- Man Group, a $200B+ alternative investment manager, has trading signals running in production — researched, back-tested, and proposed by AI, with humans reviewing all output — built on a foundation of Claude skills.
- The key lesson is that skills governance, not raw adoption, unlocks enterprise AI: shared, owned, tested, and tracked workflows give consistent, comparable results, which is essential for comparing trading signals.
- They got it wrong before getting it right — early enthusiastic adoption produced "local optimization" skills written by power users (not process owners), exemplified by a hard-coded expense-report skill; a governed skills marketplace fixed this.
Building signals that trade themselves
- Systematic trading is about trading signals: ranking thousands of securities by some factor (e.g. past 3-month returns), backing the winners and shorting the losers, then back-testing the strategy against 15+ years of history to produce statistics like annualized return, drawdown, and Sharpe ratio.
- The trading signal is just the tip of the iceberg; the hard part is the underlying workflows — cleaning data, stitching prices, detecting outliers, infrastructure, and running back-tests — and if teams use different workflows, results aren't comparable and you can't tell whether one idea was better or just measured differently.
- Out of the box, Claude is a powerful general-purpose tool but doesn't know your data, systems, or workflows; you teach it not by retraining or fine-tuning, but by giving it access to your data, capabilities, and workflows via skills.
- Organizational context is your IP and moat — frontier labs won't solve it for you because it's not on the internet; the work is exposing decades of institutional knowledge, not reinventing it, and skills are how that knowledge becomes leveraged.
- Their early "all-in" adoption (workshops, hackathons, blogs, show-and-tells) drove huge adoption but exposed cracks: power users, not process owners, wrote skills, producing local optimizations rather than common organizational solves.
- A vivid failure: a frequent traveler wrote an expense-report skill with a hard-coded cost-center code, so it kept creating expense reports under one approver's cost center for everyone — nobody had reviewed it and the author wasn't accountable for it.
- The solve was a common skills marketplace/library: every skill is visible, tagged, tested with evals, owned by the workflow owner, reviewed, usage-tracked, with a lifecycle and installable by everyone — treat skills like production code because that's what they become.
- Results: Man Group now has 750 of ~17,800 people using Claude Code across developers, quants, finance, and the people team, with over 100 governed skills (and at least as many community skills); plan ownership, review, testing, and retirement before shipping the first skill, not after the hundredth.
Systematic trading — Algorithmic trading that looks across thousands of securities and hundreds of markets to make investment decisions.
Trading signal — A ranking of securities by some factor used to decide which to back (go long) and which to short.
Back-test — Running a codified strategy against many years of historical data to estimate how it would have performed.
Drawdown — A statistical measure of how much a strategy lost when it lost money.
Sharpe ratio — A statistic comparing a strategy's volatility against its return.
Skill — Packaged, composable instructions/workflows that connect AI to an organization's data and capabilities.
Skills governance — The practice of owning, reviewing, testing, tracking, and retiring skills as production assets.
Organizational context — A company's institutional knowledge, data, and workflows — its IP and moat for AI.
Skills marketplace — A library where governed and community skills are visible, tagged, tested, and installable.
I'm Sharan Fernando. I'm head of data and AI at Man Group. Man Group are an alternative investment manager. We manage over 200 billion dollars of assets. Our clients are pension funds, sovereign wealth funds, and large institutions. We manage real people's money — thousands of people's pensions and investment capital. So when we think about AI, the stakes are high for us. Our clients are real people, from the teachers in Canada to the metal workers in Japan. We really need to get AI right. If we get this wrong, we could lose real money.
One of the really large parts of our business is systematic trading, and that represents a huge opportunity to be aided by AI. By systematic trading, I mean algorithmic trading capabilities that look across thousands of securities and hundreds of markets to make investment decisions. Systematic trading is really about trading signals. You can think of a signal like a fantasy football team: we want to pick the best players for our squad based on some intuitive strategy. The green bars would be a starting lineup, the red bars the reserve squad, and the people in the middle might come in — your subs bench. A signal is about thinking about a striker maybe hitting form, but you need to transfer them in before the Friday deadline when the price might go up. Savvy managers think about form, fixtures, and what might happen to get the most points in the season, transferring the right players in at the right time and the right cost.
