Audiences AITwo shapes, twenty industries
For most of this year we described our work as two products: a decision layer for retail demand signals and a triage system for the plant floor. In August we sat down to map what comes next and found the framing had been wrong all along. What we had was not two products. It was two recurring shapes of decision, and once the shapes had names we started seeing them in almost every industry we examined.
The first shape is a STATE problem. A system observes something correctly and still cannot say what it means. A customer stops replenishing; the CRM sees the silence perfectly and cannot tell a price defection from a doubt about authenticity. Same behavior, different underlying reason, different correct action. The second shape is a TRIAGE problem. A system records an event correctly and cannot say whether it deserves action, what actually caused it, or where it should go. A plant-floor system logs NO MATERIAL. A clinical trial system logs MISSED VISIT. A property queue logs WATER LEAK. Same code, different context, different correct response.
Both shapes resolve into the same kind of software: a small engine of rules and mappings that sits beside the system of record. Epic keeps the surgical schedule. Veeva and Medidata capture and route protocol deviations. Procore records the RFI. Buildium logs the work order. These systems do their recording jobs well, and we intend to deploy beside every one of them, handing back an interpretation with the reasoning attached. We considered a heavier path, a HealthcareOS here and a ConstructionOS there, and rejected it; the planning memo calls that route "a consulting company pretending everything is a product." The engine stays tiny. The vertical intelligence packs, the encoded judgment of each domain, are the asset worth building.
One thing should be plain before the survey begins: everything below is research territory, and Interpret and Triage remain the only systems available for pilot.
The eight-part opportunity test
Walking two shapes through twenty industries in a month forces quick decisions about which ones deserve a second week. The discipline that emerged is an eight-part test. We pursue a vertical only when most of these hold:
- The signal already exists digitally. No new sensors, no new data infrastructure to sell.
- An incumbent system records or predicts the event, yet the contextual judgment about it still happens manually.
- The same observable event can legitimately mean several different things.
- Those meanings lead to materially different actions, including, sometimes, doing nothing.
- The set of possible actions is finite enough to encode.
- The reasoning can be explained and audited after the fact.
- The wrong action costs real money, time, risk, or relationship value.
- The logic can deploy into the incumbent stack and cooperate with what is already installed.
Clinical-trial protocol deviations pass nearly every clause, the cleanest fit the sweep produced. FDA's protocol-deviation guidance notes that FDA regulations do not establish a system for classifying deviation types, while stressing consistent identification, classification and reporting. ICH E6(R3) asks sponsors to determine trial-specific criteria for classifying deviations as important, tied to participant safety and the reliability of trial data. Read together, that is an open invitation for sponsor-owned decision logic that travels with the protocol, whichever CTMS records the deviation.
The test earns its keep just as much through what it excludes. Prior authorization looked tempting until we noticed regulation rebuilding it directly, with CMS forcing electronic prior-auth APIs. Banking AML matches the triage shape almost perfectly and carries a validation burden we decline to pretend away. One caveat survived every revision of the plan: integrating decision logic can genuinely take a week, while a one-week enterprise go-live must never be claimed in a regulated industry, where security review, EHR configuration, clinical governance, GxP validation and change control can each exceed the code effort.
The three-examples rule
The test filters industries; a second rule filters our own enthusiasm. Before any vertical gets built, we require at least three worked examples of the same signal producing opposite decisions, each with a quantified account of why the wrong decision hurts. Fail to produce three and the vertical stays in the notebook.
The retail and manufacturing pairs came first. A lapsed replenishment driven by a cheaper competitor wants an offer; the same lapse driven by authenticity anxiety wants reassurance while the discount stays suppressed, because a discount can confirm the doubt. A NO MATERIAL stop with demand loaded and a supplier running late is recoverable; the identical code with no demand behind it marks a line that is correctly idle. The research verticals produced their own pairs. An open RFI that asks for information and nothing more can wait its turn; an identical entry touching the critical path threatens schedule and money the day it lands. A repeated HVAC failure with a cheap component at fault wants a repair; the same ticket on an asset past its replacement threshold wants the repairs stopped and a capital review opened.
Twenty industries went into the sweep and four investigations came out of the first pass, spanning protocol deviations, procedure readiness, RFI triage, and the maintenance-versus-capex line in property. That ratio is the point. The eight clauses now hang above our research queue as a standing filter: every candidate vertical, whether it arrives through a client conversation or our own pattern hunting, gets held against them before a line of logic is written. What gets built next is whatever passes.
If a signal in your industry keeps meaning different things to different people, tell us where it hides.