The new hot job in AI: forward-deployed engineers

Palantir wasn’t successful simply because it built cutting-edge software. It also found talented, out-of-the-box thinkers willing to be parachuted into unfamiliar territory with very little direction.

This has made the company’s alumni a hot commodity in Silicon Valley. The venture firm Sequoia Capital ranks Palantir experience as the top pedigree for startup founders, for instance.

And it’s why one of France’s hottest startups — H Company, founded by top AI researchers at companies like DeepMind — recently hired former Palantir executive Gautier Cloix as its CEO. Cloix told Semafor on stage at the RAISE event that he plans to implement the forward-deployed engineer strategy he learned at Palantir to help turn raw research into working products. “It’s a good thing to have the best technology. It’s definitely not enough to build a successful company,” Cloix said. “The enterprise world is pretty complex.”

Most companies have some kind of chatbot at this point, but there is no out-of-the-box solution for the powerful automation that AI has promised. Startups have figured out that, in order to build what companies actually want, they need to copy the Palantir playbook.

The top foundation-model companies are also embracing the strategy. OpenAI’s forward-deployed engineers, for instance, are charged with turning “research breakthroughs into production systems,” according to a job posting. Anthropic’s will “drive the adoption of frontier AI by developing bespoke LLM solutions for top enterprises.”

What makes a good forward-deployed engineer is changing with the generative AI landscape. As the cost of software plummets and AI gains new abilities, companies can build increasingly bespoke internal software.

In one sense, a forward-deployed engineer is a kind of business consultant, but without thinking like a business consultant. Founders say they want people who are not afraid to ruffle the feathers of executives inside the companies they serve. Their job is partly to help companies disrupt themselves, which requires creativity combined with technical acumen.

In the wake of the ChatGPT moment, there was a rush to develop AI models that were “trained” for specific tasks. Thousands of them emerged. Many, like Meta’s popular Llama family of models, were free to use and small enough to run economically and on local servers.

Companies then began “fine-tuning” popular, open-source models like Llama and DeepSeek on proprietary company data to create more personalized versions. But that technique only got companies so far, especially when we began moving to “agentic” AI models that need to take important actions on their own.

Today, there’s another paradigm shift underway. With a technique called “reinforcement learning with verified rewards,” models are taught to aim for a specific goal and then trained on simulations to find the most efficient route.

But that’s not how general-purpose large language models worked in the past. Instead of aiming for a goal, they predicted what would happen next. It’s partly why chatbots, if they start off in the wrong direction because of a poorly worded prompt or some quirk in the training data, will just continue along that wrong path forever.

Ultimately, the trick will be to marry the natural language capability of large language models with a goal-oriented approach.

It’s likely this is just another stop on the way to artificial general intelligence or superintelligence. At some point, frontier models like the ones made by Google, OpenAI and Anthropic will get to a place where they can reliably do almost any rote digital task without any additional training. It’s also possible that all of the customization going on now will generate some of the data necessary to achieve AGI or ASI.

Forward-deployed engineers, the connective tissue binding AI researchers with the real world, are the point at which all of these new techniques will be carried out and tested.

This thought had been kicking around in the back of my head for a while, but it really crystallized in Paris, where I talked to startup founders, public company CEOs, and investors, both publicly and in private. That was especially true when I interviewed Cloix, who’d just moved to Paris for his new H Company job, onstage.

There was a lot of speculation about H Company’s future when three of its co-founders left the company shortly after it was established about two years ago. They were some of the sharpest AI researchers in the field, and it looked as if the company had lost its primary asset.

But research alone does not make a company these days. DeepMind, OpenAI and Anthropic started as pure research efforts, but that’s not what they are today. It’s Cloix’s job to take the nucleus of AI know-how at H Company and use forward-deployed engineers to create a feedback loop between its corporate clients and its AI researchers.

There are going to be huge new breakthroughs inside corporate and academic AI research labs, but a new wave of forward-deployed engineers will determine, to a large extent, the adoption rate of those breakthroughs. And that is where humanity will most feel the impact of AI.

From one vantage point, the need for forward-deployed engineers is a tacit acknowledgement of the limitations of AI. Despite promises that the cost of software will go to zero, that models will be able to reason as well as humans, and other projections, it’s just not possible to rely on general purpose AI models for anything truly important today.

But that could change within a few years, and then all of the massive investment in implementation today becomes, essentially, technical debt. As Andreessen Horowitz argues in this piece, the tradeoff between “margin and moat” is ultimately worth it, but it is a tradeoff. Companies are giving up scale by going the bespoke route.