AI in financial services will only deliver value if embedded into workflows that reflect how people work. Intelligent integration, not just technical innovation, will determine the future of asset management.
For all the noise around AI in financial services, remarkably little is said about the people it is supposed to support. Too often, implementation focuses on capability rather than context, offering general-purpose tools to users who operate in particular, risk-sensitive environments. In asset management, where decisions carry regulatory weight and real-world financial consequences, this disconnect can prove not only unproductive but potentially dangerous.
Michael Beattie, Global Head of Product Strategy at Linedata, is part of a growing cohort in the industry that focuses not on pushing AI for its own sake but on embedding it meaningfully within the day-to-day work of portfolio managers, traders, compliance officers, and operations professionals. The core principle is disarmingly simple: start with the workflow.
When financial institutions attempt to modernise by layering advanced technologies on top of outdated processes, the result is rarely transformative. More often, it is expensive and brittle. Beattie frames the problem in plain terms. “People try and retrofit AI where it does not solve a problem,” he says. “They use technology in search of a problem. You flip that, and that’s why I discuss it from a workflow perspective. I would like to consider something highly specific to the user’s workflow. We create these things called journey maps. It is, you know, I am Mike. I am a trader. Here are the things that I do throughout the day.”
AI must adapt to the persona, not the other way around
This persona-driven view is foundational. AI is not simply an engine for pattern recognition or predictive modelling. In regulated industries, it must interpret and assist human judgement, not replace it. The nature of that assistance varies depending on whether the user is managing risk exposure, handling post-trade operations, or scanning market sentiment.
Linedata’s approach to embedding AI began with acquisitions and partnerships but has matured into a process of iterative integration. This involves developing capabilities with a deep understanding of the user’s role, the structure of their decision-making process, and the limitations of automation in high-stakes environments.
“What we are doing now is rolling AI out to customers within our current suite of applications,” Beattie adds. “We are iterating on those workflows and user personas. That was a big topic at our recent LDX event in New York. We talked about specific use cases and how the technology fits the reality of what our clients do every day.”
That reality includes vast amounts of data scattered across formats, systems, and jurisdictions. The opportunity for AI lies in sense-making, aggregating disparate inputs, highlighting anomalies, and accelerating human insight without disrupting the chain of accountability that defines asset management.
Failing fast is not a licence to fail carelessly
The software industry’s prevailing wisdom around AI and product development often encourages risk-taking and experimentation. In financial services, this ethos must be tempered by an understanding of operational risk. Beattie is not opposed to agile methodologies. But he draws a sharp boundary around what types of failure are acceptable.
“You can fail fast, just not on those mission-critical endpoints,” he explains. “We use an agile methodology. I do not think you have to abandon the agile process with AI, but there are different areas where you can and cannot fail. Do I want to fail at executing an order or settling? No. Where I think I would want to fail is at the beginning of the process. Let me provide you with access to your existing databases through natural language. Let me think about some of the guardrails I can create.”
The term ‘guardrail’ comes up often in Beattie’s explanations. He sees it as the threshold between experimentation and accountability. It is not just about technical controls but about apparent, role-specific oversight. This is where dashboards play a critical role.
“That dashboard is event-driven and persona-specific,” Beattie continues. “It is triggered by alerts where something is actually going wrong, and action needs to happen. The power in that is setting up the correct guardrails so that it is not just someone asleep at the wheel. If it is about order routing, you set a limit. Ensure that the real-time data you are using is accurate. If orders are executing far outside that limit, that should trigger intervention.”
In this model, AI is not a black box that automates decision-making. It is a co-pilot that provides visibility, alerts, and context, all anchored in workflows that are already familiar to the user. Failing fast only works when the risk of failure has been actively contained.
Transparency and auditability are not optional
A growing number of AI deployments in financial services are being challenged not for their effectiveness but for their opacity. Regulators, clients, and internal audit teams all want to know how conclusions were reached. For firms operating in global markets, this demand is not uniform. Different jurisdictions impose different disclosure requirements. But the underlying need is the same: to explain and trace every step.
Beattie sees this as central to responsible deployment. “Everyone I talk to says, OK, I can see this being an opportunity as you implement AI, but I need to be able to explain it,” he says. “To my investors, my clients, my users, and my regulators. There are regional differences, but the audit trail is vital.”
He breaks transparency into three components: post-event traceability, real-time visualisation, and pre-emptive alerts. Each has a role to play in building trust and enabling governance. “There is the post-event piece, which is about what decisions were made and why,” he explains. “But there is also the pre-event piece, which is about what is happening now and where things are going. Then there is the real-time visualisation, charts of processing time, price movement, scatter plots across sectors. That is how people make sense of the model while it is running.”
This emphasis on visibility applies as much to internal processes as it does to client-facing features. At Linedata, AI is used to support legal compliance by interpreting investment prospectuses and generating rules for portfolio management. But those outputs are not blindly accepted. “There is a whole validation set of steps that must happen with a lawyer, with a human individual. The laws themselves change, the prospectus changes, and how we create rules must change off the back of that.”
Privacy is structural, not situational
Among the most profound differences between financial services and other AI use cases is the sensitivity of the data. While tech companies often build generalised models from shared datasets, asset managers work with proprietary, confidential, and client-specific information.
For Beattie, this makes private AI not just a preference but an architectural necessity. “I do not believe public models are the right way to go in financial services,” he adds. “Our take is that these should be private models, not exposed to the public. Each client will have their own sandbox for testing purposes. Otherwise, there is too much cross-pollination of proprietary things.”
This applies not just to externally facing tools but to internal development. Linedata maintains strict isolation between datasets, model iterations, and environments to ensure that no client information is ever repurposed for any purpose. The discipline around data segregation is not simply regulatory compliance. It is essential to maintaining client trust.
There is also a pragmatic reason to favour private AI. As financial firms increase exposure to private equity, private credit, and other illiquid instruments, the available data becomes more bespoke and less structured. Public models are not trained on this material and cannot handle its complexity. In contrast, purpose-built private models can be tuned for the specific quirks and context of each client’s portfolio.
Bias, education, and selective automation
For AI to make good decisions, it must first understand the rules. But many of those rules are unwritten, shaped by precedent, institutional knowledge, or regulatory interpretation. This is where bias becomes a concern. Not because the model is malicious but because its training data and design process may be incomplete.
Beattie sees human validation as a necessary counterweight. “We must make sure that as we train some of these models, it is a diverse set of data. There is also a level of human validation that must consistently occur. That is how we train the bias out of the system. Post-event review is essential.”
He is also clear that trust cannot be enforced solely by design. It must be earned through education. Across recent trips to Europe, Asia, and the United States, he has encountered radically different levels of comfort with AI in asset management. “The educational piece is probably most important in trust. How do these models work? What is the underlying logic? That is where trust comes into play,” he continues.
The goal, he believes, is not to eliminate the human from the loop but to sharpen the human’s insight. That means identifying use cases where AI adds real value without introducing unmanageable risk. “Not every problem requires a generative model,” he concludes. “You must ask what is meaningful, measurable, and safe. We see real opportunities in compliance, summarising documents and mapping them to rules.
Additionally, natural language access to databases is also available. You can ask questions like, ‘What is my exposure to the pharmaceutical sector over the last two months?’ That kind of workflow-specific interaction is where this really helps.”
These are not headline-grabbing breakthroughs. However, they are the foundation of a more intelligent and resilient form of asset management. The tools will get smarter, but their success depends on being rooted in how people work, not just what machines can do.
