AI plans have a talent debt problem in financial services

HomeAI Talent Infrastructure

Lee Evans

Strategic Account Executive

Sep 24, 2026

6 min

AI plans have a talent debt problem in financial services

Budgets and platforms are in place at the largest banks. But there’s friction when it comes to staffing.

For example, JPMorgan Chase spent $19.8 billion on technology in 2026. Around 150,000 of its staff use LLM Suite, the bank’s own generative AI platform, every week.

The rest of the industry’s largest names are somewhere on the same track.

HSBC has signed a multi-year agreement with Google Cloud expected to carry more than 200 new AI use cases over two years. They also opened a Global AI Centre of Excellence in Singapore this year. Citi has given generative AI coding tools to 30,000 developers and built an in-house platform, Stylus Workspaces, now extending into agentic work. Goldman Sachs took its GS AI Assistant from a pilot of 10,000 people to GA for all 46,000 staff.

While the tech and platforms are set, successful deployments, adoption, and scale efforts depend on engineers who can build this kind of system inside a regulated bank. This is a tall order simply because there are not enough engineers to go around. When it comes to financial services, that staffing gap will only expand unless some drastic changes to recruitment and retraining, not small adjustments to hiring processes.

What the job actually requires

An engineer building on one of these platforms needs large language models and vector databases. They also need to know why changing an input to a credit decisioning model draws more scrutiny than changing an internal dashboard, why a model touching fair lending sits in a different risk class from one that summarises policy documents, and what happens to a change as it passes through model validation and second-line review. Very few people have both.

Regulation is moving underneath the problem. On 17 April 2026 the Federal Reserve, the OCC and the FDIC issued SR 26-2, retiring the SR 11-7 model risk guidance that had stood for fifteen years. Banks above $30 billion in assets go from fixed annual revalidation to risk-based oversight tied to materiality. It is guidance rather than an enforceable standard, and it excludes the technology everyone is deploying. Footnote 3 says generative and agentic models "are novel and rapidly evolving. As such, they are not within the scope of this guidance."

Where that leaves banks: Writing governance for their own generative and agentic systems, out of existing risk practices, while those systems are being built. Due to compliance and regulation complexity, it is not a situation where you tap in a specialist hired last month who has yet to learn how the institution runs.

Hiring conditions are poor everywhere. ManpowerGroup surveyed 39,063 employers in 41 countries this year and found AI skills harder to fill than any other skillset, ahead of engineering and IT. Average time to fill an AI role in UK financial services has gone from five and a half months to roughly nine. Robert Half put 93% of hiring managers in the sector as already struggling before anyone adds AI to the specification.

The bidding war

Paying more can work for a while. PwC’s 2026 Global AI Jobs Barometer, built from more than a billion job adverts across 27 countries, puts the wage premium for AI skills at 62% (up from 57% the year before). That figure covers every sector, and banks are competing inside it against employers with very different cost structures.

Frontier AI labs advertise base salaries above $600,000. Visa filings put median base pay at OpenAI and Anthropic close to $300,000. Mainstream enterprise hiring runs at $170,000 to $230,000. PwC found 91% of senior US financial services executives are already raising pay for AI skills. For an institution with decades-old compensation bands, that means an approved exception for every hire. A handful is manageable. However, the headcount scale these roadmaps assume converts it to a policy question rather than a hiring one.

Where the gap actually sits

The talent shortage is concentrated in one kind of engineer: someone who can take a model, a pipeline, or an agent and connect it to a core banking estate that is regulated, validated, and often decades old. A veritable needle in a haystack situation.

Those estates are already understaffed. 71% of mainframe teams report being short of people, and Futurum found 79% of organisations struggling to fill mid-career legacy roles. IBM’s z/OS 2.5 goes out of service in September 2026. When Citi rolled out AI coding tools to 30,000 developers it presented the move as part of a modernisation programme.

Banks are handling this very differently from one another. Citi has made prompt-writing training mandatory for 175,000 employees. Standard Chartered has put SC GPT in front of roughly 80,000 staff. DBS, OCBC, and UOB are retraining around 35,000 people between them, with Singapore government subsidies covering up to 90% of salary costs. Whether breadth-first training of that kind produces the integration engineers these roadmaps need is an open question. It does produce a larger pool to select from.