This is quite similar to systematic trading. A trading signal is really just this with stocks. The bars represent company stocks. We want to back the ones that would make money, and short the ones that won't. In this example, we've ranked the stocks by the past 3-month returns and run that through history to see if it would have made money. The interesting question is always: what is that factor you want to rank things by? What's the strategy to get the right stocks in your team? And does it work? The truth is you never really know. I'd love to be able to tell the future, but I can't. The best thing we can do is look at what happened historically. We codify the strategy and run it against 15 years of history or longer, through lots of macroeconomic environments and stresses, and we see how it performed. That back-test produces lots of statistical factors — how much money did it make (annualized return), how much did it lose when it lost (drawdown), and more complex factors like the Sharpe ratio, which compares the volatility of the strategy versus how much it returned.
It's this systematic trading workflow that we think we can use AI to really enhance — to come up with ideas, to run the back-tests. So there are trading signals running right now in production at Man Group, a regulated investment firm, running real capital, that were researched, back-tested, and proposed by AI. AI came up with the idea, AI got the data, AI ran the back-test, AI wrote up the strategy proposal, and AI productionized the signal. Humans, of course, reviewed all of the output to make sure it was sensible, but AI was at the center of that process. I'm sure you want to know what the signal was and how much money it made — sorry, I'm not going to tell you that, that's our IP. What I'm here to tell you about is our journey: what was the foundation that allowed us to do that, and how can you apply those learnings in your company?
It really starts with AI understanding our workflows. To do that, we use skills. Coming up with the signal is the quick bit. The hard part is everything underneath it — all of the workflows that make it happen. Think of it like an iceberg: the signal is the tip; underneath are all the workflows that make it possible. How do you clean the data? How do you stitch prices? How do you detect outliers? What's the infrastructure it runs on? How do you run those back-tests? This is where it can quickly go wrong. If different teams are running different versions of those workflows, you get different answers. One team's back-test looks amazing, another's looks average, and because they're using different workflows, you don't know whether it was the idea that was better or whether they're just measuring things differently. Shared workflows fix that. One common foundation means effort isn't duplicated and you have consistency. The outputs are comparable, which is extremely important in systematic trading when we're comparing signals.
Out of the box, Claude is an amazing general-purpose tool. It does a lot, but it doesn't know us. It doesn't know our data, our systems, or how we work. The first thing we had to do was teach it — not by retraining it, not by fine-tuning, but by giving it access to our data, our capabilities, and our workflows. That's our superpower. We have decades of institutional knowledge and systematic research and some of the best technical capabilities on the street. If we can connect that with AI, then AI can leverage that superpower. Skills are the connective layer that allows AI to leverage that superpower, so getting them right is paramount.
But we got it wrong before we got it right. We really focused on adoption. We went all-in. We were doing skills workshops — Anthropic helped us with that — hackathons, we wrote a blog, we did show-and-tell sessions. Everybody was writing skills, and the adoption was out of this world. But we started to see cracks. It was really the power users building the skills, not the process owners. That meant all the skills represented a local optimization for one user — they weren't common organizational solves.
Nothing was more symptomatic of this than a show-and-tell session one day. A guy who traveled a lot had loads of expenses, so he wrote a skill for it: he gave Claude pictures of receipts and it would do the expense report for him. It worked really well, and he shared it with a few people in his team. A few days later, the expense approver came and asked, "Why is Claude creating so many expense reports for my cost center? People from technology, the people team — why do I have to approve all of them? I'm in sales." We dug into it, and it was because the cost-center code was hard-coded. It was a local optimization. Nobody had reviewed that skill. It worked for him and his team, so it was going to work for everybody — but that's not the case, and he wasn't accountable for it. It was symptomatic of a broader problem: people were codifying their ways of doing things, not the organizational ways, and in many cases they weren't the workflow owner. This is a huge problem when it comes to back-testing and systematic trading. It becomes a blocker to scaling to enterprise. Agents can't leverage those skills; there's no commonality.
