Orange’s Miguel Alvarez does not seem like he would be a natural admirer of Cory Doctorow, a prominent AI critic who previously invented the term “enshittification” to describe the worsening of online platforms. Alvarez’s job is chief data and AI officer for Orange Business Services (OBS), the French operator’s <a href="https://myappsplus.com/3-enterprise-software-stocks-where-ai-automation-could-lift-pricing-power/” title=”3 Enterprise Software Stocks Where AI Automation Could Lift Pricing Power”>enterprise unit. People meeting him might expect to encounter one of telecom’s unapologetic AI evangelists. Doctorow, on the other hand, has recently been warning us about AI’s “reverse centaur.”
To Doctorow – a former guest onthe Telecoms.com podcast– a centaur is a person in charge of the technology, a human head and brain above a body of tools they control. A reverse centaur flips that, so the human is steered by the AI. “I think it’s a very good metaphor,” said Alvarez on the sidelines of the Techritory event held in Latvia last month. “You consider it as another tool, and it’s not the human at the service of the machine. That’s not the way that we see the world.”
It is of personal concern to Alvarez as a young parent, and yet he thinks the anxiety surrounding AI’s impact on currently essential skills is rather like the historical fear about calculators making people worse at math. His children, he says, are not learning rudimentary math by hand and in their heads because they will ever have to multiply four-digit numbers in the future.
“It’s because that kind of reasoning is important for the brain,” he said. “The same is true for other areas of life.” He firmly expects agentic systems, and not humans, to write most future code. But he thinks people will continue to learn core coding concepts “because it’s important for them to understand the architecture, because some of them will need to improve the systems, and because some of them will need to control the outcomes and build the tests and refine the overall process,” Alvarez told Light Reading.
For all his ruminations on the inadequacies of AI, Doctorow fears companies will use it to cut jobs. And aside from the scaremongering talk of a superintelligentSkynetthat exterminates humans, the prevailing concern has been about AI’s threat to livelihoods. Yet there is little evidence it has been responsible for mass layoffs in any sector. The USunemployment ratewas 4.2% in October, down from 4.5% in November last year, having peaked in the last 20 years at nearly 15% in April 2020, when the pandemic raged.
The latest proof comes from the Bank of England’sMonetary Policy Reportfor November. Evidence AI has had any impact on UK productivity growth remains “modest,” it said. Data analyzed so far puts the annual contribution at just 0.1 to 0.2 percentage points since 2024. Its survey shows “close to zero impact of AI on employment so far,” a finding “consistent with a range of empirical studies,” said the bank.
Alvarez would probably agree with the assessment about jobs. Orange has shrunk its global workforce by roughly 10,000 employees since the end of 2022, when the launch of ChatGPT heralded the dawn of AI in its generative form, to around 126,000 in September. But it shrank by around the same quantity in the preceding three years, implying factors other than AI bear responsibility for the decline. The teams within Orange that are getting the most value out of AI have not done it by reducing headcount, said Alvarez.
Instead, they have largely done it by shortening development times and speeding up activity. The effect is measurable in what Alvarez refers to as “value creation,” which considers cost savings, cost avoidance, revenue protection and additional revenue generated, he explained. The ultimate value creation target for Orange at group level is in the hundreds of millions of euros. As part of this, OBS had its own 2025 target of €50 million (US$55.5 million), which was achieved. This year, that has apparently been raised to €70 million ($77.7 million).
To help it achieve these targets, Orange is now exploring several hundred AI-based use cases. Thirty of those have been in production at OBS for a while, and 20 are being “industrialized” in a subsequent AI expansion, said Alvarez. Adoption of AI by the Orange workforce has, in the meantime, exceeded initial expectations. Originally, the operator hoped to have 10,000 to 20,000 users within a couple of years of ChatGPT’s arrival. Today, the number of active users across the organization exceeds 100,000, according to Alvarez.
What, then, might these various use cases cover? “One of my poster cases is a project for a new product that we were discussing how to develop,” said Alvarez. “We were able to get to a working version of the product in a couple of months, and a version we can propose to the market in a few months more.” Without AI, the entire process would have taken four to five times longer, he reckons.
On the coding side, he has also witnessed something of a revolution in the past year. Before then, software developers at OBS were using AI much as a teenager uses predictive text messaging on a smartphone – to autocomplete basic code. More recently, some developers appear to have relinquished major tasks to AI agents. “We still have some teams using it as autocomplete, we have some teams that are optimizing some aspects of the development, and we have some teams doing requirement-driven development that are automating most steps of the software development lifecycle,” said Alvarez.
Staying in control
But that level of automation, with humans barely in the loop, could seem to risk transforming OBS developers into Doctorow’s reverse centaurs. “Clearly, when you go to a model that is non-deterministic and where sometimes the capabilities are very big, you need to be more careful on what you allow it to do,” said Alvarez. One of his challenges is striking the appropriate balance between fixing guardrails and giving AI agents enough freedom to breathe and generate value. “You are giving them more leeway so they find better solutions, but as a result they could do more wrong,” he said.
The other visible danger is of heavy reliance for such a critical tool on external AI. An extreme illustration of that was provided earlier this year whenTrump orderedAnthropic to block access to its most advanced models by anyone not American. To avoid reinventing what is already available while mitigating the risks, Orange uses a combination of commercially available (and obviously unrestricted) frontier models as well as open-weight alternatives, deploying those on its own infrastructure, including on-premises.
The company has been guided by certain principles, the first being that it can choose where the AI runs “so you don’t create a public-cloud dependency from the beginning,” said Alvarez. Orange also demands that AI be multimodal. “You don’t want to have a solution that only runs in the most expensive models in the market, because they may be very good, but that will automatically create a scaling limitation.”
Agents, meanwhile, must remain under Orange’s control. In that area, it is using an open-source framework devised by LangChain, a nonprofit based in San Francisco, so that it can develop an agent once for multiple environments. “You can move it to a different infrastructure, have it run on Google Vertex, have it run somewhere else, and the agent itself – all the components of the agent that are key to how your company works – you can move them freely,” said Alvarez. “They are not bundled into some black box or platform where it is hard to make an export.”
The capability of open-weight models is still about six months behind that of the frontiers, in Alvarez’s view. The good news is that they have improved enough to handle more sophisticated use cases previously dependent on frontier models. It has allowed OBS to work on a contract-management system that required an on-premises model for commercial reasons. “We cannot use a cloud-based model to check – a frontier model to check – because for some contracts that’s not appropriate,” said Alvarez. “Now we have powerful enough on-prem models that we can use to do this validation.”
Using smaller models to achieve what would have required leading-edge technology just a year ago has clearly made AI deployment more economical. But competition between AI adopters is still propelling some toward the most advanced and expensive models. That has put others under growing pressure to do the same.
“If you are not able to follow that with your own transformation, then you’re going to have the cost without the benefit,” said Alvarez. “And I think that is the risk that some companies have – the technology is moving faster than their own ability to transform.” An AI that costs more than the savings it promises would hold very little value at all.
