The CIO role in an AI world
Updated: Sep 20
“I don’t know who the winners will be and quite possibly the winners might not even exist yet.” - Richard Tomlinson
Chat about AI is everywhere. Not joking, but I recently went to a meditation seminar called “Staying Grounded in an AI world”. This topic is rightly getting plenty of attention in investing too. In a field as dynamic as AI and investing, conclusions quickly become stale, so please tell me where you disagree.
On the CIO Chair podcast, where Sean Thompson and I interview CIOs, AI has come up frequently and almost exclusively positively. Most variants of the thought process are: if AI can process more information, find patterns and build models faster than humans, how can we rethink the decision process around humans plus AI, and where will humans continue to make the highest impact?
What is AI already good at?
AI is exceptional at processing digital information. It can search documents, compare disclosures, identify changes in language, populate models and highlight inconsistencies between sources.
Outside the field of investing, in closed form problems like chess, AI has shown that it is pretty much as powerful as human + AI. Performance changes when the problem is less well defined - in most cases, human + AI is better than AI alone. That's a useful context when we are considering investing.
Rob Groves described its use in credit research. Analysts review financial results, public statements, sector developments and ratings information. Specialist AI tools can make this work easier and more scalable, keep it more up to date and expand the reach of an individual human analyst.
Andy Chorlton, CIO of Fixed Income at M&G, pointed towards where he thought the value was:
“We want our analysts to spend more time thinking and less time doing.”
In his framing, the analyst’s main value is not building models or processing information. It is challenging the conclusions and deciding: what does this information mean for us?
AI is currently strongest where the task can be specified, the information exists and the output can be checked. What did the company report? Which covenants changed? What growth assumptions sit inside the rating model? These questions help us improve our understanding of “what is the case today”. There is enormous value in answering this more fully and quickly.
Jeremy Rogers, CIO of Better Society Capital, described using several AI agents with different roles rather than relying on one system. He is using AI to identify missed information and blind spots. Importantly, the person leading the work at Better Society Capital previously worked as an investment manager and had excellent context for using AI. In this case, the technology is being designed around the investment process rather than bolted on from outside.
The result should be to push investment professionals towards the margin. Once the information that we should already know has been processed, the human contribution moves towards what is not readily observable: incentives, competing agendas, misaligned interests and the credibility of the people executing the strategy.
Investment decisions
The problem with replacing entire decision-making teams with AI is that analysis is not an investment recommendation. Assumptions about what the market, company or valuation might look like in future can matter more than understanding today’s position.
AI models can help here. What has happened to similar investments or themes? Aside from the usual limitations of back-testing, this analysis cannot tell us whether the future will resemble the past.
It can, however, help humans build a broader and better-informed distribution of what the future might look like. That is valuable, even if it does not produce a reliable central forecast.
It is important that we do not gloss over this point. One component of both analysis paralysis and overconfidence is our understanding of how well we can make a prediction. For some, there is never enough analysis. For others, their back-test is enough. Because investing is a long-term, fairly idiosyncratic business, no individual investment or strategy can be relied upon to outperform indefinitely.
This was echoed by the CIOs we interviewed. None claimed an ability to make consistently accurate predictions, so the inability to predict the future is a structural problem rather than a technological one. Where technology has been directed towards markets, it has tended to focus on shorter-term anomalies. From high-frequency traders to quantitative investment firms, short-term reversions have proved fertile territory. As the same tools become more widely available, however, some of that advantage may be competed away.
Humans are still likely to be required for long-term outperformance. AI can interrogate an earnings-call transcript and explain management’s stated strategy. Alpha lies in deciding whether that strategy is achievable and, if not, how the miss feeds through to earnings and the company’s borrowing profile. The analysis can be prepared by AI, but weighing it up remains as much an art as a science.
This moves the role of humans in the loop from producing reports to scoping and scrutinising them. Huge leverage can be delivered by outsourcing the production of analysis to AI, but the value of that analysis can only be unlocked if it is critically understood.
Humans do not need to out-calculate machines. The human needs to direct it towards the important problem, recognise when the answer is incomplete and decide when human involvement is required again.