We saw that skills governance started to be the secret sauce that unlocked these enterprise use cases. If you could connect your common workflows to AI and give it access to your data and capabilities, you could allow agents to act on those skills — and allow Claude Code to do problems as complex as systematic trading. So our solve was a common marketplace. Every skill was visible, tagged, and tested with evals. Imagine a library that captures decades of institutional knowledge, with sections for finance, the people department, and research. We care for every skill. The skill is owned by the workflow owner, they're all tested, usage is tracked, they're all reviewed, they have a lifecycle, and they're all visible to everybody to install. It's that care that makes this work and moves skills from individual productivity solves to a foundation that sets you up for the agent age. Through that, we were able to apply skills to systematic trading.
[Demo video] This is Man Knowledge. It's where you'll find our collection of skills and Man Group's context store. The skill suggestions are tailored to each business unit. They have clear ownership and are organized into managed and community skills. Skills and plugins can easily be installed in Claude. Plugins are useful groups of skills — for example, a data plugin which gives us access to Man Group's data sets. We can also install skills individually, like the data-set skill which allows me to search Man Group's alternative data sets. Now those foundational skills are installed, we can get a flavor of building a systematic trading signal. We use the alternative data-set skill to search for research such as credit card data. We ask Claude what credit card data sets are available and it identifies a data set of US consumer transactions. We can plot Amazon's monthly credit card spend against its stock price returns — the blue bars are credit card spend, the line is the stock price, with spikes for seasonal spend like Black Friday and Christmas. Next we run a back-test to see if credit card spend is predictive of the stock price by comparing peaks in credit card spending with the profits and losses of the stock. The signal shows better performance than a buy-and-hold strategy: investing $1,000 in 2021 using the signal would now be worth around $2,500. This could be a fluke for Amazon, so let's run it on a broader universe of retail companies. As there are multiple companies, we run it using our distributed compute infrastructure — each company running an individual worker, then the findings are collected. In this case study we leveraged four skills to create a systematic trading signal. In reality our signal research is much more nuanced, accounting for seasonality, inflation, and broader sets of securities, with agents as well as humans exploring these ideas. The key takeaway is that the governance of these skills is key — it ensures everyone has access to the same data and uses the same workflows.
Hopefully you can see that if you get that foundation right across the board, with access to all of the data, you can leverage more capabilities — from scaling your compute to getting alternative credit card data sets, often owned by different teams. That's what allows you to scale to the agentic platform.
What did we learn? Firstly, focus on organizational context. That is your IP, it's your moat. It's one of the few safe spaces left in AI. The frontier labs are not going to solve context for you. It's not on the internet, they don't know your workflows, and you already have that context — decades of it. The work is on exposing it, not reinventing it, and skills are how those decades of institutional knowledge become leveraged. Treat those skills like production code because that's what they will become. Plan your approach before you plan the rollout: Who's going to own the skill? Who's going to review it? How does it get retired? How does it get tested? Decide this before shipping the first skill, not after the hundredth like us. Adoption is not a licensing problem, it's a people problem. Once you've got the platform in place, you need to encourage people to engage with it and rethink their workflows rather than just augmenting them. That's a training problem and an engagement problem.
Through these ideas we've been able to scale. Man Group is about 17,800 people, and we now have 750 of them using Claude Code — developers, quants, the people team, the finance team, everybody across all the departments. We're seeing a lot of engagement because they're able to use those capabilities in a simple way. They don't need to know about everything; they have the skills platform that understands our workflows. We now have over 100 governed skills and at least as many community skills that are looked after in a library and well governed. This has allowed us to unlock the capability to use AI in systematic trading. Skills governance really unlocks AI at enterprise scale. The thing I'm most proud of is that we've got our eyes on the prize — a solid bedrock built on decades of institutional knowledge. In the not-too-distant future, I can see us having swarms of agents leveraging those skills to look for new investment opportunities.
So what's the takeaway for you? Really think about how you're going to capture that context. Which department owns it? What's the process for governing them? Where will they live? How will you test them? How will you retire them? Connect a golden path from your AI platform to your capabilities and your context. Once you have that basis of knowledge, if you care for it and AI can leverage it, that will really set you up for the agent age. Thank you very much, everyone.
