Global conditions are forcing organizations to maintain growth with ever-diminishing means to do so. GenAI promises a leap in efficiency to come to everyone’s aid. However, AI has to be about much more than that, we hear from Capgemini.
We attended the Capgemini Innovation Day in the Netherlands, an event aimed squarely at “Transforming to a Digital & Sustainable Future.” Right off the bat, that has all the makings of a huge project with countless components, within which GenAI cannot be separated out. The technology has a double life: one as a glutton for already limited resources, and one as a star performer of the next transformation of our working lives. How can we reconcile those twin views?
AI shake-up
Jeannine Peek, CEO of Capgemini Netherlands, helps bring us closer to the answer. “It is true that digitalization requires energy, AI does as well. I see it as a task for the digital sector to become more sustainable and more efficient in the future.” Many lessons have already been learned and are still being turned into action: the centralization of compute by using the cloud, bringing data (back) to on-prem, opting for domain-specific solutions and avoiding vendor lock-in or being stuck with legacy.
Any IT discussion will turn monetary if you let it. Peek wants to counterbalance this. “It’s never good to run at a loss, but sometimes you have to sacrifice a percentage point of growth so you can be successful in the long run.” Especially in tough economic times, it is tempting to insist on cost-cutting and short-term thinking. Peek explains why this can actually put your organization in peril further down the line. Capgemini is discussing organizations’ AI approaches and the goals they have in mind with the tech. Any company that was’t having conversations like these whenever technology paradigms shifted, is usually out of business nowadays.
There’s an appetite for AI-generated work, it turns out. First of all, sometimes it’s effectively a requirement: as an organization, you may never find the manpower you need, for any amount of money. In addition, Peek argues that AI must revolve around an individual. What mind-numbing or easy part of a job can be automated? That improvement, somewhat achievable for decades but now a lot more tangible, can deliver higher quality of care from medical professionals and makes software developers think earlier about tricky bits of code and architectural choices.
But then we’re talking about continuity. What disruption is GenAI actually delivering? One cannot count on AI-generated labour right now without first evaluating the outputs each time, Peek believes. The frameworks of what a society finds acceptable to hand off to AI are already somewhat visible. The European Union, “good at regulating” as Peek puts it, is at the forefront of shaping these ethical and legal boundaries. The AI Act, she says, is a good start, even if government should not be a control freak. The goal should be to cultivate clear rules of the game, Peek concludes. “Those who violate trust reduce the capacity for further innovation.”
Nevertheless, many jobs are inevitably going to disappear because of AI. Perhaps there will be new ones to replace them, ones which do not currently exist, but the initial disruption by AI will cause problems no matter how you slice it. Again, Peek emphasizes that this is nothing new. “We used to have blacksmiths making horseshoes all over the place. Now, there are hardly any of those now.” For example, research by the Dutch economics board SER, of which Peek is a council member, already describes the dangers of AI for the creative sector. The impact on other industries also passes muster, such as changing customer service. “Customer response to chatbots is getting better,” Peek appoints. “You get a standardized and good answer more often now.”
At times the human factor can be withdrawn without much upheaval, even though organizations should be careful about doing so, however tempting it may be to ambitiously deploy GenAI right this second. Sustainable AI deployment requires an emphasis on its long-term impact.
In this, it is also up to the government to lead by example when it comes to the long term. Peek finds it unfortunate that the local Dutch government is jumping from topic to topic, oftentimes suddenly abandoning the previous issue. Coincidentally, that already applied to a key dtive for sustainability: energy transition. Whereas the government is busy, The Hague is now preoccupied with entirely different issues, Peek reminds us that while issues such as this one may disappear from the radar, they do not lose their relevance.
Scarce resources
Another simple fact is that GenAI in its current form cannot be scaled up to the entire world. If we were to all run LLMs through hyperscale data centers and fire off countless prompts at them, those servers would very quickly tap out. This is something that Pascal Brier, Group Chief Innovation Officer at Capgemini, acknowledges. However, he sees plenty of opportunities for further progress.
