Federico Invernizzi, our Chief Operating Officer and Board Member, joined Rupal Agarwal, Managing Director at Bernstein, for the Gen AI in Asset Management: Innovator Series. Rupal’s questions covered why bespoke portfolios are hard to scale, where generative AI fits in the investment process, and what comes next for AI in asset and wealth management.
Clients increasingly expect portfolios built around their own needs. Asset owners are asking more of their asset managers, and in wealth management, individual clients and client groups now expect bespoke solutions. With do-it-yourself options a click away, Federico argued, managers who want to stay relevant beyond low-cost ETFs have to offer active advice, active model portfolio management and bespoke portfolio construction.
Why are bespoke portfolios so hard to scale?
Most processes behind bespoke portfolios were built for a handful of them. Optimisers, spreadsheets, presentations and client intake forms sit in separate places, connected by hand to third-party delivery systems. That holds at low volumes. As portfolio numbers multiply, the same process becomes slow, error-prone and hard to oversee.
Rupal opened by asking what we set out to solve. Federico pointed to the teams that feel the strain first: investment solutions teams at asset managers, and discretionary portfolio management teams in wealth management and private banking, each running many moderately sized portfolios from one investment view. “They run portfolios, the number has grown because they’ve had some success, and things break or become ungovernable,” he said.
Asked how this compares with the optimisers clients already use, Federico was clear that they still have their place. “It’s a fantastic machine that does exactly what it’s told,” he said, and for one or a few large portfolios, iterating on every constraint is time well spent. A business running 300 smaller portfolios faces different economics. It needs a reliable tool that works at scale, accepting a little less fine control in exchange for far more in workflow and monitoring.
Where does generative AI fit in the investment process?
On top of a deterministic core. In Sphere, our platform, a large language model interprets a request made in plain language, then calls proven quantitative tools to do the work: the optimiser, the risk model, portfolio checks and controlled reporting. Performance, risk and attribution figures come from established quantitative methods.
Rupal asked which parts of the platform use generative AI and which use predictive machine learning. The portfolio checks show the split. Ask the platform whether a portfolio is in line, and it applies a preset list covering constraints, benchmark alignment, tracking error, and the dates of the last optimisation and rebalance, then reports any breaches. The portfolio manager can adjust a limit, propose a couple of trades and backtest the result before acting. The same interface can read a mandate document, set up the portfolios it describes and ask for anything missing, such as whether the initial composition should track the benchmark or the latest market view.
Machine learning sits further upstream. Our research uses neural networks and market models to detect risk regimes and generate overweight and underweight signals, and those estimates inform the risk expectations behind the optimiser. “We try to limit creativity to a text box,” Federico said. Once approved, that text becomes a fixed element on the platform.
How do you keep AI-written commentary accurate?
Limit what it can draw on, and check before it writes. Storyfolio, our controlled generative AI reporting engine, draws only on data within the platform. Its explanations come from rationales the investment team records when adding a security to a universe or setting a market view. Where it cannot explain a position, it describes it.
Rupal pressed on how a summary stays aligned with the portfolio manager’s actual reasoning. The second safeguard matters most at scale. Bespoke portfolios rarely follow the house view exactly, whether because of a client request or an added constraint. Before anything reaches the language model, Storyfolio checks whether each portfolio’s holdings and trades match the market view. Where they diverge, it sets that part aside and writes only about what the data supports. The output follows each firm’s own templates, fonts and branding.
What does true personalisation require?
Client-level inputs and an understanding of each portfolio’s history. Personalisation starts with settings, constraints, allocation bands, exclusions and ESG preferences. Federico described the next step as learning from how each portfolio has been traded, so that recommendations respect decisions the adviser and client have already made.
Rupal asked whether personalisation means the system learning from how people use it. We do not train models on client data or usage; customisation happens at the client level. Federico’s example will be familiar to anyone in private wealth. Run an optimisation on most platforms and it pulls the portfolio back towards its model, trading in and out of positions. Yet the adviser may have just bought a fund at the client’s request. “You can’t just sell it and move on,” he said. Respecting that history is where portfolio customization becomes real for the client.
Does efficiency improve investment outcomes?
Indirectly, in Federico’s view. When a team has to trade 200 portfolios and execution takes days or weeks to confirm, markets move before the work is done and portfolios fall out of sync. Automating that work returns time to investment decisions. “Less operations, more insights,” he said.
Rupal asked whether the value comes from better decisions or from time saved. Federico was careful about the limits of the claim. The technology supports the process, and performance belongs to clients, who decide how to set it up and use it. His personal view, which he acknowledged would be hard to prove, is that running portfolios tightly, monitoring proactively and realigning promptly brings a discipline that tends to show up in results.
What is the next frontier for AI in finance?
Pairing what large language models do well with tools built for reliability. Federico’s view is that the best design uses a large model to understand each request and route it to the right specialist tool, such as an optimiser or a risk model. Running everything through a frontier model would be, in his words, “unreliable and very expensive”.
Sphere by MDOTM helps asset and wealth managers deliver hundreds of tailored portfolios from a single investment view, with AI built on a quantitative engine they can trust.
The Gen AI in Asset Management: Innovator Series is hosted by Bernstein. Our thanks to Rupal Agarwal and the Bernstein team. To see how Sphere supports portfolio customization at scale, book a demo with our team.