TL;DR
- Man Group, a $200B+ alternative investment manager, has trading signals running in production — researched, back-tested, and proposed by AI, with humans reviewing all output — built on a foundation of Claude skills.
- The key lesson is that skills governance, not raw adoption, unlocks enterprise AI: shared, owned, tested, and tracked workflows give consistent, comparable results, which is essential for comparing trading signals.
- They got it wrong before getting it right — early enthusiastic adoption produced "local optimization" skills written by power users (not process owners), exemplified by a hard-coded expense-report skill; a governed skills marketplace fixed this.
Takeaways
- Systematic trading is about trading signals: ranking thousands of securities by some factor (e.g. past 3-month returns), backing the winners and shorting the losers, then back-testing the strategy against 15+ years of history to produce statistics like annualized return, drawdown, and Sharpe ratio.
- The trading signal is just the tip of the iceberg; the hard part is the underlying workflows — cleaning data, stitching prices, detecting outliers, infrastructure, and running back-tests — and if teams use different workflows, results aren't comparable and you can't tell whether one idea was better or just measured differently.
- Out of the box, Claude is a powerful general-purpose tool but doesn't know your data, systems, or workflows; you teach it not by retraining or fine-tuning, but by giving it access to your data, capabilities, and workflows via skills.
- Organizational context is your IP and moat — frontier labs won't solve it for you because it's not on the internet; the work is exposing decades of institutional knowledge, not reinventing it, and skills are how that knowledge becomes leveraged.
- Their early "all-in" adoption (workshops, hackathons, blogs, show-and-tells) drove huge adoption but exposed cracks: power users, not process owners, wrote skills, producing local optimizations rather than common organizational solves.
- A vivid failure: a frequent traveler wrote an expense-report skill with a hard-coded cost-center code, so it kept creating expense reports under one approver's cost center for everyone — nobody had reviewed it and the author wasn't accountable for it.
- The solve was a common skills marketplace/library: every skill is visible, tagged, tested with evals, owned by the workflow owner, reviewed, usage-tracked, with a lifecycle and installable by everyone — treat skills like production code because that's what they become.
- Results: Man Group now has 750 of ~17,800 people using Claude Code across developers, quants, finance, and the people team, with over 100 governed skills (and at least as many community skills); plan ownership, review, testing, and retirement before shipping the first skill, not after the hundredth.
Vocabulary
Systematic trading — Algorithmic trading that looks across thousands of securities and hundreds of markets to make investment decisions.
Trading signal — A ranking of securities by some factor used to decide which to back (go long) and which to short.
Back-test — Running a codified strategy against many years of historical data to estimate how it would have performed.
Drawdown — A statistical measure of how much a strategy lost when it lost money.
Sharpe ratio — A statistic comparing a strategy's volatility against its return.
Skill — Packaged, composable instructions/workflows that connect AI to an organization's data and capabilities.
Skills governance — The practice of owning, reviewing, testing, tracking, and retiring skills as production assets.
Organizational context — A company's institutional knowledge, data, and workflows — its IP and moat for AI.
Skills marketplace — A library where governed and community skills are visible, tagged, tested, and installable.
Transcript
I'm Sharan Fernando. I'm head of data and AI at Man Group. Man Group are an alternative investment manager. We manage over 200 billion dollars of assets. Our clients are pension funds, sovereign wealth funds, and large institutions. We manage real people's money — thousands of people's pensions and investment capital. So when we think about AI, the stakes are high for us. Our clients are real people, from the teachers in Canada to the metal workers in Japan. We really need to get AI right. If we get this wrong, we could lose real money.
One of the really large parts of our business is systematic trading, and that represents a huge opportunity to be aided by AI. By systematic trading, I mean algorithmic trading capabilities that look across thousands of securities and hundreds of markets to make investment decisions. Systematic trading is really about trading signals. You can think of a signal like a fantasy football team: we want to pick the best players for our squad based on some intuitive strategy. The green bars would be a starting lineup, the red bars the reserve squad, and the people in the middle might come in — your subs bench. A signal is about thinking about a striker maybe hitting form, but you need to transfer them in before the Friday deadline when the price might go up. Savvy managers think about form, fixtures, and what might happen to get the most points in the season, transferring the right players in at the right time and the right cost.