Preparing for an unknowable future
The message our CIOs return to is that good portfolios should not depend on correctly predicting the future. David Thompson, CIO of Zurich UK, described forecasting the next black swan as “a bit of a fool’s game”. Simon Pilcher said historical returns and correlations only take an investor so far. USS considers different futures and seeks to build a portfolio able to thrive in most and survive practically all of them.
Guillermo Donadini split risk assessment into two parts. One can be illustrated through stress tests. The other is whether decision-makers can “stomach that risk” at the bottom of a drawdown without feeling compelled to de-risk.
Guillermo highlights the role of an accountable human who has built relationships with, and communicated consistently with, an investment committee. AI cannot yet determine whether an investment committee will tolerate the risk profile of a portfolio.
It also cannot carry the accountability when the answer is wrong. Unless investors explicitly expect a fully automated strategy, they will expect someone to explain the decision, defend the assumptions and own the consequences. Agency and accountability are therefore not peripheral to the CIO role. They are integral to it.
I am not convinced that AI solves the knottiest problem: the long-term forecasting used in asset allocation, particularly where leverage is involved. It had not been solved before AI either.
AI may help us think more deeply about the range of outcomes considered and the probability attached to them, but that is quite different from removing uncertainty or knowing which outcome will occur.
So where can we point AI to help?
AI has made analysis easier, but the reality remains disconnected from some of the promise.
Area | Where AI can help | Potential impact |
Analysing history | Analyse larger datasets, include older records and test broader relationships. | High |
Understanding today | Combine data sources, compare disclosures and identify red flags, changes or opportunities. | High |
Forecasting | Build a wider distribution of outcomes, generate scenarios and test assumptions. | Medium to Low |
Portfolio interactions | Look through asset-class labels to identify common factors, concentrations and second-order risks. | Medium |
Strategy and horizon | Test mandate limits, liquidity needs and consistency with stated objectives. Less able to assess stakeholder behaviour or the institution’s real time horizon. | Low |
The analysis of history and aggregation of current information are ripe for improvement. Systematic strategies should benefit, and private-market investors can use AI for opportunity scanning, due diligence and underwriting.
As use becomes widespread, however, access to the same models and information is unlikely to remain a source of alpha. Advantage will move towards how the tools are directed and how their output is incorporated into decisions.
Overreliance
CIOs we have spoken to have concluded it would be a strategic own goal not to use AI to improve analysis. The danger arises when users fail to understand what can be inferred from that analysis and with what confidence.
Regular readers know my concerns about boiling complicated things down too far. Simplification can discourage consideration of scenarios that have not yet played out. Siddharth Chakravarty took us through the AI infrastructure build theme that is affecting several markets. Richard Tomlinson, CIO of Local Pensions Partnership Investments, calls some diversification “optical diversification”.
Using AI to draw out common factor risks that are not easily captured through industry sectors might help us understand portfolio-level behaviour. It can also make us overconfident about what happens next.
Overreliance is not inevitably a permanent weakness. As systems improve and repeatedly demonstrate their reliability, handing them more responsibility may be justified. The challenge is knowing where reliance is not only efficient by justified and where plausible output is being mistaken for understanding.
So what?
AI can read more, compare more and calculate more than an investment team could manage alone. It should improve research and remove repetitive work.
The harder part is forecasting. AI can help produce a better distribution of possible futures, but the CIO must still understand horizon, interactions, human behaviour and what has been lost when the world is reduced to factors.
The human contribution increasingly sits at the margin: framing the question, identifying what is not observable, directing further analysis and recognising when the machine needs help.
For now, humans retain primacy in understanding the institution, setting strategy and owning the decision. That boundary may move with time.
The risk is not only that investment teams trust an incorrect answer. It is that they lose the ability to challenge or probability-weight a plausible one. Machines can explore. Humans should direct and decide.
AI should help us analyse and think better. The CIO’s role is to make sure it helps as much as it can and no more, whilst ensuring that a human remains accountable for the result.
Next week's blog will be "Intuition revisited". As always get the blog delivered directly to your inbox on Home | Deciders | for mental fitness | change your mind. Other blogs in the CIO series What's happening in PE? Building winning investment teams
Comments