What is the role of humans in an AI world? Lessons from Hiring Three Data Science Students from LSE

Frederic Kalinke details Exactimo's experience of hosting three LSE Department of Statistics data science students.

There is an existential debate consuming boardrooms, lecture halls and dinner tables right now: what is our role in the workplace when artificial intelligence gets smarter by the day?
But I believe this zero-sum framing is wrong and humans are actually more important than ever in the world of work.

This past summer, my company Exactimo hired three LSE data science students to help build real products with AI’s assistance and it has only strengthened this human-first belief.

Why humans sit at the centre, not the edge

Let's start with what a Large Language Model (LLM) actually is. Strip away the interface and the slick marketing, and it is a prediction machine, whose objective is to take a user's prompt and based on a vast amount of training data guess the next word in a sequence until the output is complete. This is a remarkable capability but it is not the same thing as knowing what matters.

The bottom line is that AI is only as helpful as the prompt it is given and the information it has access to. Somebody has to decide what problem is worth solving, what inputs the AI needs to have, what constraints apply and whether the answer that is returned is helpful. None of that is a technical step. All of it requires human inquiry and as you will see, our interns discovered this for themselves when getting their hands dirty using AI-powered coding tools.

This is why we always say AI should be mastered in both senses of the word. Mastered in the sense of understanding, but also in the sense of guiding and controlling LLMs with the right information and context. Building both types of mastery across an organisation is what is meant by developing a ‘culture of knowledge’, which is the foundational step in any AI enablement process and is far more about people than technology.

So which human skills matter?

I use a framework I call ARCH:

· Attention to detail. Humans must spot the flaws in inputs they give and in the outputs they get back. Plausible and correct are not the same thing.

· Resilience. Humans must embrace change as the technology shifts, because it will, repeatedly.

· Curatorial. Cultivate taste like a curator in a museum. Filter noise, select what matters, and present things with impact.

· High agency. Be proactive and do things. Experiment. Action creates data, and data is how you learn.

The concept of the arch itself is also significant here given that arches are one of the strongest geometric shapes, helping secure structures against existential threats. Just ask the Romans.

Certainly a fitting metaphor for our times.

What the interns learnt

Advait Bijadi and Felix Wendeling joined us to build a data pipeline that extracts job and career insights from across the internet. Live scraping, structured extraction using LLMs, a database and continuous automation to keep it running.

Aarin Bhatt worked on a different question: how AI is reshaping the finance industry. From his research, he then helped compile a curriculum that provided an overview of the finance industry, including practical exercises such as a ‘market-sizing agent’ (an automated AI tool which calculates the revenue potential, customer reach, and growth trajectory for a product or service) and a ‘Monte Carlo pricer’ (an algorithm that calculates the fair value of complex financial assets such as stock options or derivatives by simulating thousands of possible future price paths), validated against Black-Scholes, both with AI’s assistance.

And within the first day of starting both projects, it was clear to the interns that human inquiry was the foundation before any technical work could begin. As Felix says:
"What surprised me was how much preparation goes into these big projects. Most of the work is actually outside the coding and the algorithms. It's in the preparation, whether that's looking at the research question, setting hypotheses and understanding constraints."

So the art of curation sits at the heart of a data scientist’s work. Yes, the model is interesting but to make it useful, you need to make the right decisions about everything that sits around the model: the cost you are willing to pay per call, the rate limit you will hit at 3am, the token budget, the schema that stops the model inventing a field. That is curatorial judgement, deciding what matters and refusing to be seduced by the flashier parts of the process.

For Advait, two specific lessons came out of the build, and both are attention to detail in its purest form. First, stopping the model from hallucinating. When you use an LLM to extract structured data, strict schema design and constrained outputs are not tidiness, they are what makes the system trustworthy enough to parse. Second, scaling affordably. Batch processing and disciplined API usage were the difference between a pipeline that worked and one that worked at a price that didn’t make it worthwhile.

Another lesson was about creating something valuable. A pipeline is only useful if someone uses it and raw extracted data alone is not an insight. Turning it into something a decision-maker can understand and get excited by is how you convert engineering into value, and it is as much a presentation skill as it is a technical one.

As Advait says:

Whilst I learned a huge variety of technical skills, such as database management, GitHub Actions and how to use agentic coding tools, the soft skills around presentation and persuasion were equally important, as I had to deliver updates weekly to the executive team, which allowed my team to iterate and course correct accordingly.

For Aarin, attention-to-detail and high agency were essential when designing learning experiences for students that wanted to use AI in finance-related workflows. Specifically, when crafting automations, it’s important to keep ‘humans-in-the-loop’ and ensure first-principles thinking is not delegated to the LLM, as Aarin says:

Early on, Frederic and I agreed the exercises had to force ‘doing over reading’. You can read a market-sizing method and follow along, but you don't actually learn it until you've written it yourself and found the one place it breaks. That's why nobody gets handed a prompt to paste in the exercise, they get a blank file and have to write the method out themselves first.

A note to employers hiring humans

We hired all three interns by advertising the roles through LSE Careers. It was a very simple process - we wrote a job description that described the actual problem rather than a list of requirements, posted it, and had talented and well qualified candidates respond in hours.

I would encourage other employers to do the same, with one specific bit of advice: give them a real problem. The temptation with early-career hires, particularly at a small company where the cost of a mistake feels high, is to scope the work down until it is safe. We did the opposite, and what we got back was working tools and three people who developed enormously over the summer.

It was clear that all three LSE students arrived with the analytical habits already in place. What they had not yet had was the chance to own something end-to-end in a real workplace, under real constraints, with someone reviewing their work weekly. Give them that and you will be surprised by what comes back, particularly if you are intent on building arches and nurturing the things that matter: detail-orientated, resilient, curatorial and high-agency humans.

By Frederic Kalinke, founder of Exactimo.