This is quite similar to systematic trading. A trading signal is really just this with stocks. The bars represent company stocks. We want to back the ones that would make money, and short the ones that won't. In this example, we've ranked the stocks by the past 3-month returns and run that through history to see if it would have made money. The interesting question is always: what is that factor you want to rank things by? What's the strategy to get the right stocks in your team? And does it work? The truth is you never really know. I'd love to be able to tell the future, but I can't. The best thing we can do is look at what happened historically. We codify the strategy and run it against 15 years of history or longer, through lots of macroeconomic environments and stresses, and we see how it performed. That back-test produces lots of statistical factors — how much money did it make (annualized return), how much did it lose when it lost (drawdown), and more complex factors like the Sharpe ratio, which compares the volatility of the strategy versus how much it returned.
It's this systematic trading workflow that we think we can use AI to really enhance — to come up with ideas, to run the back-tests. So there are trading signals running right now in production at Man Group, a regulated investment firm, running real capital, that were researched, back-tested, and proposed by AI. AI came up with the idea, AI got the data, AI ran the back-test, AI wrote up the strategy proposal, and AI productionized the signal. Humans, of course, reviewed all of the output to make sure it was sensible, but AI was at the center of that process. I'm sure you want to know what the signal was and how much money it made — sorry, I'm not going to tell you that, that's our IP. What I'm here to tell you about is our journey: what was the foundation that allowed us to do that, and how can you apply those learnings in your company?
It really starts with AI understanding our workflows. To do that, we use skills. Coming up with the signal is the quick bit. The hard part is everything underneath it — all of the workflows that make it happen. Think of it like an iceberg: the signal is the tip; underneath are all the workflows that make it possible. How do you clean the data? How do you stitch prices? How do you detect outliers? What's the infrastructure it runs on? How do you run those back-tests? This is where it can quickly go wrong. If different teams are running different versions of those workflows, you get different answers. One team's back-test looks amazing, another's looks average, and because they're using different workflows, you don't know whether it was the idea that was better or whether they're just measuring things differently. Shared workflows fix that. One common foundation means effort isn't duplicated and you have consistency. The outputs are comparable, which is extremely important in systematic trading when we're comparing signals.
Out of the box, Claude is an amazing general-purpose tool. It does a lot, but it doesn't know us. It doesn't know our data, our systems, or how we work. The first thing we had to do was teach it — not by retraining it, not by fine-tuning, but by giving it access to our data, our capabilities, and our workflows. That's our superpower. We have decades of institutional knowledge and systematic research and some of the best technical capabilities on the street. If we can connect that with AI, then AI can leverage that superpower. Skills are the connective layer that allows AI to leverage that superpower, so getting them right is paramount.
But we got it wrong before we got it right. We really focused on adoption. We went all-in. We were doing skills workshops — Anthropic helped us with that — hackathons, we wrote a blog, we did show-and-tell sessions. Everybody was writing skills, and the adoption was out of this world. But we started to see cracks. It was really the power users building the skills, not the process owners. That meant all the skills represented a local optimization for one user — they weren't common organizational solves.
Nothing was more symptomatic of this than a show-and-tell session one day. A guy who traveled a lot had loads of expenses, so he wrote a skill for it: he gave Claude pictures of receipts and it would do the expense report for him. It worked really well, and he shared it with a few people in his team. A few days later, the expense approver came and asked, "Why is Claude creating so many expense reports for my cost center? People from technology, the people team — why do I have to approve all of them? I'm in sales." We dug into it, and it was because the cost-center code was hard-coded. It was a local optimization. Nobody had reviewed that skill. It worked for him and his team, so it was going to work for everybody — but that's not the case, and he wasn't accountable for it. It was symptomatic of a broader problem: people were codifying their ways of doing things, not the organizational ways, and in many cases they weren't the workflow owner. This is a huge problem when it comes to back-testing and systematic trading. It becomes a blocker to scaling to enterprise. Agents can't leverage those skills; there's no commonality.
