both contexts. The use of AI in a professional context therefore remains limited and is progressing slowly.
However, the use of AI without informing one’s hierarchy, or “shadow AI”, has decreased, contrary to certain preconceived ideas (Half of workers use unauthorized AI at work and don’t want to quit– down from 68% to 51% in one year. This means, however, that a significant proportion of AI use at work is still carried out “in shadow”, with all the risks that this entails, and we don’t know whether the 17% who have stopped have simply abandoned AI or switched to an official enterprise platform.
Unsurprisingly, companies are struggling to keep up with the pace of innovation from vendors, and AI adoption is slower than AI development ([FR]We don’t need better AI, we need a better understanding of AI). Experimental uses are often first made in the personal sphere, which in no way presumes the ability to professionalize them later on, as we have seen with social networks.
Nuanced benefits for AI in the enterprise
Generative AI offers undeniable advantages in a number of use cases, which the study lists:
- Content generation: drafting e-mails, documents, articles…
- Analysis and synthesis: summarizing documents, analyzing data…
- Information retrieval: querying databases, finding documents…
- Translation: translate texts and documents into different languages.
- Customer support: chatbots, virtual assistance…
However, the promised productivity gains are far from being realized. While one study mentions productivity gains of 40% or even more, these gains are mainly valid for very precise and circumscribed tasks, such as summarizing a text, writing a simple e-mail or querying a document to find information. As soon as the task is part of a more complex, multi-step process, the benefit is less significant. This brings us back to the difference between gains at task level and at process level (AI in the workplace: going beyond augmentation to actually transform). What’s more, I’m always skeptical when, for this type of use, we only look at the quantitative dimension without considering the qualitative dimension (Productivity: what if quality was the new quantity?).
What’s more, you also have to take into account the time needed to converse with the AI, fine-tune the prompt, check and adjust the result, proceeding by iteration. In short, generative AI is particularly beneficial for repetitive and recurring tasks, at least to date.
In any case, we’re a long way from the revolution we’d hoped for.
The challenges of generative AI
Integrating generative AI into the workplace raises several major issues.
First of all, data. How can we integrate the company’s specific data without creating a new model? How can we ensure data confidentiality?
Then there are ethical and societal issues: can AI reproduce biases? How can we avoid discrimination? Will AI replace jobs? (The challenges posed by AI are not technological, but must be met today.).
There is also the question of regulation and how to frame the use of AI to avoid abuses, especially in Europe, where failing to be innovative leaders, we use regulation for quasi-protectionist purposes under the guise of ethics (The European AI Act for dummies).
And finally, as we saw in my previous article, there’s the environmental impact. Generative AI consumes a lot of energy and water. For example, building the GPT-3 model consumed 1,287 MWh of energy and 700 m3 of water.
The age of agents and more autonomous AI
The year 2024 marks the arrival of the “age of agents”, or agentic AI. These generative AIs are capable of perceiving their environment, interacting with it and executing complex tasks autonomously(Salesforce’s Vision for the Future of Work Is Agentforce 2.0 and Agentic AI in process management). A generative AI agent is a program that executes tasks autonomously on the basis of a set of parameters. It can be seen as an evolution of existing automation programs.
This evolution is notably due to the technical capabilities provided by LLMs, notably the contextual window (the set of information accessible and relevant to an AI at a given moment to generate a response or make a decision). AI agents can, for example, translate a text from a document, or request information about an upcoming meeting, by executing background tasks in a coordinated fashion.
In my opinion, this development is just in time. Talking about it recently with digital workplace practitioners, the same question came up again and again: “OK, text generation is fun, and there’s a magical side to it at first, but it’s not for everyone, and it’s going to be hard to justify its economic and environmental cost, so now what do we do to really have a productive impact on a large scale?”.
This is the direction in which agentic AI is heading, and sometimes even on things as simple as they are important. For example, during the presentation of this study at the HR Technologies trade show the day before, I saw a very convincing demonstration by LumApps, which makes it easy to design micro-applications that drive third-party business tools and workflows from the digital workplace, thanks to AI. As someone who has always believed that, in an increasingly fragmented working environment, the adoption of non-essential tools depends on interoperability (using them through a tool that is central, or making a non-essential tool central by enabling it to drive the others), this was the most interesting and intelligent thing to happen to the digital workplace in a very long time.
Conclusion
Generative AI is a promising technology, but we need to be cautious and realistic about integrating it into the workplace. Companies need to be aware of the potential benefits, but also of the limits and challenges associated with its use.
As we have seen many times in the past, the best solution is not the human or the machine, but the human who knows how to use the machine (The second machine age). Success lies in intelligent collaboration between human and machine, and therefore inan appropriation that is far from being effective in today’s business context, so that AI becomes a tool at the service of efficiency and innovation, while respecting ethical, legal, data security and environmental considerations.
Image: Gen IA by BOY ANTHONY via Shutterstock.
