There’s a lot of hype around agentic AI, but also a lot of questions that need answering. Where is it useful? How do you deploy it successfully? What does it mean for middle office teams?
TSAM London was the perfect place to ask these questions and more. Joining Duco’s James Maxfield on stage to discuss the agentic playbook for the middle office was GIB Asset Management Chief Operating Officer Marc Berryman.
Berryman has two decades of experience across firms such as St James's Place, Credit Suisse, UBS, and several UK banks.
Maxfield and Berryman covered a lot of ground. Their conversation touched on everything from data governance strategy to the evolving relationship between buy-side firms and their vendors.
Here are their key insights.
Benefits of AI in the middle office: Going beyond just speed
Many tasks in the middle office remain manual, due to the highly complex nature of the data teams are dealing with. The explosion of private markets is only adding to this, given the bespoke and often unstructured nature of the data.
"I think the danger initially is the assumption you make that, because it's so bespoke and everything is quite different and manual, that you can't really use AI," Berryman said.
His view is different - in fact, he sees AI's potential in this space as particularly exciting. Newer, more complex models are giving middle office teams the ability to extract data from documentation more readily. They can pull out the salient points that would otherwise require significant manual effort.
But extraction is only part of the picture. AI can also help teams assess more complex valuation scenarios and support judgement-based decisions.
"The magic for me is going to be where it can help you distill that data quicker, help you assess some of those complex scenarios, and help you make some of the more judgmental points."
This will help to conquer the ‘last mile’ of automation, Maxfield said, “removing the legwork and pushing the decisionmaking”.
Why AI pilots fail, and how to fix them
For AI to be effective anywhere - whether dealing with private markets or a more traditional middle office use case - you have to implement it properly.
Maxfield noted that many organisations struggle to move AI from proof of concept to operational reality. So what does good execution actually look like?
"I've seen plenty of AI pilots that look very impressive in isolation, but they never survive contact with real operations,” Berryman agreed. “That's usually because the person that ends up using this tool wasn’t involved in some of the early design stages."
The fix, he said, starts with involving end-users from the outset. It’s well-established that user engagement is a key factor in the success of technology adoption, so involving them in the design phase helps overcome a key blocker from the start.
As well as seeking input from the end-users early on, Berryman said a successful deployment involves accepting that the aim is not to deploy a finished product straight away.
Operations teams need to outline what good looks like, and then vendors and technology partners can work with them to build the solution. The key is accepting from the start that iteration is part of the plan, not a sign of failure.
Creating a culture of AI adoption
So, employee engagement is critical in shaping AI solutions and then ensuring their adoption. What can firms do to create an environment where their teams feel safe to experiment with AI - and excited to do so?
Berryman described several practical steps that GIB Asset Management has taken to embed AI into the wider organisation.
“We've created an AI club. It's open to all employees of the organisation for them to bring ideas about AI, but also to learn about some of the tools we've deployed. A lot of people have these [tools] available to them, but they don't use them to the full extent possible. So we use the AI club to remind people of what's available [and] share good ideas.”
Maxfield observed that many organisations struggle with the assumption that everyone should already know how AI works. Creating a safe space to explore and learn is therefore a powerful first step.
The second is embedding AI data engineers within the investment team. Their role is to identify problems firsthand and then find ways to automate them. This brings AI capability closer to the operational reality, rather than keeping it siloed in a central Technology function.
The third step is making AI part of people's performance objectives. Employees are asked to look for ways to automate tasks and use AI in their day-to-day work. This gives them a compelling reason to experiment and engage.
To build or to buy? The answer depends on what you need
Outsourcing has been a common transformation lever for COOs over the past decade. But AI is prompting many firms to re-evaluate those decisions. Is it time to bring capabilities back in-house?
Berryman's view is nuanced. He sees a clear evolution in what automation means for the middle office. Ten or fifteen years ago, automation was about doing existing manual processes more quickly. AI changes the equation because it can support judgement and tackle more complex tasks.
But the build-or-buy decision is not one-size-fits-all. It depends on the specific service.
"Where do I think I bring that unique perspective or unique position on something? But where does that partner also bring the scale and the breadth?" Berryman said. "In some cases, I want to have a big service provider provide something, because they benefit from the scale, the experience, the operational resilience, which I can't bring.”
“But, equally, I don't want to have no AI development or build capability within the organisation."
How has shopping around for AI vendors changed compared to sourcing other services? Maxfield asked. Berryman described a shift in mindset.
“I've spent a lot of time over the last few years talking to so many different providers of services, but I'm really open-minded because I see it as a collaboration,” he said.
"I have a problem at the back of my head I'm trying to solve, but I haven't seen everything, so I'm actually looking for vendors that bring that breadth, and also want to challenge me a bit and bring different perspectives."
Berryman added that he’s not just looking for off-the-shelf solutions. Customisation is a desirable feature of AI tools.
"I’m increasingly expecting that vendors bring me the start of something, and through a conversation, we will evolve to something that works for both of us."