We saw that skills governance started to be the secret sauce that unlocked these enterprise use cases. If you could connect your common workflows to AI and give it access to your data and capabilities, you could allow agents to act on those skills — and allow Claude Code to do problems as complex as systematic trading. So our solve was a common marketplace. Every skill was visible, tagged, and tested with evals. Imagine a library that captures decades of institutional knowledge, with sections for finance, the people department, and research. We care for every skill. The skill is owned by the workflow owner, they're all tested, usage is tracked, they're all reviewed, they have a lifecycle, and they're all visible to everybody to install. It's that care that makes this work and moves skills from individual productivity solves to a foundation that sets you up for the agent age. Through that, we were able to apply skills to systematic trading.
[Demo video] This is Man Knowledge. It's where you'll find our collection of skills and Man Group's context store. The skill suggestions are tailored to each business unit. They have clear ownership and are organized into managed and community skills. Skills and plugins can easily be installed in Claude. Plugins are useful groups of skills — for example, a data plugin which gives us access to Man Group's data sets. We can also install skills individually, like the data-set skill which allows me to search Man Group's alternative data sets. Now those foundational skills are installed, we can get a flavor of building a systematic trading signal. We use the alternative data-set skill to search for research such as credit card data. We ask Claude what credit card data sets are available and it identifies a data set of US consumer transactions. We can plot Amazon's monthly credit card spend against its stock price returns — the blue bars are credit card spend, the line is the stock price, with spikes for seasonal spend like Black Friday and Christmas. Next we run a back-test to see if credit card spend is predictive of the stock price by comparing peaks in credit card spending with the profits and losses of the stock. The signal shows better performance than a buy-and-hold strategy: investing $1,000 in 2021 using the signal would now be worth around $2,500. This could be a fluke for Amazon, so let's run it on a broader universe of retail companies. As there are multiple companies, we run it using our distributed compute infrastructure — each company running an individual worker, then the findings are collected. In this case study we leveraged four skills to create a systematic trading signal. In reality our signal research is much more nuanced, accounting for seasonality, inflation, and broader sets of securities, with agents as well as humans exploring these ideas. The key takeaway is that the governance of these skills is key — it ensures everyone has access to the same data and uses the same workflows.
Hopefully you can see that if you get that foundation right across the board, with access to all of the data, you can leverage more capabilities — from scaling your compute to getting alternative credit card data sets, often owned by different teams. That's what allows you to scale to the agentic platform.
What did we learn? Firstly, focus on organizational context. That is your IP, it's your moat. It's one of the few safe spaces left in AI. The frontier labs are not going to solve context for you. It's not on the internet, they don't know your workflows, and you already have that context — decades of it. The work is on exposing it, not reinventing it, and skills are how those decades of institutional knowledge become leveraged. Treat those skills like production code because that's what they will become. Plan your approach before you plan the rollout: Who's going to own the skill? Who's going to review it? How does it get retired? How does it get tested? Decide this before shipping the first skill, not after the hundredth like us. Adoption is not a licensing problem, it's a people problem. Once you've got the platform in place, you need to encourage people to engage with it and rethink their workflows rather than just augmenting them. That's a training problem and an engagement problem.
Through these ideas we've been able to scale. Man Group is about 17,800 people, and we now have 750 of them using Claude Code — developers, quants, the people team, the finance team, everybody across all the departments. We're seeing a lot of engagement because they're able to use those capabilities in a simple way. They don't need to know about everything; they have the skills platform that understands our workflows. We now have over 100 governed skills and at least as many community skills that are looked after in a library and well governed. This has allowed us to unlock the capability to use AI in systematic trading. Skills governance really unlocks AI at enterprise scale. The thing I'm most proud of is that we've got our eyes on the prize — a solid bedrock built on decades of institutional knowledge. In the not-too-distant future, I can see us having swarms of agents leveraging those skills to look for new investment opportunities.
So what's the takeaway for you? Really think about how you're going to capture that context. Which department owns it? What's the process for governing them? Where will they live? How will you test them? How will you retire them? Connect a golden path from your AI platform to your capabilities and your context. Once you have that basis of knowledge, if you care for it and AI can leverage it, that will really set you up for the agent age. Thank you very much, everyone.