The Quants Should Have Led on GenAI. They Followed.

Why the discretionary side moved first, why the quant hesitation was rational and wrong, and where the role goes next.

If anyone in finance was built to adopt generative AI first, it was the quants. They didn't. The discretionary managers, the side that doesn't build the models, got there first. That inversion is the most interesting thing happening in the industry right now, and it rewards a close look.

Last week at RavenPack's Exponential Summit in London I sat on a panel called "The Future of Quant Investing," with Peter Hafez, RavenPack's Chief Data Scientist, and Abhijeet Gaikwad, founder and CIO of Agami Capital. Peter Hafez framed the field around three sources of edge: technology, data, and intelligence. His argument was that with AI the first two are becoming table stakes, while the real disruption has moved to the third, intelligence. I want to take the inversion underneath that head-on. Why were the people built to use this technology the slower ones to pick it up?

It looks like a paradox. It isn't. The discretionary side moved first for a simple reason, and the quants hesitated over a problem they had already solved in another form.

The translation tax

The discretionary bottleneck was code, but not their own. Every time a PM needed an answer from data, the question had to be routed through a quant who could write the query, then queued, answered, and translated back into the language of the decision. That is the translation tax. By the time the dashboard came back two days later, the market had moved and the answer was worthless. I wrote the mechanism up in detail after my December keynote, so I will keep it short here.

What matters for this argument is what it explains. Natural language removes the middle layer. The PM talks to the data directly, in plain English, and gets the answer in the room. That is a workflow discretionary managers never had. Of course they moved first. They were not swapping in a better tool. They were getting something they could not do before.

The quant hesitation was rational

Quants did not lag because they are slow. They lagged for two reasons that are, on their face, sound.

The first is about trust. A language model on its own fails at things that look trivial. The week of the panel I asked a current frontier model how many days of the week have the letter D in their name. It answered two. The same failure shows up on a finance task. Ask a model on its own for a company's latest earnings-call transcript and it tends to miss twice: it hands you a call that is not actually the most recent one, and it mislabels the period, calling a February release "Q1" when that is not the quarter that was reported. It is guessing at the calendar instead of checking it. A quant sees that and draws the obvious conclusion: if I cannot trust this thing with the basics, why would I hand it a risk model?

Right about the symptom, wrong about the cause. The model on its own is not a calculator and was never meant to be. The failure is not a failure of intelligence; it is asking the intelligence to do by hand what it should be doing with a tool. Wire in the tools, the ability to run code, query a database, call a calendar, and it stops guessing and starts reasoning about which tool to call. A good part of the job, for those of us building in this space, is showing people what these systems actually do once they are connected to the things that make them exact.

Non-determinism is a problem quants already solved

The second reason is the serious one. Run the same prompt twice and you can get two different answers. For a discipline built on reproducibility, that is close to disqualifying. I understand the reflex. I also think it is overstated.

Start with the practical: turn the temperature down and the output gets far more stable. But the deeper point is that quants have lived with this exact problem for decades. Think about portfolio optimization. Convex optimization gives you different solutions from different starting points, and from small changes in the inputs. A little estimation error in the expected-return vector or the covariance matrix can move the optimal weights a long way. The field did not abandon optimization. It constrained the weights, applied robust optimization, and resampled across perturbed inputs, averaging toward something stable.

I spent years researching techniques to make portfolio optimization more stable, so I know both their promise and their limits. Every one of those techniques applies directly to language models. Run the same query several times and take the majority answer. Constrain the domain. Those tricks help, but the decisive move is not statistical, it is tooling. Ground the system in good data, and give it deterministic tools for the parts that have to be exact. If you want your portfolio volatility, you do not want "somewhere between six and eight percent." You want the number. A well-built agent does not guess it. It calls the tool that computes it and returns the exact figure. Non-determinism in the reasoning layer does not have to mean non-determinism in the answer, as long as the answer is grounded in tools and data.

What happens to the quant

So does the job survive? Five years out, do we still have a title called "quant," or do quant and discretionary collapse into a single "investor"?

My view: the title survives, but the work behind it changes beyond recognition. It stops being signal research at scale, the way it has run for thirty years. It bends toward a role I have written about before, the Context Engineer: part investor who knows what matters for markets and when, part librarian who can organise a firm's knowledge and make it machine-readable, part architect who knows how these systems work and how to ship them. The edge moves from who can build the model to who can give the model the right context.

That is the real convergence. Not quant tools replacing discretionary judgment, or the reverse, but the two skill sets fusing into one operator who does both halves well. AI is the piece that finally makes the fusion practical. I went deeper on the Context Engineer in that December post; the panel was where the convergence question finally got asked out loud.

Where this leaves a lean team

If you run a small investment team, none of this requires a twenty-person quant bench or a seven-figure infrastructure budget. The leverage is in agents that sit on top of the stack you already run, your Excel sheets, your broker files, your Bloomberg feeds, your existing Python, and take the translation tax off your desk. The PM talks to the data. The risk numbers come back grounded and exact. The judgment stays where it belongs, with the investor. That is the version we are building at Medusa for lean funds.

The quants did not lose the first round of AI adoption because they were outmatched. They hesitated at a hurdle they had cleared once already, in a different domain. The teams that see that, and wire the tools in properly, get the upside without the bench.

So, five years out: do we still have a job called "quant," or just investors who finally have the whole toolkit?