Many banks now have a reskilling programme. Very few of them have a strategy.
That sounds like a semantic distinction. It isn't. The most common framing of AI workforce transformation in banking — "we need to upskill our people on AI" — is the wrong frame for the problem the industry is actually facing. The honest version of the question is far more uncomfortable: which of the work currently being done at a bank should still be done by humans, and which of the work that doesn't yet exist will need to be?
That question has very different answers from "let's train everyone on generative AI."
The headline numbers everyone quotes — 92 million jobs displaced globally by 2030, 170 million new roles created, a net gain of 78 million (World Economic Forum, 2025) — are correct at the global level and almost useless at the institutional level. That net gain represents a structural labour-market churn equivalent to 22% of all formal jobs in the WEF dataset. The 170 million new roles will not appear at any specific bank unless that bank designs them. The 92 million displaced will show up at banks where the work being done is exposed to automation, whether the institution has decided what to do about it or not.
This is the gap most banking reskilling programmes are not addressing.
The newest empirical evidence sharpens the point. Anthropic's March 2026 study (Massenkoff & McCrory, 2026) introduced a measure called observed exposure, which combines theoretical AI capability with real-world usage data drawn from millions of Claude conversations. Two findings from that study are particularly relevant to banking. First, financial analysts are among the ten most AI-exposed occupations in their data, with task coverage materially higher than the cross-economy average. Second, the most exposed workers across the labour market are older, more educated, and higher-paid — exactly the profile of mid-career bank employees who are typically assumed to be insulated from AI displacement because their roles are "judgment-based." The data suggests otherwise: judgment-based does not mean automation-resistant.
A separate finding deserves direct attention from anyone planning a banking workforce strategy. Brynjolfsson, Chandar, and Chen (2025), using high-frequency payroll data from ADP covering millions of US workers, found a 13% relative decline in employment for workers aged 22–25 in the most AI-exposed occupations since late 2022 — even after controlling for firm-level shocks. The Anthropic study reports consistent evidence: hiring of younger workers into exposed occupations has slowed measurably since ChatGPT's release. The bottom of the organisational pyramid is already being eroded. The middle is next. Reskilling programmes built around the assumption that the existing org chart absorbs the change have the structure wrong.
THE CATEGORY ERROR
Most reskilling programmes I have reviewed fall into one of two patterns.
The first is broad-based AI literacy training — every employee gets a half-day course on what large language models are, how to write a good prompt, and the limits of what AI can do today. Several major banks have committed publicly to training their entire workforce on generative AI tools in 2025 and 2026 (IIF & EY, 2025). There is nothing inherently wrong with this kind of training. But there is little evidence — and limited mechanism — for it to change how the institution actually operates. AI literacy raises the floor of competence across the workforce. It does not, on its own, change the work.
The second pattern is technical upskilling for specific cohorts — data engineers learning to deploy models, risk analysts learning Python, compliance officers learning to audit AI systems. This is more useful, but it is solving a narrow problem. It produces individuals who know more. It does not produce an institution that works differently.
Neither of these is reskilling in the strategic sense. They are training. The difference matters: training expands the capability of the existing workforce to do the existing work. Reskilling, properly understood, is the redesign of who does what work, at what level of judgment, with what tools, against what accountability — and then the rebuilding of the workforce around that redesign.
Most banks have not done the redesign. They are training their existing workforce on the assumption that the underlying organisational structure is broadly correct and just needs to absorb new tools. It isn't, and it doesn't.
WHAT IS ACTUALLY CHANGING IN BANKING WORK
Three structural shifts are reshaping what banking work looks like over the next three to five years. Reskilling programmes that don't account for them will produce trained people doing obsolete jobs.
1. Judgment moves up the stack
The work that AI can do credibly today is the work of generating a draft, a recommendation, a first-pass analysis. The work humans will increasingly do is reviewing, challenging, contextualising, and deciding. A credit analyst in 2030 will not spend their time pulling data and building spreadsheets — that work will be done by agents. They will spend their time questioning what the agent's recommendation missed, what risks the model is structurally blind to, and whether the decision passes a judgment threshold no model can articulate.
This is not the same job with better tools. It is a different job. The skills required — model literacy, structured scepticism, the ability to interrogate without rejecting — are not the skills the role hires for today. Autor and Thompson (2025) make a related argument in their NBER paper on expertise: when automation removes some tasks from a role, the value of the remaining tasks depends on whether the expertise required for those tasks rises or falls. In banking, the residual tasks after agent-automation are concentrated in judgment, contextual interpretation, and accountability — which raises expertise requirements rather than lowering them. Banks that promote on "AI tool familiarity" rather than on depth of judgment will end up with the wrong people doing the wrong work.
2. The boundaries between roles dissolve
The traditional role architecture in banking — analyst, associate, manager, director — was built around tasks. AI commoditises the tasks. What remains is the work of orchestrating multiple AI agents, multiple data sources, and multiple stakeholders into a coherent output. That is not what "associate" describes.
The roles that will exist in the next decade — agent operations lead, model governance officer, AI risk underwriter, decisioning systems architect — do not map onto current job families. The work isn't more senior, more junior, or more technical than what exists today. It is differently shaped.
3. Senior expertise becomes more valuable, not less
The common assumption is that AI levels the playing field — anyone with the tool can produce work that previously required years of experience. In a narrow technical sense this is true. In the work that actually generates competitive advantage at a bank, it is exactly wrong.
