Field notes on vertical AI, agent systems, enterprise operations, and the engineering that holds it together.
Every claims automation project starts with the same funnel slide. Then the project goes live, and the number of claims that still need an adjuster barely moves. The hard part is the operating knowledge that lives inside your best adjusters.
8 min read
The straightforward risks move fast and automation handles them. Then a real case lands on the desk, and the senior underwriter everyone routes the hard files to is exactly the person your underwriting AI cannot replicate.
8 min read
Every operations leader eventually meets the same ceiling. One person's judgment holds the operation together, and the throughput of the entire function is capped at what that person can process in a day. You cannot hire more of them. But you can scale the judgment itself.
8 min read
Most operations leaders can name their key person risk in about ten seconds. What they do not have is a way to size the risk, defend it in a board meeting, or do something about it before the notice period. This post is the diagnostic, and the mitigation frame that comes after it.
8 min read
Most succession plans are org-chart exercises. A name in a box, a successor in the next box, a review date on the calendar. When the person in the box leaves, the successor inherits the title and none of the twenty years of judgment that made the role work.
8 min read
Every CHRO knows the number that never shows up in the deck. A new senior underwriter or demand planner will not really be productive for two to three years. The org chart says the seat is filled the day they sign the offer, and the operation knows otherwise.
13 min read
Ask a CAIO what the AI stack looks like and you will hear a familiar list: Glean, Notion AI, Copilot. Ask what the business impact has been and the answer gets quieter. People search a little faster, notes come out cleaner, and none of that is the real work.
7 min read
One leg is action, the execution layer that reads and writes to systems. The other is knowledge, the captured expertise of the people who actually know how the work gets done. The market has spent two years building the action leg. Almost nobody has built the knowledge leg.
8 min read
Every enterprise buyer asks the same question: what stops the agent from doing something we cannot undo? The market keeps selling autonomy, and autonomy is exactly the wrong picture for an operation with real money, real customers, and real regulators on the other end of every action.
9 min read
Ask a COO for a copy of how the work is done and you will get documents. Then sit next to the person who actually runs that work and watch which field they check first, which one they ignore, and which exception they escalate. None of that is in the documents, and that gap has a name.
8 min read
Every few weeks, someone in an enterprise AI meeting asks the same question: we already have a RAG stack, why would we also need expertise mining? It is a fair question. The answer is that the two are not competitors, they are different tools for different jobs, sitting on completely different substrates.
8 min read
Ask an operating leader which workflow blows up if one person disappeared tomorrow, and they answer inside three seconds. That instant recall is the problem. The risk is known, it is named, and it is unmanaged.
9 min read
Two years inside real enterprises, shipping agents that had to work on real cases under real deadlines. Almost every failure traced back to the same missing piece. Not the model, not the integration. The company could not tell you how the work actually gets done.
12 min read
Before Maria retires next quarter, someone volunteers to sit with her and document everything. Six months after she leaves, the folder is still there and nobody opens it. Most tribal knowledge capture fails because the method is wrong for the kind of knowledge it is trying to capture.
14 min read
Every enterprise has a beautiful process map now, built on real event log data. Then you sit down with the person who actually runs the process, and within ten minutes you realize the map tells you what happened but not why. Different substrate, different output.
8 min read
The board asked for an AI strategy. The strategy shipped, pilots were funded, vendors were selected. A year later, when the CFO asks what it earned, the honest answer is: not much. MIT's NANDA initiative found 95% of enterprise AI programs delivered no measurable P&L impact.
12 min readA working AI agent stands on two legs: knowledge and execution. Almost the entire market is building leg two. Almost no one is building leg one.
9 min readExpertise lives in decisions. A process without decisions is just movement. For every decision, ask three things: what did you decide, what did you check, and why did it matter?
11 min readThe system stored the records. The documents stored the rules. The experts stored the truth. This works until the expert leaves, or the company grows.
8 min readThe strongest agent deployment sequence is not model first. It is expertise first, context second, agent team third, and improvement loop after production.
15 min readThe difference between an AI answer and an AI operation is the context contract around the work.
14 min readGoverned Agent Operations turns extracted expertise into manageable product behavior instead of hiding the system inside one large prompt.
16 min readAsk any operations leader whether their processes are documented. The answer is usually "mostly, yes." Ask whether those documents describe how the work actually gets done, and the answer changes.
9 min readThe best self-improving agent systems do not change production behavior on their own. They observe, diagnose, propose, wait for approval, release, and measure.
15 min readThe best vertical AI rollouts start small, prove expertise capture, then expand into governed agent teams with clear owners and phase gates.
17 min readRAG is a useful tool. It is not enough. The faster the enterprise AI market understands why, the faster it will stop repeating the same mistake.
10 min readBetter models, better frameworks, bigger budgets, same result. After two years of agentic delivery, one pattern shows up in every project, and it was never about the technology.
10 min readMedian U.S. private-sector tenure is 3.5 years. Yet most companies treat their most experienced operators as permanent fixtures. When the expert finally gives notice, the playbook is always the same, and it never works.
9 min read