Curating the AI toolset
Maxfield was curious to know how firms can strike the right balance between speed and being thorough when it comes to sourcing AI technology. There are so many tools to evaluate that there’s a “procrastination risk”.
“There's no one-size-fits-all there, but you have to make some decision,” he said. “You can't be endlessly having sourcing discussions, otherwise you'll never do anything. You want to be agile and making the most of the market, but you need to have a path that you execute on, which is not always easy to manage.”
Berryman described a deliberate approach where GIB Asset Management plans to curate a shortlist of tools, explain what each one does, and then assess their usage and effectiveness over time.
“The business doesn't want to have every possible AI tool out there, because a lot of these models aren't really useful for every purpose,” he said.
That list is iterated on constantly, depending upon how existing tools are performing and the value offered by new tools. For example, if an existing tool stops delivering, it gets removed. The goal is to give the business direction without overwhelming it, by narrowing the scope of what they consider to a few of the most impactful tools.
And the metrics for success go beyond cost. "It's very much about, 'What are the benefits to our employees, our clients, the nature of what we do?'" Berryman explained. "If people enjoy using it and there are tangible benefits in terms of efficiencies and outcomes for clients, we're going to keep using it."
The middle office of the future: Leaner, but still human orchestrated
The conversation had already touched on the idea that agentic AI will free up middle office teams to work on analysis and decision making. But some may imagine a middle office in the future made up entirely of digital employees, led by a single person.
Berryman doesn't dismiss the scenario, but he offered an important reality check.
"Those agents can still go wrong; no technology is infallible. And also the nature of the business is evolving. We have more complex assets, we have different client relationships, we have different technology we're connecting with. So those agents too will need to be maintained."
Berryman does think middle office teams of the future will be leaner, but people will still play an important role. They will need new skills, as they will have to be accustomed to communicating with AI and good at prompt engineering. But they will still need the same business knowledge they have today: the products, the technology, the client mandates.
"When that technology doesn't work, needs to be tweaked and developed, we need them to be able to make sure it's not hallucinating, help us develop it and change and maintain it, [and] make sure our client expectations are met."
The capacity freed up by automation won't disappear. Berryman sees it being redirected towards more advisory work. The business will turn to middle office professionals for guidance on efficiency and execution when considering new products or capabilities.
Bringing people along on the journey
Not everyone is equally enthusiastic about AI. Berryman has seen both excitement and anxiety, and he shared a memorable example from his career to illustrate the point.
"In a previous role, we automated a capital calculation that had historically taken over a week to complete manually. While the efficiency gains were significant, the transition wasn’t straightforward for everyone. For some, the process had become a familiar and valued part of their role, so removing it required careful support and reassurance. Over time, that capacity was redirected towards more meaningful capital deployment discussions, but it was a good reminder that these changes need to be managed with empathy."
The 11 days of work became half a day. The remaining time was freed up for more meaningful capital deployment conversations. But helping people see that transition as an opportunity rather than a loss requires patience and empathy.
"A lot of people genuinely love it because you minimise the manual task, you give them capacity for more interesting work,” Berryman said. “And I think as humans we love to learn. So those learning new things, they love it - but it does vary depending on the person."
Governance without stifling innovation
A recurring theme in the conversation was the balance between oversight and innovation. Berryman's approach is to encourage experimentation while putting guardrails around how AI tools are developed, tested, and documented.
"Please innovate and develop your AI tools, but at the same time, let's make sure we do it in a controlled way," he said.
This means documenting how AI tools work, providing guidance on prompting, and ensuring independent testing of anything that gets built. GIB Asset Management is also working closely with its risk colleagues, looking at frameworks like model risk management as a way to govern AI without blocking progress.
Maxfield drew the discussion to a close by pointing to a familiar parallel. The current wave of AI innovation, if left ungoverned, risks becoming the equivalent of the macros and end-user computing solutions that proliferated twenty years ago. They were great ideas for tactical automation, but quickly built up a backlog of unmanaged risk.
A governed, curated approach to AI, one that balances innovation with accountability, is not just preferable - it is essential.
Building the middle office of tomorrow
Agentic AI has the potential not just to accelerate work in the middle office, but to fundamentally shift what that work is.
Innovation is everyone’s responsibility, but you need to provide them with guardrails to ensure they can innovate safely.
Crossing the hurdles between a successful AI pilot and a working in-production implementation requires including the end-users early on. They are vital in the design process, but also this gives you the opportunity to get them on board with agentic AI.
Be empathetic with people who are unsure about AI and how it may change their roles; help them to see the possibilities for more meaningful work it is creating.
The promise of AI is enormous, but be prepared to iterate. Strike the right balance between keeping AI capabilities in house and harnessing the scale, experience and resilience of vendors. Partner with the right firms to build the solution you need. Look for customisation, rather than off-the-shelf solutions.
Follow these and you’re well on the way to transforming your middle office.
