At the MDOTM Future Intelligence Forum in Troutbeck, New York, COO and Board Member Federico Invernizzi set out what 220 conversations with asset and wealth managers reveal about where growth is going, and why most firms are building AI agents in a way that will not pay back.
Over the past twelve months we have spoken with roughly 220 asset and wealth managers across North America and Europe about what they are building, what has worked and what has quietly been shelved. Those conversations point in a consistent direction.
The next phase of growth in this industry will come from services and solutions rather than products. AI agents have removed the cost barrier that kept that shift out of reach. And a large share of the agent programmes now running inside investment firms will not survive contact with their own economics.
The firms that connect all three are the ones pulling ahead.
What an AI agent actually is, and where it returns value
An agent is AI software that goes beyond a chatbot. It pulls in context and data, and works toward a defined goal with a degree of independence. By nature it is non-deterministic, which is the whole point and also a problem.
Most agents today are built on large language models, because LLMs made natural interaction between software and people possible. We do not expect that to hold. Not every agentic workflow will be LLM-driven, and assuming it will is already shaping some expensive decisions.
Agents are being developed across every function, almost always bottom-up, by teams who want to see what the technology can do. Looking at how financial services firms are actually using them, the work falls into two groups.
The first is efficiency. Back-office processes such as KYC, reconciliation, trade matching and document handling. Every industry is automating this layer, and the business case is easily justified.
The second is growth. The places where profit and loss are made and how firms choose to engage with clients. Portfolio construction and rebalancing, frictionless portfolio distribution, adviser enablement, personalised commentary and client reporting. For investment managers, this is the group that matters, and it is the one receiving less attention.
Why growth is moving from products to solutions
The industry has done well in recent years selling higher-margin products, alternatives in particular. The larger opportunity over the next five years sits somewhere else: in the move from a product-led model to a service-led one, built around model portfolios, customised solutions and a closer relationship with the end client.
Three structural forces are pushing in that direction, and none of them are reversible.
The first is retailization. Asset growth now comes predominantly from retail investors, which fragments a client base that used to sit in a small number of large mandates. BCG puts retail AUM growth at 61% between 2020 and 2025, and argues that distribution rather than product now decides which managers grow [BCG, Global Asset Management Report 2026].
The second is demographic. Cerulli projects $124 trillion of wealth transferring through 2048, most of it to younger generations [Cerulli, via ASPPA]. HSBC expects women to control more than 40% of global wealth by 2030 [HSBC, March 2026]. The mass affluent segment keeps expanding. The client base is far less homogeneous than the one most operating models were designed around.
The third is choice. After the mutual fund era and the ETF era, an investor can access almost any asset class cheaply from a phone, usually without the expertise to assemble those parts into a portfolio. Demand for guidance rises as access gets easier.
Most firms are running operating models built for a client base that no longer exists.
The opportunity is obvious. Delivering it at scale is where the bottleneck is
Investment solutions have been a conference topic for years. Adoption stayed limited because the work was labour-intensive, so genuinely tailored service stayed reserved for institutional and ultra-high-net-worth relationships. Everyone else got a product off the shelf.
AI changes that arithmetic by cutting the cost of three specific activities.
Producing timely, context-aware content for clients and advisers, rather than a quarterly commentary written once and sent to everyone. Customising portfolios at ten or a hundred times the current scale, against each client's constraints, preferences and tax position. Analysing and explaining each portfolio to the person who owns it, in language they can act on.
Those three costs are what kept bespoke service expensive. Lower them and the economics of serving a fragmented, demanding client base change completely.
Commercially, the return compounds. Firms capture more of the value in the chain, moving from product provider to the partner the client relies on. Relationships become harder to displace, because a personalised portfolio and the service around it carry switching costs that a fund does not. Revenue diversifies away from dependence on a handful of large mandates.
The effect on people is more positive than the headlines suggest. Professionals who currently spend their weeks writing RFPs or assembling documentation hold deep experience and client knowledge. Move that capacity toward client-facing work, client portfolio management in particular, and the firm gets more from it. As Federico put it in Troutbeck: "Clients, clients, clients."
Real momentum, and real cost
Two numbers suggest this is infrastructure rather than hype. When Anthropic donated the Model Context Protocol to the Linux Foundation's Agentic AI Foundation in December 2025, it reported more than 10,000 active public MCP servers and over 97 million monthly SDK downloads across Python and TypeScript [Anthropic, December 2025]. Each download represents somebody inside an organisation building or testing something. MCP was co-founded as a foundation project with Block and OpenAI, with Google, Microsoft, AWS, Cloudflare and Bloomberg behind it, which makes it industry plumbing rather than a vendor bet. Financial services, as a data-intensive industry, is moving on it faster than most.
