Espen Skogen has spent two decades inside the machinery of financial services technology, and he's watched a pattern emerge that many advisory leaders still describe as a surprise. AI arrived in consulting and law before it arrived almost anywhere else, and the reason has little to do with carelessness. The structure of the work itself made these sectors the first point of contact. Professional services firms built their economics on layers of junior staff performing structured, language-intensive tasks. Research memos, due diligence summaries, first-draft contracts, discovery review, slide decks. That is precisely the category of work large language models were built to do.
The exposure was baked into the business model long before the technology existed.
The Anatomy of the First Strike
Consider the logic chain. Fact one: advisory and legal work runs on documents, precedents, and structured argument. Fact two: large language models process and generate exactly that material at near-zero marginal cost. Fact three: the traditional leverage model in these firms depends on billing junior hours against that material. The conclusion follows without much strain. When the cost of producing a competent first draft collapses, the economics underneath the pyramid start to compress.
Skogen, who operates at the intersection of law, software engineering, and finance as chief executive of RocketFin, sees this as a diagnosis problem rather than a technology problem. In his view, firms that frame AI adoption as a tooling decision have misread the situation. The question sits deeper, in how the work gets staffed and who holds the judgment.
💡 The pattern he identifies: the tasks most exposed to AI are the tasks firms used to train their people on. Removing them without a replacement plan creates a judgment gap five years out.
The Courtroom as an Early Warning System
The same dynamic is playing out in law, and it reaches beyond the firms themselves. Tools like ChatGPT now allow litigants in person, people representing themselves without a solicitor, to engage with court proceedings at a level that would have been unheard of five years ago. They draft submissions, research precedent, and structure arguments with a fluency that previously required paid representation.
As a qualified solicitor who writes code, Skogen reads this as a signal rather than a curiosity. Access to structured legal language stopped being scarce. The scarce asset shifted to judgment: knowing which argument holds under scrutiny, which precedent actually applies, which procedural step carries risk. That distinction between production and judgment now defines the entire professional services landscape.
Production got cheap. Judgment stayed expensive.
Firms that internalize this early are restructuring around it. Firms that deliberate are, in effect, accruing debt against their future relevance, and the interest compounds quietly.
The Speed Differential Is Already Measurable
The gap between adopters and deliberators shows up in delivery data, and the baseline was already fragile before AI entered the picture. BCG research shows that more than two-thirds of large-scale tech programs are not expected to be delivered on time, within budget, or to their defined scope. In regulated finance, the stakes of getting this wrong are concrete. Nine major UK banks and building societies accumulated 803 hours of tech outages across 2023 and 2024, the equivalent of 33 days of downtime.
Heavily regulated sectors also face the steepest adoption path. Research on AI project outcomes notes that in financial services, explainability and validation requirements lengthen timelines, and the cost of a wrong decision keeps approval bars stringent.
Skogen's reading of these numbers is characteristically structural. The failure rate correlates with team composition and decision layers, a pattern research supports. Studies have found that project team size was significantly negatively correlated with team performance, with larger teams experiencing decreased productivity and weaker problem solving.
⚠️ The implication for AI adoption is direct: an organisation that already struggles to ship conventional technology on time will struggle even more with technology that demands rapid iteration and senior judgment at every step.
Small Senior Teams Absorb AI Faster
Here the industry trend intersects with the thesis Skogen built RocketFin around. Small teams of senior specialists outperform large groups of generalists, and AI amplifies that advantage rather than eroding it. Research published in Harvard Business Review, examining millions of papers, patents, and software projects, found that while large teams advance and develop existing fields, small teams are the ones that disrupt them.
The mechanism is straightforward. A senior specialist reviewing AI output applies judgment immediately, catches the subtle error, and moves on. A large generalist team routes the same output through approval layers, and each layer adds latency without adding insight. Large teams commonly suffer from decision paralysis, too many opinions and too much bureaucracy, while small teams communicate directly and decide without waiting.
AI removes the production bottleneck. It leaves the coordination bottleneck fully intact.
This explains why the firms operating at a fundamentally different speed today tend to share a profile. They kept their teams small, staffed them with people who can evaluate output rather than merely generate it, and treated AI as leverage on existing competence. Firms that bolted AI onto a headcount-heavy delivery model found that the technology simply produced mediocre drafts faster, which their review structures then absorbed at the usual pace.
What Deliberation Actually Costs
The deliberating firm rarely perceives itself as falling behind. Its revenue looks stable, its clients remain, its pipeline holds. The erosion happens in relative terms. Competitors compress research cycles from weeks to days. They prototype client solutions during the engagement rather than after it. They redeploy senior time from document production to the diagnostic work clients actually pay premium rates for.
In finance specifically, the risk compounds. Change activity already drives a meaningful share of operational incidents. In 2019, nearly 1,000 material incidents were reported to the FCA, with 17% attributed to change activity. Introducing AI into that environment without senior oversight raises exposure. Introducing it with senior oversight, in small accountable teams, is what the evidence suggests works.
Skogen frames the strategic question the way he frames delivery questions for tier-one institutions. The problem underneath the AI problem is a people problem. Which people evaluate the output, who holds accountability for the judgment call, and how many layers sit between the model and the decision. Firms that answer those questions with precision are compounding an advantage. Firms that answer them with committees are paying an opportunity cost that grows each quarter.
The Question That Determines the Next Five Years
The evidence points in a consistent direction. Structured, language-intensive work is exactly where AI performs, which is why professional services felt the impact first. Access to that capability has democratised, as litigants in person are demonstrating in courtrooms right now. The differentiator has moved decisively toward judgment, and judgment concentrates in small senior teams that can absorb new capability without drowning it in process.
The firms already operating this way didn't get lucky. They designed for it.
So the question worth sitting with is this: if the production layer of your firm's work became free tomorrow, would your structure amplify your best people's judgment, and would you know how to prove it?
Most Large-Scale Tech Programs Fail—Here’s How to Succeed
BCG’s latest research shows that more than two-thirds of large-scale tech programs are not expected to be delivered on time or within budget or to meet their defined scope.
https://www.bcg.com/publications/2024/most-large-scale-tech-programs-fail-how-to-succeed