“It is always difficult to judge a technology without having a perspective on what it brings about,” Brier argues. GenAI, he says, is “clearly part of the problem as well as part of the solution.” For example, prompting models in all sorts of ways is extremely intuitive and full of potential. But what if you don’t produce the same carbon footprint each time you prompt?
Capgemini RAISE (Reliable AI solution engineering) has worked on a solution that addresses this, which is to reuse AI outputs. In essence, the process relies on a dynamic database based on actual usage. Call it a kind of FAQ 2.0. If staff continuously ask questions about one particular piece in the knowledge base, but with somewhat different wording, RAISE’s solution can recognize that. As a result, the chatbot will initially offer a general, pre-calculated answer. If that fails to provide enough information, there is always a new output to generate. Brier thinks an intermediate promoting step like this will already save a lot of energy consumption.
Capgemini is also considering alternatives to GenAI. One such alternative is the technology behind liquid neural networks, an AI technique that operates much more efficiently than GenAI models. We wrote about it back in 2023: much like a biological brain, it operates on the basis of neurons occupying various functions within a network, which can even help build a car’s self-driving ‘brain’, for example. It has not yet had its “ChatGPT moment,” if you will, but Capgemini is patient.
Brier admits that the energy issue is not just about optimizations. After all, energy is becoming more expensive and scarce. At Capgemini, CO2 emissions have already fallen from 600,000 tons p/y prior to Covid to 200,000 p/y nowadays. Similar reductions are impossible for some sectors, such as heavy industry. As a result, we will all be requiring more energy sources. “We will need wind, we will need solar. I also think we will need nuclear”, Brier says.
As for nuclear power, Brier believes our society is still stuck in the social acceptance debate of the 1970s. “We need to restart the social debate on this on a solid foundation.” As for the alternatives, he does feel somewhat disappointed with the progression of solar and wind power over the past few decades. It leads to even greater motivation to utilize multiple sources, not betting the farm on one or two.
When AI?
It is also important that AI not be used indiscriminately. “With AI, you have to start with a business problem,” says Brier. “Do you need it and why? What is your organization doing today that could be done much more efficiently? What business process would have a much greater impact if it was supported by AI?” And finally, a “blue ocean” suggestion: “What can’t we do now and with AI we could?”
Organizations must also keep their expectations in check. Brier refers to the supposedly immense efficiency gains AI assistants promise for programmers. “There’s a big difference between generating code and building an application.” One example: AI-driven coding can be, say, 50 percent faster than unassisted programming, but that process itself only takes up 20 percent of the time one spends on a project as a programmer. Consequently, we’re not talking about a 50 percent acceleration, but one consisting of a 10 percent gain. Still meaningful, but not as overwhelming as it seems.
Capgemini itself already uses AI regularly. In addition to having already completed 1,000 projects with GenAI at organizations, GenAI models help its employees with code generation, code reviews, building tests and preparing for annual reports. Sometimes Capgemini uses an LLM from large parties like OpenAI or Anthropic, but regularly the company chooses small models trained on domain-specific information and designed to accomplish highly targeted tasks.
Don’t commission
Although Capgemini works closely with organizations, it does not impose any choices on them. “We will never tell a client that they have to put all their eggs in our basket or someone else’s,” Peek explains. Brier also stresses that a quarterly look at progress with the partner in question to improve an ongoing project. Regularly, the AI playing field will have shifted by then.
We see this during Innovation Day as well. Take Scotty Technologies, aimed at recruiters and HR, which Capgemini partners with. Scotty’s AI chatbot (not entirely surprisingly called Scotty AI) is already being deployed by PostNL, the Dutch national post service. Scotty AI, operating within WhatsApp, brought candidate satisfaction to 91 percent and delivered cost savings of 75 percent for the postal company. It is an example of a startup with a specific tool that Capgemini is capitalizing on.
Peek points out that startups are necessary for a sustainable ecosystem. “They are more agile in terms of finding solutions to problems than we can ever be,” she says. That includes the larger clients Capgemini works for, which can benefit from emerging AI specialists through its partner. This keeps organizations equipped with the most up-to-date technologies. In a world that demands more and more from companies with fewer and fewer resources, that’s a crucial advantage.
Also read: OutSystems introduces GenAI tool AI Builder for low-code apps