Spending, though, is not the same as returning. McKinsey's Performance Lens Global Asset Management Survey plots technology spend as a share of cost against revenue per full-time equivalent, and the spread at every level of investment is wide [McKinsey]. Some firms convert spend into performance and many do not. The variable is not the size of the budget.
Agents add two risks that are specific to them. They can cost far more to run in production than anyone modelled at the proof-of-concept stage. And because they are probabilistic, they make mistakes.
Why most agent programmes will not pay back
The common pattern looks productive and rarely is. A capable team gets a budget. They start from a large language model, connect a database, then a second data provider, then bring a person in to resolve the ambiguities the model cannot. Because the newest frontier model is the most capable, they use it at every step, consuming expensive tokens on work that does not need them. The proof of concept performs.
However, two problems sit underneath this.
The first is direction. These projects get shaped by whatever problem sat in front of the people who built them, not by the firm's most significant P&L opportunities. The work is real and the relevance is accidental.
The second is unit economics. Cost per task detaches from the value of the task, and because every step in an agent's work is probabilistic, errors and costs compound as the task gets more complex. The more you run it, the worse the economics look. That is the opposite of what scale is supposed to do.
Our concern is that agents become the new spreadsheet. Everybody builds one, each partly works, nobody can see across them, and the industry spends the next twenty years unpicking tools that never moved revenue or cost. Firms are still paying people to maintain spreadsheets written in 2009.
What the firms getting it right do differently
They think like architects, and they work top-down. Before choosing a tool they answer three questions. Where is the growth. What is genuinely differentiating for this firm. What has to stay human.
They deploy agents only where the economics justify it, because an agent costs money to build and more money to run. And they favour specialised solutions over token maxing: vertical tools that already understand the data structures and the problem, combined into a coherent workflow, rather than a general-purpose model grinding through an entire process.
Across our conversations, around seven in ten firms were experimenting with AI in some form, and most started bottom-up. The ones making the best progress now are frequently on their second attempt, having paid for the lesson the first time.
Reliability is where this gets serious. In a regulated business you cannot take an LLM at its word. Our approach is to build a core that stays deterministic where it needs to be, with rules that verify the quality of what comes out, and connect language models only at the points where flexibility earns its place. A plain-language instruction to adjust 200 portfolios becomes a set of rules, executed by specialised optimisation models, then checked back against the original intent.
Without that structure, the familiar pattern follows. A firm tests an agent on a handful of cases, likes what it sees, and then will not put it near 10,000 portfolios, because nobody can say what it will do there. The test succeeded and the deployment never happens.
The ambition we are working toward is an end-to-end portfolio management workflow. People make the investment decisions and remain accountable for them. Machines tailor those decisions to tens of thousands of client profiles and constraints, explain the resulting portfolios, and feed client responses back into the loop. The human stays in the picture rather than in the loop.
A window that will not stay open
Very little of what investment management needs has to be invented. The pieces already exist, and the opportunity lies in assembling them at the right moment. What the industry has been searching for is within reach using technology that is available today.
The practical consequence is that firms no longer have to wait for a multi-year data migration before they can build anything. If the prioritisation work is done properly, you can invest in a small number of areas with fast, measurable impact rather than rewiring everything first.
That creates a problem for incumbents. For decades, scale was the moat. More AUM meant more resources, more data, better negotiating power and more technology leverage. That correlation is weakening. Smaller, more focused firms that move early are capturing gains out of proportion to their size, and we see European wealth managers already growing faster than larger competitors. Large organisations carry more complexity and more stakeholders, and risk adapting too slowly.
Within this ecosystem, MDOTM is building the layer where portfolios are hosted, managed and optimised, and opening our MCP tools and agents so that other agents, workflows and people can work with those portfolios directly.
What is left to decide is not whether any of this happens, but who captures it. Treat agents as a run of disconnected experiments and you end up with a portfolio of expensive tools and an unchanged cost line. Act as an architect, start from where the growth is, decide deliberately what stays human and choose specialised building blocks over brute computing power, and a mid-sized firm can move faster than a much larger one.
Much of what we described at Troutbeck changed in the preceding three to six months, and organisations are not adapting at the pace the technology is. The window is open now. It will not stay open indefinitely, and the firms designing carefully today are the ones most likely to define what investment management looks like next.
The Future Intelligence Forum is MDOTM's invitation-only event series for senior leaders in asset and wealth management. To discuss how Sphere supports portfolio customization at scale, get in touch.
