Building Agents with Intuition
Scaling the expertise of the most capable professionals
Imagine if the highest quality legal or accounting advice was accessible to anyone, not just those who can afford premium fees. Historically, white-collar professions have offered highly variable service quality based on affordability because elite experts—the top 1%—can manage only a handful of clients annually. Could that change if startup professional services firms embed decades of tacit knowledge into reasoning agents and make it widely accessible?
Like many, I’m spending significant time thinking about how AI will impact white collar industries by automating much of the manual, repetitive work. The most obvious starting points are roles governed by clear Standard Operating Procedures, such as customer support and claims management. These roles, with their detailed sequences of decisions, actions, and identifiable exceptions, lend themselves readily to automation.
However, the truly transformative opportunity may be in professional fields where knowledge is not strictly rule-based, but instead deeply intuitive, developed over years of practice and experience. These professions often have wildly varying levels of service depending on affordability, because the top 1% of experts can handle only a limited number of clients annually, charging premium prices and creating significant barriers to access.
Consider senior lawyers, who rely extensively on tacit knowledge gained through years of practice, careful observation, and personal relationships to navigate complex legal challenges.
What if startups could create AI reasoning agents embedded with the nuanced judgment of these top-tier experts, making that expertise broadly accessible? What if startups could enable Linklaters at the cost of LegalVision?
To effectively democratise this level of expertise, we must find ways to embed tacit knowledge reliably into AI systems. In 1998, Laura Militello and Robert Hutton published a landmark paper titled “Applied Cognitive Task Analysis (ACTA): A Practitioner’s Toolkit for Understanding Cognitive Task Demands.” ACTA systematically extracts expert knowledge through several phases:
Creating a task diagram to outline the broader scope of the activity.
Conducting a knowledge audit to pinpoint exactly where and how expert judgment is applied.
Performing simulation interviews to understand how experts approach specific incidents.
Synthesising insights from multiple experts into a cognitive demands table.
Approaches like ACTA could help us begin to document the lived experience of the most experienced professionals.
In November last year, I wrote about how AI products should progressively automate as users build trust in their capabilities. This strategy would also apply when building products for use in professional services firms.
Initially, a product might target 60% accuracy, leveraging the founders’ industry expertise combined with insights from ACTA performed alongside early design partners. This first step could allow a startup professional services firm get off the ground.
Subsequently, different forms of reinforcement learning (RL) would allow professionals to correct the model through regular usage. Over time, the product would gradually automate decision-making, continually refining its handling of complex, less predictable scenarios. While you would still need people for client management and to make final decisions in regulated and trust-based industries, the number of clients those professionals could manage would increase dramatically.
By training AI reasoning agents on the expertise of the top 1% of professionals in their fields, previously exclusive knowledge could become widely accessible. This democratisation would break down barriers related to cost and availability, transforming entire industries traditionally constrained by limited access to elite human expertise.