Plexo papers
Research for organizations learning to operate with AI.
Ideas, evidence, and models for understanding how work changes when machines become more than tools.
Integration Is a Discovery Problem
What a deal model cannot see about how the acquired team works
U.S. oil and gas is consolidating at a pace not seen in a decade, and the most repeated statistic about acquisitions, that 70 to 90 percent fail, has no source. The recent upstream megadeals tell a more useful story. The acquirers report meeting or exceeding their synergy targets, and the clearest account of where the excess came from is one acquirer’s own: a synergy that “could not be modeled in our spreadsheet”, delivered by the acquired team. The industry’s standard operating agreement transfers “all records and data” to a new operator and says nothing about how the work is actually done. The research on transferring practices explains why that gap does not close by itself: the main barriers are knowledge, not motivation, and what predicts integration performance is codification, not experience. We argue that operational integration is a discovery problem, scope the claim to the operating and administrative layer, and say where the evidence runs out.
Who Answers Changes the Number
Why operational diagnostics should not average the office and the field
Ask the same question of different positions in the same organization and the answers differ, systematically. It happens in official U.S. statistics, where the job title of the person filling out a Census survey moves the reported rate of AI use by ten points after controlling for firm size, industry, and state. It happens in the largest process safety culture survey conducted at U.S. refineries, where operators and operations managers at the same site, answering the same item, were more than twenty points apart. The aggregation methods that justify averaging assume sources estimating the same quantity with independent errors. A process owner and an operator meet neither assumption. We propose assigning authority by type of claim, and a working criterion for telling positional disagreement, which is a finding, from noise, which is not.
Said, Seen, and Unseen
Which evidence to trust when most of the work is not on a screen
Before an organization hands work to AI, someone has to find out how that work is actually done. The methods sold for this have split in two: ask the people who do the work, or record their screens and treat observation as the corrective to self-report. In oil and gas, where three of every four jobs in the industry that includes drilling and well services are field occupations, neither is sufficient on its own. We show that the research on verbal reports does not say interviews are unreliable; it says which questions produce data and which produce inference. We show that the authors of process mining do not say that what is missing from a log did not happen; they require the opposite assumption. From both we derive a rule for assigning authority by question rather than by source, and a three-state partition of every process step — on-screen, off-screen, and unobserved — whose purpose is to keep what no source saw from being counted as evidence.
The Agent-Ready Company
Organizational complementarities and the limits of delegating work to autonomous agents
Firms are deploying AI agents into operating models built on the assumption that only humans can interpret context, handle exceptions and exercise judgment. We argue that what limits agent deployment is organizational rather than technical, and we ground the argument in three literatures: the measured complementarity between information technology and organizational capital, the knowledge-based theory of hierarchy, and field experiments on generative AI. They converge on a narrow claim. Measured returns concentrate where operational knowledge was already codified, and delegating past the boundary of what a system does reliably makes performance worse. We also look at why observational shortcuts fail to substitute for asking people directly, and we say which parts of the argument nobody has measured yet.