The AI Skills Gap Isn't Where You Think It Is
(17)
Overview
Everyone wants AI in the front office first. The readiness data says start in the back. Here's the tier map that shows why.
Year
2026
Industry
inancial Services / COOs & Operations
Challenge
Ask most banks where they want AI first, and the answer is the front office: a trading copilot, a deal-advisory assistant, something a client or a banker will actually see. That's where the budget goes and where the demo gets built. It's also, almost always, the office least ready for it. Not because front-office people are behind. Because front-office work is judgment work such as pricing discretion, deal structuring, relationship calls, and judgment is exactly what an AI Skill can't safely encode on its own. Meanwhile the office nobody pitches at the offsite, the back-office operations, accounting, data management, is sitting on the most rule-based, highest-volume, easiest-to-encode work in the building. It's just not exciting to talk about. That mismatch is the actual AI Skills gap. Not “we don't have enough AI.” A gap between where the enthusiasm is and where the operating model can actually support it. Why “Workflow” isn't the same thing as “AI Skill” A function map usually has a column called something like “Workflow”: trade execution, pricing, client coverage. That's not one AI Skill, it's three, at minimum, each with a different owner, a different definition of “done,” and a different cost of getting it wrong. Treating the workflow label as the skill is how pitches end up vague. The workflow is the raw material. The actual candidate skills are underneath it, and they need to be pulled out one at a time before anyone can say whether the operating model is ready to support them.
Impact
Where the twelve gaps actually sit Once you decompose each function into its candidate AI Skills, you can sort them by how ready the operating model underneath each one actually is, and not by how interesting the use case sounds. Tier 1 means rule-based, high-volume, with a clear “done” state: automate now. Tier 2 means real judgment is involved, but it's bounded enough that AI can draft and a person signs off. Tier 3 means the judgment is relationship- or negotiation-driven and the cost of a wrong call is high, encode later, if at all. Lay it out this way and the pattern is obvious: readiness runs almost exactly opposite to excitement. Back office is nearly all tier 1. Front office is nearly all tier 2 and 3. The office with the least glamorous AI use cases is the one that can absorb them fastest, and the office everyone wants to lead with is the one where the operating model has the least margin for error. What this table is, and isn't This is a template, not a diagnostic. It shows the method, decompose the workflow, sort by how bounded the judgment is, let that sequence the rollout and not a specific company's findings. A real prioritization has to come from the client's own numbers: transaction volume, regulatory materiality, how mature the process already is in each function. Two banks with the same org chart can have completely different tier assignments once you look at their actual data. Presenting this table as if it were someone's real diagnostic would overstate what it is. Where to actually start The instinct to lead with the front office isn't wrong because the front office doesn't matter. It's wrong because it asks the least-ready part of the operating model to go first. The sequencing that actually works is the reverse: start where the process is already rule-based and the “done” state is already clear, prove the model there, and let that build the case, and the operating discipline for tackling the judgment-heavy work later. That sequencing is itself a diagnostic question, not a technology one. Before any of these twelve get built, the question worth answering is which ones your operating model can actually support today, and what specifically has to change in the rest. That's the list and not a slide with “AI roadmap” on it, the list with the gaps named. Fix the operating model. AI works after that and not before.