Most of the conversation about AI in financial services focuses on what the technology will replace: analysts, compliance functions, middle-office operations. It is a reasonable frame, but it keeps landing on the wrong problem. Having spent years building and running operations across fintech and regulated investment management, the more pressing issue is not what AI eliminates. It is what running an AI-native, tokenised operating model actually requires from the people still in the room. The industry is trying to run infrastructure it has not yet staffed for. AI is already reshaping what work looks like inside fund management, operations, compliance and client services. Tokenisation is changing the infrastructure those functions sit on. And the workforce, in most firms, was hired and structured for a world that looked nothing like this. That gap is not a future planning concern. It is showing up now in organisational friction, operational risk and the inability of functions that used to work in sequence to work simultaneously.
The Financial Services Skills Commission’s Future Skills Report 2025 puts a number on it: demand for the 13 skills the industry needs most is running approximately 20% ahead of supply. Those skills span data analytics, machine learning, cybersecurity, software development and coaching. The FSSC data shows that AI-related skills carry the largest supply-demand gap at 35 percentage points. Coaching and empathy sit alongside it. The institutions that navigate this well will bring the whole workforce through the transition, not bolt specialists onto an unchanged structure underneath. The World Economic Forum’s Future of Jobs Report 2025 sets this in broader context. 170 million new roles will be created globally by 2030 while 92 million are displaced. Employers expect 39% of workers’ core skills to change or become obsolete by 2030. Of those needing retraining, 11 in every 100 are unlikely to receive it. That is an organisational risk with a name and a number.
Source: Financial Services Skills Commission, Future Skills Report 2025|percentage-point gap between demand and supply
From answering to executing what AI agents change
We have spent the past two years adapting to AI that answers questions. The shift now underway is more consequential: AI that executes tasks. Perplexity’s pivot from search to autonomous agents produced a 50% revenue surge in a single month, pushing ARR past $450 million in March 2026. That is not a product metric. It is a market signal about where value is migrating, from AI that summarises to AI that acts. For financial services, the operational implications are immediate. Portfolio teams are already working with AI-generated investment signals requiring human validation before any decision is made. Compliance functions receive automated monitoring alerts at a volume no team was sized to review manually. Client reporting is increasingly AI-assisted, creating quality control obligations rather than removing them. These are still human-supervised workflows. Agentic AI changes that equation. An agent authorised to execute trades, file regulatory reports or manage client communications on behalf of a firm is not answering a question. It is making consequential decisions at machine speed, and the accountability has to land somewhere.
That somewhere is people. The human checkpoint becomes more demanding, not less, as AI becomes more autonomous. Two infrastructure problems sit underneath this. The first is machine identity: existing digital systems were built to verify humans, not agents acting on their behalf. When an AI agent initiates a transaction or files a report, proving it has the authority to do so and that the human or institution behind it has consented, is not a solved problem. The second is programmable money: traditional payment rails were built for human-speed friction. Stablecoins and smart contracts provide the real-time, rule-based execution agentic workflows require but most firms are operating those alongside legacy fiat infrastructure, not instead of it. Both problems require people who understand what the systems are doing and can identify where governance breaks down. This is why the skills gap matters beyond headcount. Data privacy and security obligations do not diminish because an agent is handling the data. Bias testing requirements do not disappear because a model is acting autonomously. Continual learning is not optional when agentic systems are making consequential decisions: models need retraining as markets shift and regulations evolve. A firm treating AI governance as a one-time implementation rather than an ongoing discipline is already carrying risk it has not priced.
Running two worlds simultaneously
Asset managers, hedge funds and private market operators are navigating versions of the same structural shift at different speeds. Institutional managers are compressing settlement toward real time. Private market firms are grappling with tokenised secondaries and on-chain NAV reporting. Hedge funds are exploring tokenised structures that could reshape how liquidity terms and redemption rights are managed. What makes this hard is that none of these firms get to leave the old world behind. SWIFT does not disappear because Kinexys exists. T+2 settlement does not stop because some funds move toward T+1 or T+0. Custody arrangements built on decades of legal precedent do not get replaced overnight by smart contracts. The foreseeable future is parallel infrastructure: digital asset rails and legacy systems co-existing, interacting and occasionally conflicting. The FCA, SEC and ESMA are all working through how existing rules apply to tokenised instruments, not replacing them wholesale. Firms need people who can operate across both simultaneously, understand where the two systems interact and where a decision on one rail has consequences on the other.
Tokenisation compounds the pressure
The tokenisation of real-world assets adds a structural layer on top of the AI and agentic challenge. The tokenised RWA market had surpassed $30billion, with BlackRock’s BUIDL fund now having grown to $2.5 billion in AUM and JPMorgan’s Kinexys having processed over $1.5 trillion in tokenised repos. These are not pilots. Settlement compressing from days to seconds means reconciliation becomes instantaneous. It introduces oracle-grade data requirements and new risk categories. Larry Fink’s 2026 Chairman’s letter made tokenisation a central strategic theme. When the world’s largest asset manager says this publicly, workforce planning that ignores it is a governance problem. The skill demand sits on top of the AI gap, not instead of it: people who can read a smart contract; explain the difference between a tokenised deposit and a stablecoin to a pension trustee; and monitor cross-chain oracle reliability alongside conventional requirements. The demand is cumulative.
Human skills and governance are not the soft part
The FSSC’s research is consistent: demand for relationship management and empathy outpaces demand for most technical skills. The WEF identifies analytical thinking, creative thinking, resilience and leadership among the fastest-rising skills globally, alongside AI literacy rather than instead of it. As AI agents take on more execution, the human role shifts toward oversight, judgment and accountability. Automated processes can handle volume. They cannot handle the conversation where a client’s confidence is in question, or the moment a regulator asks who was responsible for a decision an agent made. Under the EU AI Act, transparency about when AI is used and what it is doing is a legal requirement for high-risk applications. The PRA’s SS1/23 requires AI incident reporting within 72 hours. The FCA’s approach to AI emphasises human accountability for high-impact decisions. These obligations require people close enough to the systems to know when something has gone wrong and senior enough to act quickly. Human oversight and autonomy become more demanding, not less, as AI agents execute more.
Moving people rather than replacing them
Reskilling an existing employee costs an average of £31,800. Making a role redundant and rehiring costs £80,900. The FSSC reports twice as many people were reskilled in 2024 as in 2022, but cohort sizes averaging 223 per firm are nowhere near the scale the transition requires. The WEF estimates 59 out of every 100 workers will need reskilling by 2030. Eleven of those are unlikely to receive it. The questions that belong at board level are not complicated. Are AI systems, including agentic ones, being tested for bias and fairness across client groups? Can the firm explain to a regulator what human oversight looks like when an agent is executing rather than advising? Do compliance officers understand smart contract risk well enough to escalate it? What does the operating model look like supporting T+0 and T+1 settlement alongside legacy T+2 processes? The firms building cross-functional teams, running meaningful internal mobility programmes and embedding AI governance into roles that were not previously technical will find the transition manageable. The firms treating it as something to address in the next planning cycle are already running late.
The technology is not waiting. The regulation is not waiting. The clients moving assets on chain are not waiting. The workforce strategy needs to move at the same pace.
This article first appeared in Digital Bytes (19th of May, 2026), a weekly newsletter by Jonny Fry of Team Blockchain.