The integration work itself can be taught. Nobody arrives knowing how to connect a retrieval system to transaction monitoring, or an agent to a credit review process. There is a decent population of engineers with the fundamentals to learn it through structured exposure.

Now, the regulatory judgement aspect takes considerably longer. Developing a feel for how a change moves through validation, and for where the second line will push back, accumulates over years in one institution. The market cannot supply that at short notice. Fortunately, a fair number of people already on the payroll have it.

What an external hire does not bring

Every offer made to close a vacancy is visible to the competitors chasing the same shortlist, and it raises the price of the next one.

A new hire, however strong, also does not know the ins and outs of bringing AI into a heavily regulated, institution-specific environment. They lack knowledge about things like: how this particular bank validates a model, who signs off a change to a monitoring rule, or which committee meets monthly and which meets quarterly. Building that knowledge takes time, and on a nine-month hiring cycle the clock has not started. That’s why losing an engineer who already has it is devastating. There’s a 99% chance their replacement costs market rate, and arrives at their first day without the critical knowledge.

The internal option

A more useful first question is which of the engineers already employed sit closest to the profile the roadmap needs. Answering it properly means assessing people against a defined target profile across technical, delivery, and regulatory dimensions. Putting the whole department through a generic AI course and seeing who engages produces a different result.

Standard Chartered has cost mapped out the comparison. Tanuj Kapilashrami, the bank’s Chief Strategy and Talent Officer, puts savings at roughly $49,000 for each employee reskilled and redeployed internally instead of hired externally. Multiply that across the headcount an AI roadmap assumes and it stops being a training decision.

There’s a talent drain expectation across financial services. Nearly eight in ten of the same leaders expect their workforce to shrink by at least 20% over five years, with entry-level and middle-management roles most exposed. Reskilling and headcount reduction are being planned in the same rooms, which makes the internal pitch harder to land than it looks on paper. It remains one of the few routes by which some of those people end up inside the AI roadmap rather than outside the company.

Placement and reskilling answer different questions. Placement covers what cannot wait: a platform build already running, a regulatory deadline, a programme behind schedule. Reskilling is slower and shapes the team over two or three years.

Most banks intend to do a bit of everything. In PwC’s survey, 61% plan to upskill existing employees, 62% to hire AI-specific skills and 57% to partner with external providers. What is usually missing is any sequence between the three. In the same survey, 77% said most of their AI investment has yet to show measurable return.

The resourcing gap

JPMorgan has around 500 AI use cases in production and is targeting 1,000. Most large institutions can point to something comparable. However, resourcing plans of the same quality are harder to find. Writing one means putting a number on how many of these engineers the bank needs, then admitting that number is not available to hire. This is a more uncomfortable metric to take to a board than a roadmap.

Banks that come through the next two years with their plans intact will mostly have done it by retraining people they already employ and hiring externally only where they can win.

Where Andela fits

Andela places vetted specialist engineers where a timeline will not move, and runs structured reskilling programmes for teams already in place.

Goldman Sachs offers a placement example. As the bank extended its digital transformation into emerging markets, Andela built six specialised teams covering SREs, full-stack developers, data engineers, UI/UX designers, and delivery leads. In total, we placed more than 30 engineers across cloud, compliance, CRM and digital lending. The technical evaluations were written around Goldman’s own bar, rather than a generic screen. Placement success ran at 93%.

Reskilling talent in place looks different. A GitHub Copilot programme delivered with GitHub Learn started with 200 technologists and is scaling towards 3,000 across Microsoft’s FY25, run as a certification pathway rather than an open catalogue of courses. Completion has been 92%; 84% of learners have gone on to sit the exam; 70% have passed; and 80% are using Copilot in daily work.

One of Egypt’s largest private banks ran the same programme in a regulated setting. They put 65 engineers through 12 weeks of Copilot certification with Andela’s AI Academy: 50 of its own staff and 15 from Andela. Developers moved Copilot into production workflows during the programme rather than waiting until it finished.

Both routes start in the same place: An honest assessment of where the engineering team already sits against what it has been asked to deliver.

Sources