X Marks the Spot: Unearth Your Company's Buried Expertise While Building an AI Agent

Read this article on LinkedIn.
Building an excellent AI agent is not a technology problem. It’s a DIKW–data, information, knowledge and wisdom (read: expertise)--capture problem. You can use the best AI models, but if the information you are feeding those models is incomplete, guess what? Your agent will irrevocably produce poor results. This problem is time consuming and tedious: How much of what your best people know is actually captured anywhere? How much precious intel has been lost when one of them leaves? How can you preserve that information?
Tacit knowledge–the knowledge that your best people know because, well, they’ve been doing this for a helluva long time– is the real buried treasure you’re sitting on. Uncover it and you’ll unlock the real potential of AI.
Your Treasure Map: Complexity and Codification
Before you can dig, you need a map. Every process in your business–from product technical support to internal payroll–sits somewhere along these two coordinates: 1) Level of Complexity and 2) Extent of Codification.
Level of Complexity: How much real variation exists in what the process is supposed to handle?
Extent of Codification: How much of the process already exists somewhere, in a manual, a system of record, a transcript–anything besides someone's memory? And is that knowledge accessible to the people who need it?

Low complexity, already codified. The steps don't vary much and someone already captured them. This is the easiest ground to build on, and rules-based systems or simple agents tend to work well here immediately. UPS didn't need an LLM to figure this out for delivery operations. Its "340 Methods," some dating back to the 1920s, spell out the most efficient way to do nearly everything a driver does, down to buckling a seatbelt with one hand while starting the ignition with the other. Every one of those methods started as something an experienced driver just knew or a time and motion study uncovered. UPS wrote it down and turned individual habits into a company-wide standard, one method at a time.
Low complexity, uncodified. The work itself doesn't vary that much, and it's never been captured properly or simply isn’t accessible. Think of equipment troubleshooting across a chain of restaurants. The knowledge might reside somewhere in part–a dusty manual–but it’s not accessible to the front line. Expert judgment isn't required here because the task doesn’t vary much. You can take what’s been codified and give it to a model, where it can be easily accessed by the frontline, say via a chat or search interface.
High complexity, already codified. Someone was forced to write this down because the cost of relying on memory and ad hoc execution was too high. Rocket launch or surgical procedure checklists are the obvious examples, but so is what Google did with infrastructure reliability. Keeping systems running used to be treated as an art, dependent on a handful of engineers who had internalized how things tended to break. Google turned that into site reliability engineering: defined practices, error budgets, playbooks that didn't live or die with any one person's memory. Palantir built an entire business on the same idea applied to enterprise data. Mapping how a large company's systems and processes actually fit together used to require a small army of consultants rebuilding that picture from scratch at every client. Palantir turned the mapping process itself into reusable structures, so the next engagement didn't start from zero. Knowledge repositories like this can be made readily accessible via agents or chat interfaces. However, often the high consequence of these processes still require guardrails.
High complexity, uncodified. These are things like HVAC repair, or the kind of auto engine diagnosis that takes a mechanic twenty years to get a feel for. Nobody captured this knowledge because the cost was deemed too high to capture (too many nuances, too many branches on the decision tree, too much context required to perform) and/or too high to make that knowledge easily accessible to a lay user. It was thought to be cheaper to hire, develop, and retain experts. For example the London regulated black taxi “green badge” typically required 3 years of study and fast recall knowledge of 25,000 streets and 20,000 points of interest. Traditional software hasn’t successfully codified this because the number of eventualities is too great. This is where AI is a major improvement. AI can think through an infinite number of scenarios and support codification. Nevertheless, a human-in-the-loop is still required.
The shovel just got a lot bigger
Here's what's changed. What took those companies years and serious budgets, AI is compressing into months, sometimes weeks.
In the low complexity, already codified quadrant, the ground is already firm, so rules-based systems or simple agents tend to work well immediately. Here’s a useful framework to determine where to start.
In the low complexity, uncodified quadrant, AI's biggest contribution is speed of cleanup. A generative system can ingest what scattered documentation exists, fill in the obvious gaps, and get a process usable fast, since there was never much tacit judgment standing in the way.
In the high complexity, already codified quadrant, AI can absolutely help, but it can’t run alone. Codification reduces the risk, it doesn't eliminate the judgment required. You still want rules and a human in the loop.
In the high complexity, uncodified quadrant, AI's real contribution is both extraction, and accessibility. This is where an agent can sit with your most experienced people, ask the questions a good apprentice would ask, and turn their reasoning into something explicit for the first time.
It's early. Most companies haven't done this yet. Industry surveys keep landing on the same finding: a large share of enterprise AI efforts stall not because the model is weak, but because the knowledge the model would need was never captured anywhere. That gap is the opportunity.
An agent isn't smart about your business by default. It only knows what you've handed it. Building one isn't summoning a new kind of intelligence into your company, it's figuring out how to capture what your best people know already. Stay tuned for the next piece where Aaron and I explain how.