AI amplifies the difference between a great senior practitioner and a mediocre one. The senior asks better questions, recognises when the AI is confidently wrong, knows what the model has been trained out of saying, and has the judgment to override the recommendation when it matters. The mediocre practitioner with the same tool produces fluent-sounding nonsense at scale — something I have written about in the context of LLMs specifically.
This dynamic is consistent with two empirical observations from the recent literature. Massenkoff and McCrory (2026) find that the most AI-exposed workers in the economy are also among the most highly-paid and educated. Hampole and colleagues (2025), using a different methodology applied to AI exposure measured from 2010–2023, show that mean task exposure reduces labour demand within firms but that the concentration of exposure across tasks offsets this — workers whose exposure is concentrated in a few automatable tasks can reallocate effort to the remaining, higher-judgment work. The implication for banks: under-investing in senior talent — assuming AI will close the gap — risks losing exactly the people whose remaining work is most valuable.
THE RESKILLING MATH DOES NOT WORK FOR MOST BANKS
The headline statistic — 170 million new roles created globally — masks an institutional problem. Those new roles do not appear at a bank by default. They appear at banks that design them. The displacement, on the other hand, is the default. If a role's work can be automated and the institution has not redesigned it, the role will erode whether the incumbent was reskilled or not.
This produces a specific failure mode that is emerging in real time across the industry: institutions that have invested heavily in AI literacy training, run change-management programmes, and given every employee a generative AI tool — and that have made no measurable progress on workforce transformation because they have not changed the work.
The output of the training does not match the input the institution actually needs. People are AI-literate. The jobs are unchanged. The decisioning structure is unchanged. The performance metrics are unchanged. The only thing that has changed is that the same work now takes slightly less time, which produces slightly more capacity, which is absorbed by the next quarter's expanded workload. Capacity gets harvested. Capability does not.
This is what an unsuccessful AI transformation looks like. Not a failed pilot. A successful AI literacy programme inside an unchanged operating model.
WHAT BANKS SHOULD ACTUALLY DO
Four shifts, in order of where most institutions are weakest.
→ Identify the net-new capability roles before training anyone. Twenty to fifty roles that don't exist in the current org chart but will need to in three years. Agent operations, AI risk, model governance, decisioning architecture, prompt engineering for production systems. These are not vague — each can be specified in a job description today. If they cannot be specified, the institution does not yet have a workforce strategy.
→ Map current senior talent against those roles, not against generic AI skills. Some of the strongest senior people are 80% of the way to roles that have not yet been named. They are the highest-leverage internal transitions. The remaining gaps will need external hires — accept this, plan for it, and stop pretending training will close every gap.
→ Stop running generic AI literacy training and start running role-specific transition programmes. A 100-person cohort of credit risk analysts being redesigned into a 60-person team of AI-augmented credit governance professionals is a different programme from a 10,000-person company-wide course. The first one changes the bank. The second one is a line item.
→ Restructure performance metrics and incentives. Most bank performance reviews still reward output metrics — files processed, deals closed, reports filed — that AI is commoditising fastest. If the metrics do not change, the behaviour does not change, and the reskilling does not stick. This is the part most transformation programmes underestimate and the part most institutions cannot do from the inside, because it requires confronting performance frameworks built up over decades.
Some of this work is internal. Some benefits from outside perspective — particularly the question of what the role architecture should look like, which is difficult to do from inside an organisational structure one is immersed in. Digital Bank Expert's IT strategy consulting and team extension model are designed for exactly this kind of redesign work: bringing senior banking technologists into the conversation alongside the in-house team, then leaving the institution with the capability to evolve the model on its own.
THE WINDOW
This is not an academic argument, even if the academic evidence is now starting to confirm it. The banks that handle this well over the next 24 months will be operating with workforces redesigned around what AI can and cannot do. The banks that do not will be operating with AI-literate workforces doing the same jobs they did before, generating slightly more output and the same value, while their best people leave and their decisioning structures decay.
The displacement curve is steeper than the creation curve in the short term. The 92 million displaced will hit before the 170 million created. Institutions that have already started designing the new roles will be hiring from the displaced pool. Institutions that have not will be displaced into it.
This is the same dynamic I described in a different context — AI lowers the floor and raises the ceiling at the same time, and the middle gets hollowed out. The reskilling decision is the workforce-level expression of the same logic. The banks that build the right capability set will dominate. The banks that train their existing workforce on tools without redesigning the work will not.
The window for getting this right is open. It is not unlimited.
Bibliography
All entries verified accessible as of May 2026.
- Autor, D. H., & Thompson, N. (2025). Expertise (NBER Working Paper No. 33941). National Bureau of Economic Research. https://www.nber.org/papers/w33941
- Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/
- Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). GPTs are GPTs: An early look at the labor market impact potential of large language models (arXiv:2303.10130). https://arxiv.org/abs/2303.10130
- Hampole, M., Papanikolaou, D., Schmidt, L. D. W., & Seegmiller, B. (2025). Artificial Intelligence and the Labor Market (NBER Working Paper No. 33509). National Bureau of Economic Research. https://www.nber.org/papers/w33509
- Institute of International Finance & EY. (2025). 2025 IIF–EY Annual Survey Report on AI Use in Financial Services. Institute of International Finance. https://www.iif.com/Publications/ID/6322/2025-IIF-EY-Annual-Survey-Report-on-AI-Use-in-Financial-Services
- Massenkoff, M., & McCrory, P. (2026). Labor market impacts of AI: A new measure and early evidence. Anthropic. https://www.anthropic.com/research/labor-market-impacts
- World Economic Forum. (2025). The Future of Jobs Report 2025. World Economic Forum. https://www.weforum.org/publications/the-future-of-jobs-report-2025/
